test: 暂存代码

electron.vite.config.ts

@@ -8,7 +8,8 @@ export default defineConfig({
       rollupOptions: {
         input: {
           index: resolve('src/main/index.ts'),
-          voiceRecognitionWorker: resolve('src/main/voice-pipeline/voice-recognition-worker.ts')
+          voiceRecognitionWorker: resolve('src/main/voice-pipeline/voice-recognition-worker.ts'),
+          knowledgeWorker: resolve('src/main/knowledge/knowledge-worker.ts')
         },
         output: {
           entryFileNames: '[name].js'

package.json

@@ -34,7 +34,10 @@
     "test:unit": "vitest run --config vitest.unit.config.ts",
     "test:component": "vitest run --config vitest.component.config.ts",
     "test:integration": "vitest run --config vitest.integration.config.ts",
+    "benchmark:knowledge": "vitest run --config vitest.knowledge-benchmark.config.ts --reporter=verbose",
+    "benchmark:knowledge:capacity": "cross-env KNOWLEDGE_CAPACITY=1 vitest run --config vitest.knowledge-benchmark.config.ts --reporter=verbose",
     "test:e2e:build": "electron-vite build",
+    "test:knowledge-worker": "pnpm test:e2e:build && node scripts/test-knowledge-worker.cjs",
     "test:e2e": "pnpm test:e2e:build && playwright test --grep-invert @visual",
     "test:visual": "pnpm test:e2e:build && playwright test tests/e2e/visual.spec.ts",
     "test:smoke": "node --test tests/smoke/native-environment.test.mjs",
@@ -59,6 +62,7 @@
     "@electron-toolkit/preload": "^3.0.2",
     "@electron-toolkit/utils": "^4.0.0",
     "@koromix/koffi-win32-x64": "3.1.0",
+    "@radix-ui/react-popover": "^1.1.23",
     "@tanstack/react-virtual": "^3.14.6",
     "archiver": "^8.0.0",
     "cross-env": "^10.1.0",

pnpm-lock.yaml

@@ -13,6 +13,7 @@ specifiers:
   '@electron-toolkit/utils': ^4.0.0
   '@koromix/koffi-win32-x64': 3.1.0
   '@playwright/test': ^1.62.1
+  '@radix-ui/react-popover': ^1.1.23
   '@rollup/rollup-darwin-arm64': ^4.62.2
   '@tanstack/react-virtual': ^3.14.6
   '@testing-library/dom': ^10.4.1
@@ -58,6 +59,7 @@ dependencies:
   '@electron-toolkit/preload': 3.0.2_electron@43.1.0
   '@electron-toolkit/utils': 4.0.0_electron@43.1.0
   '@koromix/koffi-win32-x64': 3.1.0
+  '@radix-ui/react-popover': 1.1.23_eijghdl4n2x4hz6j4cg7ctgbuu
   '@tanstack/react-virtual': 3.14.6_bokjwhiew3ov3ffvbmafuwoalq
   archiver: 8.0.0
   cross-env: 10.1.0
@@ -888,6 +890,34 @@ packages:
         optional: true
     dev: true
 
+  /@floating-ui/core/1.8.0:
+    resolution: {integrity: sha512-0CIZ5itps/8x7BG8dEIhs53BvCUH2PCoogtakwRTut+Arm58sJooJ0AuZhLw2HJYIR5cMLNPBSS728sPho2khQ==}
+    dependencies:
+      '@floating-ui/utils': 0.2.12
+    dev: false
+
+  /@floating-ui/dom/1.8.0:
+    resolution: {integrity: sha512-yXSrzeHZBTZadLOlfyhCkJHNeLJnHRnRInwdZ40L7ZiaAtrBwoYlsDrX3v5zB1Utk7CLfzcOVnVVWoXEky7Ceg==}
+    dependencies:
+      '@floating-ui/core': 1.8.0
+      '@floating-ui/utils': 0.2.12
+    dev: false
+
+  /@floating-ui/react-dom/2.1.9_bokjwhiew3ov3ffvbmafuwoalq:
+    resolution: {integrity: sha512-JDjEFGCpImxDCA7JJKviA0M9+RtmJdj0m/NVU5IMgBK+AmZouAQQ7/+2GLH0GXXY0YMw9oXPB8hKdbPYg5QLYg==}
+    peerDependencies:
+      react: '>=16.8.0'
+      react-dom: '>=16.8.0'
+    dependencies:
+      '@floating-ui/dom': 1.8.0
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@floating-ui/utils/0.2.12:
+    resolution: {integrity: sha512-HpCo8tmWzLVad5s2d19EhAz5zqrrQ6s69qd6moPMQvkOuSwDT1YgRfWSVuc4ennqrgv3OHppiOGMQ7oC13yIww==}
+    dev: false
+
   /@gar/promisify/1.1.3:
     resolution: {integrity: sha512-k2Ty1JcVojjJFwrg/ThKi2ujJ7XNLYaFGNB/bWT9wGR+oSMJHMa5w+CUq6p/pVrKeNNgA7pCqEcjSnHVoqJQFw==}
     dev: true
@@ -1247,6 +1277,355 @@ packages:
       playwright: 1.62.1
     dev: true
 
+  /@radix-ui/primitive/1.1.7:
+    resolution: {integrity: sha512-rqWnm76nYT8HoNNqEjpgJ7Pw/DrBj5iBTrmEPo6HTX5+VJyBNOqTdv4g89G63HuR5g0AaENoAcH7Is5fF2kZ8Q==}
+    dev: false
+
+  /@radix-ui/react-arrow/1.1.15_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-v4zggRcjadnI+ClKDuijlQEW4tw3NoaeHc/PwpKnLoLLKNUG4InLegkstooLcRIUWCs+8L22dGURCVuFfOKfnA==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/react-primitive': 2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-compose-refs/1.1.5_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-+48PbAAbq3didjJxa+OaWY2ZwgAKsNiRGyeHKszblZMQ+kcpd9pAaT11cMkGEie0vsOi3QdeTE6d5Fe3Gn61kA==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-context/1.2.2_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-RHCUGwKHDr0hDGg4X7ma4JG4/+12qxw8rkh5QKdDldlCvtja6nUx1Ef/8HVrJze81lEsgLQlqjzjGNHantgnQA==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-dismissable-layer/1.1.19_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-8g4pfOL9HoKKLWGiypT+dphVqjFfmcXO5GBnhsG6zI+lxAx/8feQpr+1LSN8Re3hiZ+XkLNS4O9ztK11/LzQ6w==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/primitive': 1.1.7
+      '@radix-ui/react-compose-refs': 1.1.5_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-primitive': 2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-use-callback-ref': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-use-effect-event': 0.0.5_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-focus-guards/1.1.6_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-RNOJjfZMTyBM6xYmV3IVGXkPjIhcBAuv48POevAXwrGJhkWZ9p1rFoIS1JFooPuT193AZmRsCPhpoVJxx6OPoQ==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-focus-scope/1.1.16_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-wmRZ2WWLvmt6KHy2rNPOdPUjwq5xOHY02+m+udwJTn0aNIox/rkskAvJTyTLGhPK6KgrUjlJUJpgmx/+wFiFIQ==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/react-compose-refs': 1.1.5_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-primitive': 2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-use-callback-ref': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-id/1.1.4_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-TMQp2llA+RYn7JcjnrMnz7wN4pcVttPZnRZo52PLQsoLVKzNlVwUeHmfePgTgRluXFvlD3GD5g5MOVVTJCO0qA==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-popover/1.1.23_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-mw58MrBlyHWFisTOYignD0vf/3gdcgAR+9of1s9G/38CbFiUwH1nCDkc0AUM9IrXFgN5Ue8n45j9WCgyM1sbiQ==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/primitive': 1.1.7
+      '@radix-ui/react-compose-refs': 1.1.5_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-context': 1.2.2_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-dismissable-layer': 1.1.19_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-focus-guards': 1.1.6_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-focus-scope': 1.1.16_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-id': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-popper': 1.3.7_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-portal': 1.1.17_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-presence': 1.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-primitive': 2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-slot': 1.3.3_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-use-controllable-state': 1.2.6_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      aria-hidden: 1.2.6
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+      react-remove-scroll: 2.7.2_2lt27s3vun674ldqoia6newrwa
+    dev: false
+
+  /@radix-ui/react-popper/1.3.7_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-UsJrrd7w4wuKKTdvd/DNERVlwSlUcyXzjhyDwBk+3aPOsCjOY6ZSbxuw8E6lZTjjfP8Cpd0J8VVkrYUWyGYXyg==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@floating-ui/react-dom': 2.1.9_bokjwhiew3ov3ffvbmafuwoalq
+      '@radix-ui/react-arrow': 1.1.15_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-compose-refs': 1.1.5_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-context': 1.2.2_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-primitive': 2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-use-callback-ref': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-use-rect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-use-size': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/rect': 1.1.3
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-portal/1.1.17_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-vKQLcWypUnwZVvfV7UkGahH2g6ySe8M8R+zYBwPrv5byZ9QAW6cQVvNKo7GgmD+p8aYb6D9JBuvy8/WhOno2wQ==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/react-primitive': 2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-presence/1.1.10_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-3wyzCQ6+ubRA+D4uv9m95JYLXxmOHp05qjrkjeA7uKHHtjpPggQzc6DAb0URl7j67oR0K2foO4ip27TiX037Bw==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-primitive/2.1.10_eijghdl4n2x4hz6j4cg7ctgbuu:
+    resolution: {integrity: sha512-MucOnzh6hR5mid6VpkbglRAMYMjKLqRnGBbjXkzjK52fuQDd1qbkx78a5P40mkcnVXJdEVxm26E9OPAiUq7nBg==}
+    peerDependencies:
+      '@types/react': '*'
+      '@types/react-dom': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+      react-dom: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+    dependencies:
+      '@radix-ui/react-slot': 1.3.3_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      '@types/react-dom': 19.2.3_@types+react@19.2.7
+      react: 19.2.1
+      react-dom: 19.2.1_react@19.2.1
+    dev: false
+
+  /@radix-ui/react-slot/1.3.3_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-qx7oqnYbxnK9kYI9m317qmFmEgo6ywqWvbTogdj7cL9p3/yx4M48p7Rnw5z3H890cL/ow/EeWJsuTykeZVXP5Q==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@radix-ui/react-compose-refs': 1.1.5_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-use-callback-ref/1.1.4_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-R6OUY2e2fA6Yn6s+VSx5KBV6Nx8LQEhu+cz7LCej18rQ1HLyg9PSC9jP/ZNx0o6FAIK9c0F1kHylzSxKsdlkrQ==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-use-controllable-state/1.2.6_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-uEQJGT97ZA/TgP/Hydw47lHu+/vQj6z/0jA+WeTbK1o9Rx45GImjpD0tc3W5ad3D6XTSR6e1yEO0FvGq6WQfVQ==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@radix-ui/primitive': 1.1.7
+      '@radix-ui/react-use-effect-event': 0.0.5_2lt27s3vun674ldqoia6newrwa
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-use-effect-event/0.0.5_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-7cshFL8HGS/7HEiHH+9kL9HBwp2sa9yX18Knwek6KYWmXwM7pegMgta2AXMQKI+rq3JnfSj9x8wYqFMTdG1Jgg==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-use-layout-effect/1.1.4_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-K20DkRkUwDnxEYMBPcg3Y6voLkEy5p5QQmszZgLngKKiC7dzBR/aEuK3w1qlx2JWDUNH6FluahYdgR3BP+QbYw==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-use-rect/1.1.4_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-cSOCh6JlkmfjLyNcLiu2nB4v+nm+dkZ+Q5KHWk/soo4U7ZLiEQFKHK9/YmtBHjfCEaU43IBKQOc4/uJmCaiCTQ==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@radix-ui/rect': 1.1.3
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/react-use-size/1.1.4_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-D3anSY15EJoxrihpsXI6SMrmmonnQtR2ni7arO+Lfdg3O95b9hNXxONk8jA5C8ANdF/h5HMAxejgs8PWJ6rlhw==}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8 || ^17.0 || ^18.0 || ^19.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@radix-ui/react-use-layout-effect': 1.1.4_2lt27s3vun674ldqoia6newrwa
+      '@types/react': 19.2.7
+      react: 19.2.1
+    dev: false
+
+  /@radix-ui/rect/1.1.3:
+    resolution: {integrity: sha512-JtyZR+mqgBibTo8xea3B6ZRmzZiM/YeVBtUkas6zMuXjAlfIFIW2FgqeM9eLyvEaYX66vr6DJMK+4U6LV0KhNw==}
+    dev: false
+
   /@rolldown/pluginutils/1.0.0-beta.53:
     resolution: {integrity: sha512-vENRlFU4YbrwVqNDZ7fLvy+JR1CRkyr01jhSiDpE1u6py3OMzQfztQU2jxykW3ALNxO4kSlqIDeYyD0Y9RcQeQ==}
     dev: true
@@ -1655,13 +2034,11 @@ packages:
       '@types/react': ^19.2.0
     dependencies:
       '@types/react': 19.2.7
-    dev: true
 
   /@types/react/19.2.7:
     resolution: {integrity: sha512-MWtvHrGZLFttgeEj28VXHxpmwYbor/ATPYbBfSFZEIRK0ecCFLl2Qo55z52Hss+UV9CRN7trSeq1zbgx7YDWWg==}
     dependencies:
       csstype: 3.2.3
-    dev: true
 
   /@types/readdir-glob/1.1.5:
     resolution: {integrity: sha512-raiuEPUYqXu+nvtY2Pe8s8FEmZ3x5yAH4VkLdihcPdalvsHltomrRC9BzuStrJ9yk06470hS0Crw0f1pXqD+Hg==}
@@ -2095,6 +2472,13 @@ packages:
   /argparse/2.0.1:
     resolution: {integrity: sha512-8+9WqebbFzpX9OR+Wa6O29asIogeRMzcGtAINdpMHHyAg10f05aSFVBbcEqGf/PXw1EjAZ+q2/bEBg3DvurK3Q==}
 
+  /aria-hidden/1.2.6:
+    resolution: {integrity: sha512-ik3ZgC9dY/lYVVM++OISsaYDeg1tb0VtP5uL3ouh1koGOaUMDPpbFIei4JkFimWUFPn90sbMNMXQAIVOlnYKJA==}
+    engines: {node: '>=10'}
+    dependencies:
+      tslib: 2.8.1
+    dev: false
+
   /aria-query/5.3.0:
     resolution: {integrity: sha512-b0P0sZPKtyu8HkeRAfCq0IfURZK+SuwMjY1UXGBU27wpAiTwQAIlq56IbIO+ytk/JjS1fMR14ee5WBBfKi5J6A==}
     dependencies:
@@ -2723,7 +3107,6 @@ packages:
 
   /csstype/3.2.3:
     resolution: {integrity: sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ==}
-    dev: true
 
   /data-urls/7.0.0:
     resolution: {integrity: sha512-23XHcCF+coGYevirZceTVD7NdJOqVn+49IHyxgszm+JIiHLoB2TkmPtsYkNWT1pvRSGkc35L6NHs0yHkN2SumA==}
@@ -2832,6 +3215,10 @@ packages:
     engines: {node: '>=8'}
     dev: true
 
+  /detect-node-es/1.1.0:
+    resolution: {integrity: sha512-ypdmJU/TbBby2Dxibuv7ZLW3Bs1QEmM7nHjEANfohJLvE0XVujisn1qPJcZxg+qDucsr+bP6fLD1rPS3AhJ7EQ==}
+    dev: false
+
   /dir-compare/4.2.0:
     resolution: {integrity: sha512-2xMCmOoMrdQIPHdsTawECdNPwlVFB9zGcz3kuhmBO6U3oU+UQjsue0i8ayLKpgBcm+hcXPMVSGUN9d+pvJ6+VQ==}
     dependencies:
@@ -3661,6 +4048,11 @@ packages:
       math-intrinsics: 1.1.0
     dev: true
 
+  /get-nonce/1.0.1:
+    resolution: {integrity: sha512-FJhYRoDaiatfEkUK8HKlicmu/3SGFD51q3itKDGoSTysQJBnfOcxU5GxnhE1E6soB76MbT0MBtnKJuXyAx+96Q==}
+    engines: {node: '>=6'}
+    dev: false
+
   /get-proto/1.0.1:
     resolution: {integrity: sha512-sTSfBjoXBp89JvIKIefqw7U2CCebsc74kiY6awiGogKtoSGbgjYE/G/+l9sF3MWFPNc9IcoOC4ODfKHfxFmp0g==}
     engines: {node: '>= 0.4'}
@@ -5118,6 +5510,57 @@ packages:
     engines: {node: '>=0.10.0'}
     dev: true
 
+  /react-remove-scroll-bar/2.3.8_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-9r+yi9+mgU33AKcj6IbT9oRCO78WriSj6t/cF8DWBZJ9aOGPOTEDvdUDz1FwKim7QXWwmHqtdHnRJfhAxEG46Q==}
+    engines: {node: '>=10'}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+      react-style-singleton: 2.2.3_2lt27s3vun674ldqoia6newrwa
+      tslib: 2.8.1
+    dev: false
+
+  /react-remove-scroll/2.7.2_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-Iqb9NjCCTt6Hf+vOdNIZGdTiH1QSqr27H/Ek9sv/a97gfueI/5h1s3yRi1nngzMUaOOToin5dI1dXKdXiF+u0Q==}
+    engines: {node: '>=10'}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+      react-remove-scroll-bar: 2.3.8_2lt27s3vun674ldqoia6newrwa
+      react-style-singleton: 2.2.3_2lt27s3vun674ldqoia6newrwa
+      tslib: 2.8.1
+      use-callback-ref: 1.3.3_2lt27s3vun674ldqoia6newrwa
+      use-sidecar: 1.1.3_2lt27s3vun674ldqoia6newrwa
+    dev: false
+
+  /react-style-singleton/2.2.3_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-b6jSvxvVnyptAiLjbkWLE/lOnR4lfTtDAl+eUC7RZy+QQWc6wRzIV2CE6xBuMmDxc2qIihtDCZD5NPOFl7fRBQ==}
+    engines: {node: '>=10'}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      get-nonce: 1.0.1
+      react: 19.2.1
+      tslib: 2.8.1
+    dev: false
+
   /react/19.2.1:
     resolution: {integrity: sha512-DGrYcCWK7tvYMnWh79yrPHt+vdx9tY+1gPZa7nJQtO/p8bLTDaHp4dzwEhQB7pZ4Xe3ok4XKuEPrVuc+wlpkmw==}
     engines: {node: '>=0.10.0'}
@@ -5912,6 +6355,10 @@ packages:
       typescript: 5.9.3
     dev: true
 
+  /tslib/2.8.1:
+    resolution: {integrity: sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w==}
+    dev: false
+
   /type-check/0.4.0:
     resolution: {integrity: sha512-XleUoc9uwGXqjWwXaUTZAmzMcFZ5858QA2vvx1Ur5xIcixXIP+8LnFDgRplU30us6teqdlskFfu+ae4K79Ooew==}
     engines: {node: '>= 0.8.0'}
@@ -6053,6 +6500,37 @@ packages:
       punycode: 2.3.1
     dev: true
 
+  /use-callback-ref/1.3.3_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-jQL3lRnocaFtu3V00JToYz/4QkNWswxijDaCVNZRiRTO3HQDLsdu1ZtmIUvV4yPp+rvWm5j0y0TG/S61cuijTg==}
+    engines: {node: '>=10'}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      react: 19.2.1
+      tslib: 2.8.1
+    dev: false
+
+  /use-sidecar/1.1.3_2lt27s3vun674ldqoia6newrwa:
+    resolution: {integrity: sha512-Fedw0aZvkhynoPYlA5WXrMCAMm+nSWdZt6lzJQ7Ok8S6Q+VsHmHpRWndVRJ8Be0ZbkfPc5LRYH+5XrzXcEeLRQ==}
+    engines: {node: '>=10'}
+    peerDependencies:
+      '@types/react': '*'
+      react: ^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0 || ^19.0.0-rc
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+    dependencies:
+      '@types/react': 19.2.7
+      detect-node-es: 1.1.0
+      react: 19.2.1
+      tslib: 2.8.1
+    dev: false
+
   /utf8-byte-length/1.0.5:
     resolution: {integrity: sha512-Xn0w3MtiQ6zoz2vFyUVruaCL53O/DwUvkEeOvj+uulMm0BkUGYWmBYVyElqZaSLhY6ZD0ulfU3aBra2aVT4xfA==}
     dev: true

scripts/test-knowledge-worker.cjs

@@ -0,0 +1,153 @@
+const { fork } = require('node:child_process')
+const { existsSync, mkdtempSync } = require('node:fs')
+const { rm } = require('node:fs/promises')
+const { tmpdir } = require('node:os')
+const { join } = require('node:path')
+const { randomUUID, createHash } = require('node:crypto')
+
+const workerPath = join(__dirname, '..', 'out', 'main', 'knowledgeWorker.js')
+if (!existsSync(workerPath)) throw new Error(`Knowledge worker build is missing: ${workerPath}`)
+
+const root = mkdtempSync(join(tmpdir(), 'wxe-knowledge-worker-'))
+const child = fork(workerPath, [], {
+  stdio: ['ignore', 'ignore', 'ignore', 'ipc'],
+  serialization: 'advanced',
+  env: { ...process.env, ELECTRON_RUN_AS_NODE: '1' }
+})
+const pending = new Map()
+
+function request(type, payload) {
+  const requestId = randomUUID()
+  return new Promise((resolve, reject) => {
+    pending.set(requestId, { resolve, reject })
+    child.send({ version: 1, type, requestId, payload }, (error) => {
+      if (error) reject(error)
+    })
+  })
+}
+
+child.on('message', (message) => {
+  if (!message || message.type === 'progress') return
+  const current = pending.get(message.requestId)
+  if (!current) return
+  pending.delete(message.requestId)
+  if (message.type === 'error') current.reject(new Error(message.error))
+  else current.resolve(message.payload)
+})
+
+function fts(profileId) {
+  return {
+    profileId,
+    tokenizer: 'trigram',
+    contentMode: 'external',
+    detail: 'full',
+    columnsize: 1
+  }
+}
+
+function conversation(accountId, id) {
+  return {
+    conversationId: `conversation-${id}`,
+    completeSnapshot: true,
+    messages: [
+      {
+        accountId,
+        conversationId: `conversation-${id}`,
+        messageId: `message-${id}`,
+        createTime: 1,
+        senderId: 'fixture-member',
+        senderName: '脱敏成员',
+        kind: 'text',
+        text: `脱敏索引内容 ${id}`
+      }
+    ]
+  }
+}
+
+function accountPath(accountId) {
+  const key = createHash('sha256')
+    .update(`knowledge-account-v1:${accountId}`)
+    .digest('hex')
+    .slice(0, 32)
+  return join(root, key, 'knowledge.sqlite')
+}
+
+async function main() {
+  try {
+    const chunker = {
+      version: 'conversation-v1',
+      maxGapMs: 600000,
+      maxMessages: 12,
+      maxCharacters: 1200,
+      overlapMessages: 3
+    }
+    const accountA = 'worker-fixture-a'
+    const accountB = 'worker-fixture-b'
+    const first = await request('index', {
+      accountId: accountA,
+      databaseRoot: root,
+      conversations: [conversation(accountA, 'a')],
+      chunker,
+      fts: fts('worker-a')
+    })
+    await request('index', {
+      accountId: accountB,
+      databaseRoot: root,
+      conversations: [conversation(accountB, 'b')],
+      chunker,
+      fts: fts('worker-b')
+    })
+    if (
+      !first ||
+      first.cancelled ||
+      !existsSync(accountPath(accountA)) ||
+      !existsSync(accountPath(accountB))
+    ) {
+      throw new Error('Knowledge worker did not create isolated derived databases')
+    }
+    const search = await request('search', {
+      accountId: accountA,
+      databaseRoot: root,
+      fts: fts('worker-a'),
+      text: '查询脱敏索引内容 a',
+      terms: ['脱敏索引内容', 'a'],
+      limit: 10
+    })
+    const evidence = search?.evidence?.[0]
+    if (
+      search?.state !== 'ready' ||
+      !evidence ||
+      evidence.messageId !== 'message-a' ||
+      evidence.conversationId !== 'conversation-a' ||
+      evidence.sender !== '脱敏成员' ||
+      typeof evidence.timestamp !== 'number'
+    ) {
+      throw new Error('Knowledge worker search did not return message-level evidence')
+    }
+    await request('remove', { accountId: accountA, databaseRoot: root })
+    if (existsSync(accountPath(accountA)) || !existsSync(accountPath(accountB))) {
+      throw new Error('Knowledge worker removal crossed an account boundary')
+    }
+    const unavailable = await request('search', {
+      accountId: accountA,
+      databaseRoot: root,
+      fts: fts('worker-a'),
+      text: '查询脱敏索引内容 a',
+      terms: ['脱敏索引内容'],
+      limit: 10
+    })
+    if (unavailable?.state !== 'unavailable' || unavailable.evidence?.length) {
+      throw new Error('Knowledge worker did not report unavailable index after removal')
+    }
+    await request('close', {})
+    console.log('Knowledge worker integration check passed')
+  } finally {
+    child.kill()
+    await rm(root, { recursive: true, force: true })
+  }
+}
+
+main().catch((error) => {
+  console.error(error)
+  process.exitCode = 1
+})

src/main/index.ts

@@ -104,6 +104,10 @@ import type { ExportRequest } from '../shared/export'
 import { discoverAccounts } from './services/account-discovery'
 import { VoiceRecognitionUseCase } from './voice-pipeline/voice-recognition-use-case'
 import type { VoiceMessageReference } from '../shared/voice-recognition'
+import type { AiSearchPipelineRequest } from '../shared/ai-search'
+import type { KnowledgeSearchIpcRequest, KnowledgeSearchIpcResult } from '../shared/knowledge'
+import { KnowledgeSearchService } from './knowledge/knowledge-search-service'
+import { AiSearchPipelineService } from './services/ai-search-pipeline-service'
 
 // electron-vite can close the child's stdout/stderr after spawning Electron.
 // Plain console.error then throws EPIPE on a closed pipe and crashes the IPC
@@ -112,6 +116,8 @@ installSafeConsole()
 
 let voiceService: VoiceService | null = null
 let voiceRecognition: VoiceRecognitionUseCase | null = null
+let knowledgeSearchService: KnowledgeSearchService | null = null
+let aiSearchPipelineService: AiSearchPipelineService | null = null
 let imageDecryptService: ImageDecryptService | null = null
 let stickerService: StickerService | null = null
 let videoAssetService: VideoAssetService | null = null
@@ -430,6 +436,16 @@ app.whenReady().then(async () => {
     databasePath: join(app.getPath('userData'), 'cache', 'voice-transcripts.sqlite'),
     workerPath: join(__dirname, 'voiceRecognitionWorker.js')
   })
+  knowledgeSearchService = new KnowledgeSearchService(
+    app.getPath('userData'),
+    join(__dirname, 'knowledgeWorker.js')
+  )
+  aiSearchPipelineService = new AiSearchPipelineService(knowledgeSearchService, aiProviderService)
+  knowledgeSearchService.onStatusChange((status) => {
+    for (const window of BrowserWindow.getAllWindows()) {
+      if (!window.isDestroyed()) window.webContents.send('knowledge:status', status)
+    }
+  })
   voiceRecognition.modelManager.setProgressListener((status) => {
     for (const window of BrowserWindow.getAllWindows()) {
       if (!window.isDestroyed()) window.webContents.send('voice:modelProgress', status)
@@ -501,10 +517,13 @@ app.whenReady().then(async () => {
   ipcMain.handle('app-update:install', () => appUpdateService.install())
   ipcMain.handle('cache:getSummary', () => getCacheSummary())
   ipcMain.handle('cache:clear', async (_, scope: CacheClearScope) => {
-    const allowedScopes: CacheClearScope[] = ['bootstrap', 'electron', 'all']
+    const allowedScopes: CacheClearScope[] = ['bootstrap', 'electron', 'knowledge', 'all']
     if (!allowedScopes.includes(scope)) return getCacheSummary()
     imageDecryptService = null
-    return clearCache(scope)
+    return clearCache(scope, {
+      beforeClearKnowledge: () =>
+        knowledgeSearchService?.prepareForCacheClear() || Promise.resolve()
+    })
   })
 
   ipcMain.handle('db:init', async (_, key: string, accountRoot?: string) => {
@@ -891,6 +910,29 @@ app.whenReady().then(async () => {
   })
 
   ipcMain.handle('db:search', (_, keyword: string) => chat.searchMessages(keyword))
+  ipcMain.handle(
+    'knowledge:search',
+    (_, request: KnowledgeSearchIpcRequest): Promise<KnowledgeSearchIpcResult> => {
+      if (!knowledgeSearchService) {
+        throw new Error('本地知识库服务尚未初始化')
+      }
+      return knowledgeSearchService.search(request)
+    }
+  )
+  ipcMain.handle('knowledge:getStatus', () => {
+    if (!knowledgeSearchService) throw new Error('本地知识库服务尚未初始化')
+    return knowledgeSearchService.getStatus()
+  })
+  ipcMain.handle('knowledge:startIndex', () => {
+    if (!knowledgeSearchService) throw new Error('本地知识库服务尚未初始化')
+    return knowledgeSearchService.startCurrentAccountIndex()
+  })
+  ipcMain.handle('ai-search:run', (event, request: AiSearchPipelineRequest) => {
+    if (!aiSearchPipelineService) throw new Error('本地搜索服务尚未初始化')
+    return aiSearchPipelineService.run(request, (progress) => {
+      if (!event.sender.isDestroyed()) event.sender.send('ai-search:progress', progress)
+    })
+  })
 
   ipcMain.handle(
     'ai:chat',
@@ -1444,7 +1486,8 @@ app.on('before-quit', (event) => {
     const [, nativeCallsDrained] = await Promise.all([
       apiServer.stop().catch(() => undefined),
       chat.closeChatDbForQuit().catch(() => false),
-      voiceRecognition?.dispose().catch(() => undefined)
+      voiceRecognition?.dispose().catch(() => undefined),
+      knowledgeSearchService?.dispose().catch(() => undefined)
     ])
     if (!nativeCallsDrained) {
       console.warn('[Shutdown] WCDB async calls did not fully drain before quit')

src/main/knowledge/chunker.ts

@@ -0,0 +1,88 @@
+import { createHash } from 'crypto'
+import type {
+  KnowledgeChunk,
+  KnowledgeChunkerConfig,
+  KnowledgeNormalizedMessage
+} from '../../shared/knowledge'
+import { isIndexableKnowledgeMessage } from './normalizer'
+
+function digest(value: string): string {
+  return createHash('sha256').update(value).digest('hex')
+}
+
+function formatChunkText(messages: KnowledgeNormalizedMessage[]): string {
+  return messages
+    .map((message) => {
+      const sender = message.senderName || message.senderId || '未知成员'
+      return `[${new Date(message.createTime).toISOString()}] ${sender}: ${message.searchableText}`
+    })
+    .join('\n')
+}
+
+function buildChunk(
+  messages: KnowledgeNormalizedMessage[],
+  config: KnowledgeChunkerConfig
+): KnowledgeChunk {
+  const first = messages[0]
+  const last = messages[messages.length - 1]
+  const text = formatChunkText(messages)
+  const messageIds = messages.map((message) => message.messageId)
+  const participantIds = Array.from(
+    new Set(messages.map((message) => message.senderId).filter((value): value is string => Boolean(value)))
+  )
+  const messageKinds = Array.from(new Set(messages.map((message) => message.kind)))
+  const identity = `${first.accountId}|${first.conversationId}|${config.version}|${messageIds.join('|')}`
+  return {
+    chunkId: digest(identity),
+    accountId: first.accountId,
+    conversationId: first.conversationId,
+    startTime: first.createTime,
+    endTime: last.createTime,
+    text,
+    messageIds,
+    participantIds,
+    messageKinds,
+    contentHash: digest(`${identity}|${text}`),
+    chunkerVersion: config.version
+  }
+}
+
+/** Chunks one conversation only; cross-conversation chunks are never allowed. */
+export function chunkConversation(
+  messages: KnowledgeNormalizedMessage[],
+  config: KnowledgeChunkerConfig
+): KnowledgeChunk[] {
+  const sorted = messages
+    .filter(isIndexableKnowledgeMessage)
+    .slice()
+    .sort((left, right) => left.createTime - right.createTime || left.messageId.localeCompare(right.messageId))
+  if (!sorted.length) return []
+
+  const conversationId = sorted[0].conversationId
+  const accountId = sorted[0].accountId
+  if (sorted.some((message) => message.conversationId !== conversationId || message.accountId !== accountId)) {
+    throw new Error('Conversation chunker received messages from multiple accounts or conversations')
+  }
+
+  const chunks: KnowledgeChunk[] = []
+  let current: KnowledgeNormalizedMessage[] = []
+  let currentCharacters = 0
+  for (const message of sorted) {
+    const previous = current[current.length - 1]
+    const nextCharacters = currentCharacters + message.searchableText.length
+    const shouldSplit =
+      current.length > 0 &&
+      (message.createTime - previous.createTime > config.maxGapMs ||
+        current.length >= config.maxMessages ||
+        nextCharacters > config.maxCharacters)
+    if (shouldSplit) {
+      chunks.push(buildChunk(current, config))
+      current = []
+      currentCharacters = 0
+    }
+    current.push(message)
+    currentCharacters += message.searchableText.length
+  }
+  if (current.length) chunks.push(buildChunk(current, config))
+  return chunks
+}

src/main/knowledge/knowledge-search-service.ts

@@ -0,0 +1,611 @@
+import * as chat from '../services/chat-service'
+import type {
+  KnowledgeAttachmentMetadata,
+  KnowledgeEvidence,
+  KnowledgeMessageKind,
+  KnowledgeRuntimeStatus,
+  KnowledgeSearchRequest,
+  KnowledgeSearchIpcRequest,
+  KnowledgeSearchIpcResult,
+  KnowledgeSearchResult,
+  KnowledgeSourceMessage
+} from '../../shared/knowledge'
+import {
+  DEFAULT_KNOWLEDGE_CHUNKER,
+  DEFAULT_KNOWLEDGE_FTS_CONFIG,
+  emptyKnowledgeSearchTimings
+} from '../../shared/knowledge'
+import { KnowledgeService } from './knowledge-service'
+
+const FALLBACK_LIMIT = 240
+const MAX_SENDER_NAME_CONVERSATIONS = 8
+const MAX_CONVERSATION_FILTERS_PER_WORKER_SEARCH = 700
+
+function looksLikeOpaqueSenderId(value: string | undefined): boolean {
+  const normalized = value?.trim() || ''
+  return (
+    normalized.startsWith('wxid_') ||
+    normalized.endsWith('@chatroom') ||
+    /^\d{6,}$/.test(normalized)
+  )
+}
+
+function groupMemberDisplayName(member: chat.GroupSnapshot['members'][number]): string {
+  return (
+    [member.groupNickname, member.wechatNickname, member.nickname, member.remark]
+      .map((value) => value.trim())
+      .find((value) => value && !looksLikeOpaqueSenderId(value)) || ''
+  )
+}
+
+function sourceMessageId(message: chat.FormattedMessage): string {
+  if (message.localId) return `local:${message.localId}`
+  if (message.id) return String(message.id)
+  return `${message.createTime || 0}:${message.serverId || message.content}`
+}
+
+function sourceKind(message: chat.FormattedMessage): KnowledgeMessageKind {
+  if (message.voiceTranscript || message.type === '语音') return 'voice'
+  if (message.contentData?.type === 'share' || message.contentData?.type === 'miniProgram') {
+    return message.contentData.type === 'share' && message.contentData.typeVal === '6'
+      ? 'file'
+      : 'link'
+  }
+  if (message.contentData?.type === 'system') return 'system'
+  return message.content?.trim() ? 'text' : 'other'
+}
+
+function sourceTextAndAttachment(message: chat.FormattedMessage): {
+  text?: string
+  attachment?: KnowledgeAttachmentMetadata
+} {
+  const text = message.content?.trim() || ''
+  const content = message.contentData
+  if (!content) {
+    return {
+      text: text || undefined,
+      attachment: message.exportMediaName
+        ? {
+            name: message.exportMediaName,
+            kind: message.exportMediaType === 'file' ? 'file' : 'other'
+          }
+        : undefined
+    }
+  }
+  if (content.type === 'share') {
+    const title = content.title?.trim() || ''
+    const description = content.des?.trim() || ''
+    return {
+      text: [text, title, description].filter(Boolean).join('\n') || undefined,
+      attachment:
+        title || content.url
+          ? {
+              name: title || content.url,
+              kind: content.typeVal === '6' ? 'file' : 'link',
+              url: content.url
+            }
+          : undefined
+    }
+  }
+  if (content.type === 'miniProgram') {
+    return {
+      text: [text, content.title, content.description].filter(Boolean).join('\n') || undefined,
+      attachment: content.title ? { name: content.title, kind: 'link' } : undefined
+    }
+  }
+  if (content.type === 'quote') {
+    return {
+      text:
+        [text, content.title, content.content, content.quotedContent].filter(Boolean).join('\n') ||
+        undefined
+    }
+  }
+  if (content.type === 'forwardBundle') {
+    return {
+      text: [text, content.title, content.description, ...content.items.map((item) => item.text)]
+        .filter(Boolean)
+        .join('\n')
+    }
+  }
+  return { text: text || undefined }
+}
+
+function toSourceMessage(
+  accountId: string,
+  conversationId: string,
+  message: chat.FormattedMessage
+): KnowledgeSourceMessage | null {
+  if (!message.createTime) return null
+  const extracted = sourceTextAndAttachment(message)
+  const voiceTranscript = message.voiceTranscript?.trim() || undefined
+  if (!extracted.text && !extracted.attachment && !voiceTranscript) return null
+  return {
+    accountId,
+    conversationId,
+    messageId: sourceMessageId(message),
+    // Existing chat messages use Unix seconds; the knowledge contract uses milliseconds.
+    createTime: message.createTime * 1000,
+    senderId: message.senderId || message.from || undefined,
+    senderName: message.isSender ? '我' : message.name || undefined,
+    kind: sourceKind(message),
+    text: extracted.text,
+    attachment: extracted.attachment,
+    voiceTranscript
+  }
+}
+
+function normalizeComparable(value: string): string {
+  return value.toLocaleLowerCase().replace(/\s+/g, '')
+}
+
+function fallbackTermScore(message: chat.FormattedMessage, terms: string[]): number {
+  const source = toSourceMessage('fallback', 'fallback', message)
+  const text = `${source?.text || ''}\n${source?.voiceTranscript || ''}\n${source?.attachment?.name || ''}`
+  const normalized = normalizeComparable(text)
+  return terms.reduce((score, term) => {
+    const normalizedTerm = normalizeComparable(term)
+    return normalizedTerm && normalized.includes(normalizedTerm)
+      ? score + normalizedTerm.length
+      : score
+  }, 0)
+}
+
+/**
+ * Main-process adapter for the read-only chat archive. It never passes source
+ * database handles or keys to the worker; only normalized serializable values.
+ */
+export class KnowledgeSearchService {
+  private readonly service: KnowledgeService
+  private readonly indexing = new Map<string, Promise<void>>()
+  private readonly statusByAccount = new Map<string, KnowledgeRuntimeStatus>()
+  private readonly statusListeners = new Set<(status: KnowledgeRuntimeStatus) => void>()
+  private wcdbReadTail: Promise<void> = Promise.resolve()
+
+  constructor(userDataPath: string, workerPath: string) {
+    this.service = new KnowledgeService(userDataPath, workerPath)
+  }
+
+  startCurrentAccountIndex(): KnowledgeRuntimeStatus {
+    const accountId = this.currentAccountId()
+    if (!accountId) return this.emptyStatus('')
+    const current = this.statusByAccount.get(accountId) || this.emptyStatus(accountId)
+    if (this.indexing.has(accountId)) return current
+    const started: KnowledgeRuntimeStatus = {
+      ...current,
+      state: current.indexedMessageCount ? 'syncing' : 'building',
+      processedMessages: 0,
+      totalMessages: current.sourceMessageCount,
+      estimatedRemainingMs: null,
+      lastError: undefined
+    }
+    this.publishStatus(started)
+    const task = this.indexAccount(accountId)
+      .catch((error) => {
+        const previous = this.statusByAccount.get(accountId)
+        this.publishStatus({
+          ...(previous || this.emptyStatus(accountId)),
+          state: 'error',
+          lastError: error instanceof Error ? error.message : String(error)
+        })
+        throw error
+      })
+      .finally(() => {
+        this.indexing.delete(accountId)
+        void this.refreshStatus(accountId).catch(() => undefined)
+      })
+    this.indexing.set(accountId, task)
+    void task.catch((error) => {
+      console.warn('[Knowledge] background index failed:', error)
+    })
+    return started
+  }
+
+  async search(request: KnowledgeSearchIpcRequest): Promise<KnowledgeSearchIpcResult> {
+    const accountId = this.currentAccountId()
+    if (!accountId) return this.searchFallback(request, 'unavailable')
+    try {
+      const searchRequest: Omit<KnowledgeSearchRequest, 'databaseRoot'> = {
+        accountId,
+        fts: DEFAULT_KNOWLEDGE_FTS_CONFIG,
+        text: request.text,
+        terms: request.terms,
+        limit: Math.max(1, Math.min(request.limit || FALLBACK_LIMIT, FALLBACK_LIMIT)),
+        conversationIds: request.conversationIds,
+        senderIds: request.senderIds,
+        startTime: request.startTime === undefined ? undefined : request.startTime * 1000,
+        endTime: request.endTime === undefined ? undefined : request.endTime * 1000
+      }
+      const result = await this.searchKnowledge(searchRequest)
+      // An existing derived database can answer while its next incremental pass is running.
+      // Never turn an interactive global search into another full WCDB scan during that pass.
+      if (result.state === 'ready' || result.evidence.length) {
+        return this.toKnowledgeResult(result)
+      }
+      if (this.indexing.has(accountId)) {
+        return {
+          ...result,
+          source: 'knowledge',
+          totalMessages: result.indexedMessageCount
+        }
+      }
+      return this.searchFallback(request, 'unavailable')
+    } catch (error) {
+      console.warn('[Knowledge] search failed, using legacy fallback:', error)
+      return this.searchFallback(request, 'error')
+    }
+  }
+
+  async dispose(): Promise<void> {
+    await this.service.dispose()
+  }
+
+  /** Safely release derived SQLite handles before the cache screen removes them. */
+  async prepareForCacheClear(): Promise<void> {
+    if (this.indexing.size) {
+      throw new Error('本地知识库正在同步，请等待同步完成后再清理')
+    }
+    await this.service.dispose()
+    const accountIds = Array.from(this.statusByAccount.keys())
+    this.statusByAccount.clear()
+    accountIds.forEach((accountId) => this.publishStatus(this.emptyStatus(accountId)))
+  }
+
+  onStatusChange(listener: (status: KnowledgeRuntimeStatus) => void): () => void {
+    this.statusListeners.add(listener)
+    return () => this.statusListeners.delete(listener)
+  }
+
+  async getStatus(): Promise<KnowledgeRuntimeStatus> {
+    const accountId = this.currentAccountId()
+    if (!accountId) return this.emptyStatus('')
+    return this.refreshStatus(accountId)
+  }
+
+  private currentAccountId(): string {
+    if (!chat.isReady()) return ''
+    return chat.getSelfAccountInfo()?.wxid || chat.getCurrentAccountRoot()
+  }
+
+  private async indexAccount(accountId: string): Promise<void> {
+    const contacts = await this.listContacts()
+    let processedMessages = 0
+    const startedAt = Date.now()
+    this.publishStatus({
+      ...(this.statusByAccount.get(accountId) || this.emptyStatus(accountId)),
+      state: this.statusByAccount.get(accountId)?.indexedMessageCount ? 'syncing' : 'building',
+      processedMessages: 0,
+      totalMessages: null,
+      estimatedRemainingMs: null
+    })
+    for (const [index, contact] of contacts.entries()) {
+      // WCDB rejects overlapping async pagination. Queue every archive read so
+      // background indexing and an interactive fallback search can interleave safely.
+      const messages = await this.listMessages(contact.md5)
+      const sourceMessages = messages
+        .map((message) => toSourceMessage(accountId, contact.md5, message))
+        .filter((message): message is KnowledgeSourceMessage => Boolean(message))
+      await this.service.index(
+        {
+          accountId,
+          conversations: [
+            {
+              conversationId: contact.md5,
+              completeSnapshot: true,
+              messages: sourceMessages
+            }
+          ],
+          chunker: DEFAULT_KNOWLEDGE_CHUNKER,
+          fts: DEFAULT_KNOWLEDGE_FTS_CONFIG,
+          sourceMessageCount:
+            index === contacts.length - 1 ? processedMessages + sourceMessages.length : undefined
+        },
+        (progress) => {
+          const current = this.statusByAccount.get(accountId) || this.emptyStatus(accountId)
+          this.publishStatus({
+            ...current,
+            state: current.indexedMessageCount ? 'syncing' : 'building',
+            processedMessages: processedMessages + progress.processedMessages,
+            totalMessages: null,
+            currentConversationId: progress.conversationId,
+            estimatedRemainingMs: null
+          })
+        }
+      )
+      processedMessages += sourceMessages.length
+      const current = this.statusByAccount.get(accountId) || this.emptyStatus(accountId)
+      this.publishStatus({
+        ...current,
+        state: current.indexedMessageCount ? 'syncing' : 'building',
+        processedMessages,
+        totalMessages: null,
+        currentConversationId: contact.md5,
+        estimatedRemainingMs: null
+      })
+    }
+    await this.refreshStatus(accountId, {
+      processedMessages,
+      totalMessages: processedMessages,
+      startedAt
+    })
+  }
+
+  private async searchFallback(
+    request: KnowledgeSearchIpcRequest,
+    fallbackReason: 'unavailable' | 'indexing' | 'error'
+  ): Promise<KnowledgeSearchIpcResult> {
+    const startedAt = Date.now()
+    const contacts = await this.listContacts()
+    const allowedConversations = new Set(request.conversationIds || [])
+    const sourceContacts = allowedConversations.size
+      ? contacts.filter((contact) => allowedConversations.has(contact.md5))
+      : contacts
+    const senderIds = new Set(request.senderIds || [])
+    const terms = request.terms.filter((term) => term.trim().length >= 2)
+    const matches: Array<{
+      contact: (typeof sourceContacts)[number]
+      message: chat.FormattedMessage
+      score: number
+    }> = []
+    let totalMessages = 0
+
+    for (const contact of sourceContacts) {
+      const messages = await this.listMessages(contact.md5, request.startTime, request.endTime)
+      totalMessages += messages.length
+      for (const message of messages) {
+        matches.push({
+          contact,
+          message,
+          score: fallbackTermScore(message, terms)
+        })
+      }
+    }
+    const filtered = matches
+      .filter(({ message, score }) => {
+        const senderMatches = !senderIds.size || senderIds.has(message.senderId || message.from)
+        const termMatches = !terms.length || score > 0
+        return senderMatches && termMatches
+      })
+      .sort(
+        (left, right) =>
+          right.score - left.score ||
+          (right.message.createTime || 0) - (left.message.createTime || 0)
+      )
+      .slice(0, Math.max(1, Math.min(request.limit || FALLBACK_LIMIT, FALLBACK_LIMIT)))
+    const result: KnowledgeSearchIpcResult = {
+      source: 'fallback',
+      fallbackReason,
+      state: fallbackReason === 'indexing' ? 'indexing' : 'unavailable',
+      indexedMessageCount: 0,
+      indexedChunkCount: 0,
+      totalMessages,
+      timings: {
+        ...emptyKnowledgeSearchTimings(),
+        messageLoadMs: Date.now() - startedAt,
+        totalMs: Date.now() - startedAt
+      },
+      evidence: filtered.map(({ contact, message, score }) => ({
+        chunkId: `fallback:${contact.md5}:${sourceMessageId(message)}`,
+        conversationId: contact.md5,
+        startTime: (message.createTime || 0) * 1000,
+        endTime: (message.createTime || 0) * 1000,
+        messageId: sourceMessageId(message),
+        senderId: message.senderId || message.from || undefined,
+        sender: message.isSender ? '我' : message.name || '未知成员',
+        timestamp: (message.createTime || 0) * 1000,
+        messageIds: [sourceMessageId(message)],
+        text: sourceTextAndAttachment(message).text || message.content || `[${message.type}]`,
+        score: -score
+      }))
+    }
+    return {
+      ...result,
+      evidence: await this.enrichEvidenceSenders(result.evidence)
+    }
+  }
+
+  /**
+   * SQLite has a finite bind-parameter limit. Group/one-to-one scope filters
+   * can contain over one thousand conversations, so split only the Worker
+   * query and merge real Evidence instead of dropping the selected scope.
+   */
+  private async searchKnowledge(
+    request: Omit<KnowledgeSearchRequest, 'databaseRoot'>
+  ): Promise<KnowledgeSearchResult> {
+    const conversationIds = Array.from(new Set(request.conversationIds || []))
+    if (conversationIds.length <= MAX_CONVERSATION_FILTERS_PER_WORKER_SEARCH) {
+      return this.searchWorker(request)
+    }
+    const partialResults: KnowledgeSearchResult[] = []
+    for (
+      let start = 0;
+      start < conversationIds.length;
+      start += MAX_CONVERSATION_FILTERS_PER_WORKER_SEARCH
+    ) {
+      partialResults.push(
+        await this.searchWorker({
+          ...request,
+          conversationIds: conversationIds.slice(
+            start,
+            start + MAX_CONVERSATION_FILTERS_PER_WORKER_SEARCH
+          )
+        })
+      )
+    }
+    const evidenceByIdentity = new Map<string, KnowledgeEvidence>()
+    partialResults
+      .flatMap((result) => result.evidence)
+      .forEach((item) => {
+        const identity = `${item.conversationId}:${item.messageId}`
+        const existing = evidenceByIdentity.get(identity)
+        if (!existing || (item.score || 0) < (existing.score || 0)) {
+          evidenceByIdentity.set(identity, item)
+        }
+      })
+    const mergeStartedAt = Date.now()
+    const mergedEvidence = Array.from(evidenceByIdentity.values())
+      .sort(
+        (left, right) => (left.score || 0) - (right.score || 0) || right.timestamp - left.timestamp
+      )
+      .slice(0, request.limit)
+    const timings = partialResults.reduce(
+      (total, result) => ({
+        workerIpcMs: total.workerIpcMs + (result.timings?.workerIpcMs || 0),
+        workerBootMs: total.workerBootMs + (result.timings?.workerBootMs || 0),
+        dispatchMs: total.dispatchMs + (result.timings?.dispatchMs || 0),
+        workerSqlMs: total.workerSqlMs + (result.timings?.workerSqlMs || 0),
+        responseTransferMs: total.responseTransferMs + (result.timings?.responseTransferMs || 0),
+        responseSerializeMs: total.responseSerializeMs + (result.timings?.responseSerializeMs || 0),
+        ftsMs: total.ftsMs + (result.timings?.ftsMs || 0),
+        messageLoadMs: total.messageLoadMs + (result.timings?.messageLoadMs || 0),
+        chunkExpandMs: total.chunkExpandMs + (result.timings?.chunkExpandMs || 0),
+        rankingMs: total.rankingMs + (result.timings?.rankingMs || 0),
+        totalMs: total.totalMs + (result.timings?.totalMs || 0)
+      }),
+      emptyKnowledgeSearchTimings()
+    )
+    const mergeRankingMs = Date.now() - mergeStartedAt
+    timings.rankingMs += mergeRankingMs
+    timings.totalMs += mergeRankingMs
+    return {
+      state: partialResults.some((result) => result.state === 'ready')
+        ? 'ready'
+        : partialResults.some((result) => result.state === 'indexing')
+          ? 'indexing'
+          : 'unavailable',
+      indexedMessageCount: Math.max(...partialResults.map((result) => result.indexedMessageCount)),
+      indexedChunkCount: Math.max(...partialResults.map((result) => result.indexedChunkCount)),
+      evidence: mergedEvidence,
+      timings
+    }
+  }
+
+  private async searchWorker(
+    request: Omit<KnowledgeSearchRequest, 'databaseRoot'>
+  ): Promise<KnowledgeSearchResult> {
+    const startedAt = Date.now()
+    const result = await this.service.search(request)
+    const timings = result.timings || emptyKnowledgeSearchTimings()
+    return {
+      ...result,
+      timings: {
+        ...timings,
+        workerIpcMs: timings.workerIpcMs || Math.max(0, Date.now() - startedAt - timings.totalMs),
+        workerSqlMs: timings.workerSqlMs || timings.totalMs
+      }
+    }
+  }
+
+  private listContacts(): ReturnType<typeof chat.listContactsAsync> {
+    return this.enqueueWcdbRead(() => chat.listContactsAsync())
+  }
+
+  private listMessages(
+    conversationId: string,
+    startTime?: number,
+    endTime?: number
+  ): ReturnType<typeof chat.listMessagesAsync> {
+    return this.enqueueWcdbRead(() => chat.listMessagesAsync(conversationId, startTime, endTime))
+  }
+
+  private async toKnowledgeResult(
+    result: KnowledgeSearchResult
+  ): Promise<KnowledgeSearchIpcResult> {
+    return {
+      ...result,
+      evidence: await this.enrichEvidenceSenders(result.evidence),
+      source: 'knowledge',
+      totalMessages: result.indexedMessageCount
+    }
+  }
+
+  private async enrichEvidenceSenders(evidence: KnowledgeEvidence[]): Promise<KnowledgeEvidence[]> {
+    const candidateConversationIds = Array.from(
+      new Set(
+        evidence
+          .filter((item) => item.senderId && looksLikeOpaqueSenderId(item.sender))
+          .map((item) => item.conversationId)
+      )
+    ).slice(0, MAX_SENDER_NAME_CONVERSATIONS)
+    if (!candidateConversationIds.length) return evidence
+
+    const contacts = await this.listContacts()
+    const groupConversationIds = new Set(
+      contacts.filter((contact) => contact.type === 'group').map((contact) => contact.md5)
+    )
+    const memberNamesByConversation = new Map<string, Map<string, string>>()
+    for (const conversationId of candidateConversationIds) {
+      if (!groupConversationIds.has(conversationId)) continue
+      const snapshot = await this.enqueueWcdbRead(() => chat.getGroupSnapshotAsync(conversationId))
+      const memberNames = new Map(
+        (snapshot?.members || [])
+          .map((member) => [member.wxid, groupMemberDisplayName(member)] as const)
+          .filter(([, name]) => Boolean(name))
+      )
+      if (memberNames.size) memberNamesByConversation.set(conversationId, memberNames)
+    }
+
+    return evidence.map((item) => {
+      const sender = memberNamesByConversation.get(item.conversationId)?.get(item.senderId || '')
+      return sender ? { ...item, sender } : item
+    })
+  }
+
+  private enqueueWcdbRead<T>(operation: () => Promise<T>): Promise<T> {
+    const result = this.wcdbReadTail.then(operation, operation)
+    // Keep the queue usable after a read failure while returning that failure to its caller.
+    this.wcdbReadTail = result.then(
+      () => undefined,
+      () => undefined
+    )
+    return result
+  }
+
+  private emptyStatus(accountId: string): KnowledgeRuntimeStatus {
+    return {
+      accountId,
+      state: 'unavailable',
+      indexedMessageCount: 0,
+      indexedChunkCount: 0,
+      sourceMessageCount: null,
+      processedMessages: 0,
+      totalMessages: null,
+      estimatedRemainingMs: null,
+      databaseBytes: 0,
+      walBytes: 0,
+      shmBytes: 0
+    }
+  }
+
+  private async refreshStatus(
+    accountId: string,
+    progress?: Pick<KnowledgeRuntimeStatus, 'processedMessages' | 'totalMessages'> & {
+      startedAt?: number
+    }
+  ): Promise<KnowledgeRuntimeStatus> {
+    const remote = await this.service.status({ accountId, fts: DEFAULT_KNOWLEDGE_FTS_CONFIG })
+    const current = this.statusByAccount.get(accountId)
+    const indexing = this.indexing.has(accountId)
+    const processedMessages =
+      progress?.processedMessages ?? current?.processedMessages ?? remote.processedMessages
+    const totalMessages = progress?.totalMessages ?? remote.sourceMessageCount
+    const state = indexing
+      ? remote.indexedMessageCount > 0
+        ? 'syncing'
+        : 'building'
+      : remote.state
+    const status: KnowledgeRuntimeStatus = {
+      ...remote,
+      state,
+      processedMessages,
+      totalMessages,
+      estimatedRemainingMs: null
+    }
+    this.publishStatus(status)
+    return status
+  }
+
+  private publishStatus(status: KnowledgeRuntimeStatus): void {
+    this.statusByAccount.set(status.accountId, status)
+    for (const listener of this.statusListeners) listener(status)
+  }
+}

src/main/knowledge/knowledge-service.ts

@@ -0,0 +1,54 @@
+import { join } from 'path'
+import type {
+  KnowledgeCapacityPreflight,
+  KnowledgeCapacityPreflightRequest,
+  KnowledgeIndexProgress,
+  KnowledgeIndexRequest,
+  KnowledgeIndexResult,
+  KnowledgeRuntimeStatus,
+  KnowledgeSearchRequest,
+  KnowledgeSearchResult,
+  KnowledgeStatusRequest
+} from '../../shared/knowledge'
+import { KnowledgeWorkerHost } from './knowledge-worker-host'
+
+/** Minimal main-process service; no renderer API is exposed in Task 0～Task 2. */
+export class KnowledgeService {
+  private readonly worker: KnowledgeWorkerHost
+
+  constructor(userDataPath: string, workerPath: string) {
+    this.worker = new KnowledgeWorkerHost(workerPath)
+    this.databaseRoot = join(userDataPath, 'knowledge')
+  }
+
+  private readonly databaseRoot: string
+
+  index(
+    request: Omit<KnowledgeIndexRequest, 'databaseRoot'>,
+    onProgress?: (progress: KnowledgeIndexProgress) => void
+  ): Promise<KnowledgeIndexResult> {
+    return this.worker.index({ ...request, databaseRoot: this.databaseRoot }, onProgress)
+  }
+
+  preflight(
+    request: Omit<KnowledgeCapacityPreflightRequest, 'databaseRoot'>
+  ): Promise<KnowledgeCapacityPreflight> {
+    return this.worker.preflight({ ...request, databaseRoot: this.databaseRoot })
+  }
+
+  remove(accountId: string): Promise<{ removed: true }> {
+    return this.worker.remove(accountId, this.databaseRoot)
+  }
+
+  search(request: Omit<KnowledgeSearchRequest, 'databaseRoot'>): Promise<KnowledgeSearchResult> {
+    return this.worker.search({ ...request, databaseRoot: this.databaseRoot })
+  }
+
+  status(request: Omit<KnowledgeStatusRequest, 'databaseRoot'>): Promise<KnowledgeRuntimeStatus> {
+    return this.worker.status({ ...request, databaseRoot: this.databaseRoot })
+  }
+
+  dispose(): Promise<void> {
+    return this.worker.dispose()
+  }
+}

src/main/knowledge/knowledge-worker-host.ts

@@ -0,0 +1,182 @@
+import { fork, type ChildProcess } from 'child_process'
+import { randomUUID } from 'crypto'
+import type {
+  KnowledgeCapacityPreflight,
+  KnowledgeCapacityPreflightRequest,
+  KnowledgeIndexProgress,
+  KnowledgeIndexRequest,
+  KnowledgeIndexResult,
+  KnowledgeRuntimeStatus,
+  KnowledgeSearchRequest,
+  KnowledgeSearchResult,
+  KnowledgeStatusRequest,
+  KnowledgeWorkerRequest,
+  KnowledgeWorkerResponse
+} from '../../shared/knowledge'
+
+type WorkerResult =
+  | KnowledgeIndexResult
+  | KnowledgeCapacityPreflight
+  | KnowledgeSearchResult
+  | KnowledgeRuntimeStatus
+  | { removed: true }
+type PendingRequest = {
+  resolve: (result: WorkerResult) => void
+  reject: (error: Error) => void
+  onProgress?: (progress: KnowledgeIndexProgress) => void
+  sentAt: number
+  workerBootStartedAt?: number
+}
+
+/**
+ * Main-process boundary for the derived knowledge database. The child runs
+ * with ELECTRON_RUN_AS_NODE so synchronous node:sqlite calls never block UI.
+ */
+export class KnowledgeWorkerHost {
+  private child: ChildProcess | null = null
+  private childStartedAt = 0
+  private readonly pending = new Map<string, PendingRequest>()
+
+  constructor(private readonly workerPath: string) {}
+
+  index(
+    payload: KnowledgeIndexRequest,
+    onProgress?: (progress: KnowledgeIndexProgress) => void
+  ): Promise<KnowledgeIndexResult> {
+    return this.request('index', payload, onProgress) as Promise<KnowledgeIndexResult>
+  }
+
+  preflight(payload: KnowledgeCapacityPreflightRequest): Promise<KnowledgeCapacityPreflight> {
+    return this.request('preflight', payload) as Promise<KnowledgeCapacityPreflight>
+  }
+
+  search(payload: KnowledgeSearchRequest): Promise<KnowledgeSearchResult> {
+    return this.request('search', payload) as Promise<KnowledgeSearchResult>
+  }
+
+  status(payload: KnowledgeStatusRequest): Promise<KnowledgeRuntimeStatus> {
+    return this.request('status', payload) as Promise<KnowledgeRuntimeStatus>
+  }
+
+  remove(accountId: string, databaseRoot: string): Promise<{ removed: true }> {
+    return this.request('remove', { accountId, databaseRoot }) as Promise<{ removed: true }>
+  }
+
+  cancel(targetRequestId: string): Promise<{ removed: true }> {
+    return this.request('cancel', { targetRequestId }) as Promise<{ removed: true }>
+  }
+
+  async dispose(): Promise<void> {
+    const child = this.child
+    if (!child) return
+    try {
+      await this.request('close', {})
+    } catch {
+      // The child is about to be stopped; its only job is a derived local index.
+    }
+    if (this.child === child) this.child = null
+    if (!child.killed) child.kill()
+  }
+
+  private request(
+    type: KnowledgeWorkerRequest['type'],
+    payload: KnowledgeWorkerRequest['payload'],
+    onProgress?: (progress: KnowledgeIndexProgress) => void
+  ): Promise<WorkerResult> {
+    const hadWorker = Boolean(this.child?.connected)
+    const child = this.ensureChild()
+    const requestId = randomUUID()
+    const sentAt = Date.now()
+    const request: KnowledgeWorkerRequest = { version: 1, type, requestId, sentAt, payload }
+    return new Promise((resolve, reject) => {
+      this.pending.set(requestId, {
+        resolve,
+        reject,
+        onProgress,
+        sentAt,
+        workerBootStartedAt: hadWorker ? undefined : this.childStartedAt
+      })
+      child.send(request, (error) => {
+        if (error) this.finish(requestId, undefined, error)
+      })
+    })
+  }
+
+  private ensureChild(): ChildProcess {
+    if (this.child?.connected) return this.child
+    const child = fork(this.workerPath, [], {
+      stdio: ['ignore', 'ignore', 'ignore', 'ipc'],
+      serialization: 'advanced',
+      env: { ...process.env, ELECTRON_RUN_AS_NODE: '1' }
+    })
+    child.on('message', (message: KnowledgeWorkerResponse) => {
+      if (message?.version !== 1) return
+      if (message.type === 'progress') {
+        const pending = this.pending.get(message.requestId)
+        if (pending && message.payload)
+          pending.onProgress?.(message.payload as KnowledgeIndexProgress)
+        return
+      }
+      this.finish(
+        message.requestId,
+        message.payload as WorkerResult | undefined,
+        message.type === 'error'
+          ? new Error(message.error || 'Knowledge worker failed')
+          : undefined,
+        message.transport
+      )
+    })
+    child.once('error', (error) => this.failAll(error))
+    child.once('exit', (code) => {
+      if (this.child === child) this.child = null
+      this.failAll(new Error(`Knowledge worker exited (${code ?? 'unknown'})`))
+    })
+    this.child = child
+    this.childStartedAt = Date.now()
+    return child
+  }
+
+  private finish(
+    requestId: string,
+    result?: WorkerResult,
+    error?: Error,
+    transport?: KnowledgeWorkerResponse['transport']
+  ): void {
+    const pending = this.pending.get(requestId)
+    if (!pending) return
+    this.pending.delete(requestId)
+    if (error) pending.reject(error)
+    else if (result) pending.resolve(this.applyTransportTimings(result, pending, transport))
+    else pending.reject(new Error('Knowledge worker returned no result'))
+  }
+
+  private applyTransportTimings(
+    result: WorkerResult,
+    pending: PendingRequest,
+    transport?: KnowledgeWorkerResponse['transport']
+  ): WorkerResult {
+    if (!('timings' in result) || !transport) return result
+    const receivedAt = Date.now()
+    const workerBootMs = pending.workerBootStartedAt
+      ? Math.max(0, transport.workerReceivedAt - pending.workerBootStartedAt)
+      : 0
+    const dispatchMs = Math.max(0, transport.workerReceivedAt - pending.sentAt)
+    const responseTransferMs = Math.max(0, receivedAt - transport.workerCompletedAt)
+    return {
+      ...result,
+      timings: {
+        ...result.timings,
+        workerBootMs,
+        dispatchMs,
+        workerSqlMs: result.timings.totalMs,
+        responseSerializeMs: transport.responseSerializeMs,
+        responseTransferMs,
+        workerIpcMs: workerBootMs + dispatchMs + responseTransferMs
+      }
+    }
+  }
+
+  private failAll(error: Error): void {
+    for (const requestId of this.pending.keys()) this.finish(requestId, undefined, error)
+  }
+}

src/main/knowledge/knowledge-worker.ts

@@ -0,0 +1,200 @@
+import type {
+  KnowledgeCapacityPreflightRequest,
+  KnowledgeIndexRequest,
+  KnowledgeRuntimeStatus,
+  KnowledgeSearchRequest,
+  KnowledgeStatusRequest,
+  KnowledgeWorkerRequest,
+  KnowledgeWorkerResponse
+} from '../../shared/knowledge'
+import { emptyKnowledgeSearchTimings } from '../../shared/knowledge'
+import {
+  KnowledgeStore,
+  estimateKnowledgeCapacityPreflight,
+  getKnowledgeDatabasePath,
+  removeKnowledgeDatabase
+} from './knowledge-store'
+import { existsSync } from 'fs'
+import { serialize } from 'v8'
+
+const stores = new Map<string, KnowledgeStore>()
+const controllers = new Map<string, AbortController>()
+
+function send(
+  message: KnowledgeWorkerResponse,
+  transport?: KnowledgeWorkerResponse['transport']
+): void {
+  if (process.send) process.send({ ...message, transport })
+}
+
+function sendSearchResult(
+  request: KnowledgeWorkerRequest,
+  payload: KnowledgeWorkerResponse['payload'],
+  workerReceivedAt: number
+): void {
+  const serializeStartedAt = Date.now()
+  // This measures the actual payload encoding workload before Node IPC performs
+  // its own transfer. It lets diagnostics separate payload cost from SQL time.
+  serialize(payload)
+  const responseSerializeMs = Date.now() - serializeStartedAt
+  send(
+    { version: 1, type: 'result', requestId: request.requestId, payload },
+    { workerReceivedAt, workerCompletedAt: Date.now(), responseSerializeMs }
+  )
+}
+
+function storeKey(databaseRoot: string, accountId: string): string {
+  return getKnowledgeDatabasePath(databaseRoot, accountId)
+}
+
+function getStore(
+  request: Pick<KnowledgeIndexRequest, 'databaseRoot' | 'accountId' | 'fts'>
+): KnowledgeStore {
+  const key = storeKey(request.databaseRoot, request.accountId)
+  let store = stores.get(key)
+  if (!store) {
+    store = new KnowledgeStore(request.databaseRoot, request.accountId, request.fts)
+    stores.set(key, store)
+  }
+  return store
+}
+
+function closeStore(databaseRoot: string, accountId: string): void {
+  const key = storeKey(databaseRoot, accountId)
+  const store = stores.get(key)
+  if (store) store.close()
+  stores.delete(key)
+}
+
+async function handleIndex(
+  request: KnowledgeWorkerRequest,
+  payload: KnowledgeIndexRequest
+): Promise<void> {
+  const controller = new AbortController()
+  controllers.set(request.requestId, controller)
+  try {
+    const result = await getStore(payload).index(payload, controller.signal, (progress) => {
+      send({ version: 1, type: 'progress', requestId: request.requestId, payload: progress })
+    })
+    send({ version: 1, type: 'result', requestId: request.requestId, payload: result })
+  } finally {
+    controllers.delete(request.requestId)
+  }
+}
+
+async function handlePreflight(
+  request: KnowledgeWorkerRequest,
+  payload: KnowledgeCapacityPreflightRequest
+): Promise<void> {
+  const result = await estimateKnowledgeCapacityPreflight(payload)
+  send({ version: 1, type: 'result', requestId: request.requestId, payload: result })
+}
+
+async function handleSearch(
+  request: KnowledgeWorkerRequest,
+  payload: KnowledgeSearchRequest
+): Promise<void> {
+  const workerReceivedAt = Date.now()
+  const path = getKnowledgeDatabasePath(payload.databaseRoot, payload.accountId)
+  if (!existsSync(path)) {
+    sendSearchResult(
+      request,
+      {
+        state: 'unavailable',
+        evidence: [],
+        indexedMessageCount: 0,
+        indexedChunkCount: 0,
+        timings: emptyKnowledgeSearchTimings()
+      },
+      workerReceivedAt
+    )
+    return
+  }
+  const result = getStore(payload).searchWithStatus(payload)
+  sendSearchResult(request, result, workerReceivedAt)
+}
+
+async function handleStatus(
+  request: KnowledgeWorkerRequest,
+  payload: KnowledgeStatusRequest
+): Promise<void> {
+  const path = getKnowledgeDatabasePath(payload.databaseRoot, payload.accountId)
+  if (!existsSync(path)) {
+    const unavailable: KnowledgeRuntimeStatus = {
+      accountId: payload.accountId,
+      state: 'unavailable',
+      indexedMessageCount: 0,
+      indexedChunkCount: 0,
+      sourceMessageCount: null,
+      processedMessages: 0,
+      totalMessages: null,
+      estimatedRemainingMs: null,
+      databaseBytes: 0,
+      walBytes: 0,
+      shmBytes: 0
+    }
+    send({ version: 1, type: 'result', requestId: request.requestId, payload: unavailable })
+    return
+  }
+  send({
+    version: 1,
+    type: 'result',
+    requestId: request.requestId,
+    payload: getStore(payload).getRuntimeStatus()
+  })
+}
+
+async function handle(request: KnowledgeWorkerRequest): Promise<void> {
+  try {
+    if (request.type === 'cancel') {
+      const payload = request.payload as { targetRequestId: string }
+      controllers.get(payload.targetRequestId)?.abort()
+      send({ version: 1, type: 'result', requestId: request.requestId, payload: { removed: true } })
+      return
+    }
+    if (request.type === 'close') {
+      for (const controller of controllers.values()) controller.abort()
+      for (const store of stores.values()) store.close()
+      stores.clear()
+      send({ version: 1, type: 'result', requestId: request.requestId, payload: { removed: true } })
+      process.disconnect?.()
+      return
+    }
+    if (request.type === 'remove') {
+      const payload = request.payload as { accountId: string; databaseRoot: string }
+      closeStore(payload.databaseRoot, payload.accountId)
+      removeKnowledgeDatabase(payload.databaseRoot, payload.accountId)
+      send({ version: 1, type: 'result', requestId: request.requestId, payload: { removed: true } })
+      return
+    }
+    if (request.type === 'preflight') {
+      await handlePreflight(request, request.payload as KnowledgeCapacityPreflightRequest)
+      return
+    }
+    if (request.type === 'search') {
+      await handleSearch(request, request.payload as KnowledgeSearchRequest)
+      return
+    }
+    if (request.type === 'status') {
+      await handleStatus(request, request.payload as KnowledgeStatusRequest)
+      return
+    }
+    if (request.type === 'index') {
+      await handleIndex(request, request.payload as KnowledgeIndexRequest)
+      return
+    }
+    throw new Error(`Unsupported knowledge worker request: ${String(request.type)}`)
+  } catch (error) {
+    send({
+      version: 1,
+      type: 'error',
+      requestId: request.requestId,
+      error: error instanceof Error ? error.message : String(error)
+    })
+  }
+}
+
+process.on('message', (message: KnowledgeWorkerRequest) => {
+  if (message?.version !== 1) return
+  void handle(message)
+})

src/main/knowledge/normalizer.ts

@@ -0,0 +1,55 @@
+import { createHash } from 'crypto'
+import type {
+  KnowledgeNormalizedMessage,
+  KnowledgeSourceMessage
+} from '../../shared/knowledge'
+
+const compact = (value: string | undefined): string => value?.replace(/\s+/g, ' ').trim() || ''
+
+function digest(value: string): string {
+  return createHash('sha256').update(value).digest('hex')
+}
+
+/**
+ * Converts a read-only archive record into text safe for local search. Paths,
+ * binary media and raw voice data are deliberately excluded.
+ */
+export function normalizeKnowledgeMessage(
+  source: KnowledgeSourceMessage
+): KnowledgeNormalizedMessage {
+  const sections: string[] = []
+  const messageText = compact(source.text)
+  if (messageText) sections.push(messageText)
+
+  const transcript = compact(source.voiceTranscript)
+  if (transcript) sections.push(`语音转写：${transcript}`)
+
+  const attachmentName = compact(source.attachment?.name)
+  if (attachmentName) {
+    const label = source.attachment?.kind === 'link' ? '链接' : '附件'
+    sections.push(`${label}：${attachmentName}`)
+  }
+  const url = compact(source.attachment?.url)
+  if (url) sections.push(`地址：${url}`)
+
+  const searchableText = sections.join('\n')
+  return {
+    ...source,
+    text: messageText || undefined,
+    voiceTranscript: transcript || undefined,
+    searchableText,
+    contentHash: digest(
+      JSON.stringify({
+        messageId: source.messageId,
+        createTime: source.createTime,
+        senderId: source.senderId || '',
+        kind: source.kind,
+        searchableText
+      })
+    )
+  }
+}
+
+export function isIndexableKnowledgeMessage(message: KnowledgeNormalizedMessage): boolean {
+  return Boolean(message.searchableText.trim())
+}

src/main/knowledge/worker-protocol.ts

@@ -0,0 +1,6 @@
+import type { KnowledgeWorkerRequest, KnowledgeWorkerResponse } from '../../shared/knowledge'
+
+export const KNOWLEDGE_WORKER_PROTOCOL_VERSION = 1 as const
+
+export type WorkerKnowledgeRequest = KnowledgeWorkerRequest
+export type WorkerKnowledgeResponse = KnowledgeWorkerResponse

src/main/services/ai-search-agent.ts

@@ -0,0 +1,185 @@
+import type { AiSearchAgentToolName, AiSearchAgentTraceItem } from '../../shared/ai-search'
+
+export const MAX_AGENT_TOOL_CALLS = 5
+
+export type AgentAction =
+  | { action: 'tool'; tool: AiSearchAgentToolName; arguments: Record<string, unknown> }
+  | { action: 'finalize'; reason: string }
+
+export interface AgentToolResult {
+  summary: Record<string, unknown>
+  candidateCount: number
+  /** A host-owned coverage signal, never supplied by the model. */
+  finalizeReason?: string
+}
+
+export interface ControlledSearchAgentOptions {
+  question: string
+  scopeLabel: string
+  rangeLabel: string
+  maxToolCalls?: number
+  decide: (prompt: string) => Promise<string | undefined>
+  execute: (action: Extract<AgentAction, { action: 'tool' }>) => Promise<AgentToolResult>
+  onTrace: (item: Omit<AiSearchAgentTraceItem, 'sequence'>) => void
+}
+
+export interface ControlledSearchAgentResult {
+  status: 'finalized' | 'exhausted' | 'invalid'
+  toolCalls: number
+  reason: string
+}
+
+const TOOL_NAMES = new Set<AiSearchAgentToolName>([
+  'search_conversations',
+  'search_people',
+  'search_messages',
+  'get_conversation_messages',
+  'get_messages_by_time',
+  'get_message_context'
+])
+
+const parseAction = (value: string | undefined): AgentAction | null => {
+  if (!value) return null
+  const match = value.match(/\{[\s\S]*\}/)
+  if (!match) return null
+  try {
+    const parsed = JSON.parse(match[0]) as Record<string, unknown>
+    if (parsed.action === 'finalize' && typeof parsed.reason === 'string' && parsed.reason.trim()) {
+      return { action: 'finalize', reason: parsed.reason.trim().slice(0, 240) }
+    }
+    if (
+      parsed.action === 'tool' &&
+      typeof parsed.tool === 'string' &&
+      TOOL_NAMES.has(parsed.tool as AiSearchAgentToolName) &&
+      parsed.arguments &&
+      typeof parsed.arguments === 'object' &&
+      !Array.isArray(parsed.arguments)
+    ) {
+      return {
+        action: 'tool',
+        tool: parsed.tool as AiSearchAgentToolName,
+        arguments: parsed.arguments as Record<string, unknown>
+      }
+    }
+  } catch {
+    // Invalid model output is rejected by the caller and triggers legacy fallback.
+  }
+  return null
+}
+
+const agentSystemPrompt = (
+  question: string,
+  scopeLabel: string,
+  rangeLabel: string
+): string => `你是 WechatExplorer 的受控本地聊天搜索代理，只负责决定下一步检索，不回答用户问题。
+用户问题：${question}
+允许范围：${scopeLabel}；时间范围：${rangeLabel}。
+
+你只能输出一个 JSON 对象，不能输出 Markdown、解释、代码、SQL、文件路径或任何系统操作。
+唯一合法格式：
+{"action":"tool","tool":"search_people|search_conversations|search_messages|get_conversation_messages|get_messages_by_time|get_message_context","arguments":{...}}
+或：
+{"action":"finalize","reason":"已有足够证据"}
+
+规则：
+- 只能使用此前 Tool 返回的 conversationRef/messageRef；不得猜测或创建引用。
+- 问“我和某人最近聊了什么”时，优先 search_people 或 search_conversations，再 get_conversation_messages；不要把联系人名当消息关键词。
+- 搜索会话没有结果时，可根据结果自行尝试更短或更自然的名称表达，但最多五次 Tool 调用。
+- Tool 结果不足时可以改 Tool 或查询策略；结果充分时 finalize。
+- 不要请求全部聊天记录；遵守 Tool 返回的受限结果。`
+
+const traceArguments = (
+  argumentsValue: Record<string, unknown>
+): Record<string, string | number | boolean> => {
+  const result: Record<string, string | number | boolean> = {}
+  if (typeof argumentsValue.query === 'string') result.query = argumentsValue.query.slice(0, 80)
+  if (typeof argumentsValue.limit === 'number') result.limit = argumentsValue.limit
+  if (typeof argumentsValue.startTime === 'number') result.startTime = argumentsValue.startTime
+  if (typeof argumentsValue.endTime === 'number') result.endTime = argumentsValue.endTime
+  if (typeof argumentsValue.conversationRef === 'string') result.target = '已选择会话'
+  if (typeof argumentsValue.messageRef === 'string') result.context = '已选择消息'
+  return result
+}
+
+export async function runControlledSearchAgent(
+  options: ControlledSearchAgentOptions
+): Promise<ControlledSearchAgentResult> {
+  let toolCalls = 0
+  let previousResult = '尚未执行 Tool。'
+  options.onTrace({ event: 'agentStart', label: '开始规划本次本地检索' })
+
+  const maxToolCalls = options.maxToolCalls || MAX_AGENT_TOOL_CALLS
+  while (toolCalls < maxToolCalls) {
+    const decisionStartedAt = Date.now()
+    const decisionInput = `${agentSystemPrompt(options.question, options.scopeLabel, options.rangeLabel)}\n\n上一次 Tool 结果：${previousResult}`
+    const output = await options.decide(decisionInput)
+    const decisionElapsedMs = Date.now() - decisionStartedAt
+    const action = parseAction(output)
+    if (!action) return { status: 'invalid', toolCalls, reason: 'Agent 返回的控制协议无效' }
+    if (action.action === 'finalize') {
+      options.onTrace({
+        event: 'agentDecision',
+        label: 'Agent 判断现有结果足够',
+        decision: action.reason,
+        decisionInput: decisionInput.slice(0, 8_000),
+        elapsedMs: decisionElapsedMs
+      })
+      return { status: 'finalized', toolCalls, reason: action.reason }
+    }
+
+    options.onTrace({
+      event: 'agentDecision',
+      label: 'Agent 选择下一次检索',
+      toolName: action.tool,
+      elapsedMs: decisionElapsedMs,
+      decisionInput: decisionInput.slice(0, 8_000)
+    })
+    toolCalls += 1
+    options.onTrace({
+      event: 'toolCallStart',
+      label: '正在执行本地检索',
+      toolName: action.tool,
+      arguments: traceArguments(action.arguments)
+    })
+    const toolStartedAt = Date.now()
+    try {
+      const result = await options.execute(action)
+      const elapsedMs = Date.now() - toolStartedAt
+      options.onTrace({
+        event: 'toolCallEnd',
+        label: '本地检索完成',
+        toolName: action.tool,
+        resultCount: result.candidateCount,
+        elapsedMs
+      })
+      previousResult = JSON.stringify(result.summary)
+      if (result.finalizeReason) {
+        options.onTrace({
+          event: 'agentDecision',
+          label: '本地资料已覆盖所选时间范围，可直接整理回答',
+          decision: result.finalizeReason,
+          elapsedMs: 0
+        })
+        return { status: 'finalized', toolCalls, reason: result.finalizeReason }
+      }
+    } catch (error) {
+      const elapsedMs = Date.now() - toolStartedAt
+      const message = error instanceof Error ? error.message : '本次本地检索不可用'
+      options.onTrace({
+        event: 'toolCallEnd',
+        label: '本地检索未返回结果',
+        toolName: action.tool,
+        resultCount: 0,
+        elapsedMs,
+        decision: message.slice(0, 160)
+      })
+      previousResult = JSON.stringify({ error: message.slice(0, 160), results: [] })
+    }
+  }
+  options.onTrace({
+    event: 'agentDecision',
+    label: '已达到本次检索上限',
+    decision: `最多允许 ${maxToolCalls} 次本地检索`
+  })
+  return { status: 'exhausted', toolCalls, reason: '已达到本次检索上限' }
+}

src/main/services/ai-search-evidence.ts

@@ -0,0 +1,217 @@
+import type {
+  AiSearchAggregation,
+  AiSearchFinalEvidence,
+  AiSearchPipelineEvidence
+} from '../../shared/ai-search'
+
+export type EvidenceBuildResult = {
+  evidence: AiSearchFinalEvidence[]
+  aggregation: AiSearchAggregation
+  candidateCount: number
+  deduplicatedCount: number
+  candidateRankingMs: number
+  evidenceBuildMs: number
+  aggregationMs: number
+}
+
+export type CitationValidationResult = {
+  answer: string
+  invalidCitationIds: string[]
+  status: 'valid' | 'sanitized'
+}
+
+export const evidenceIdentity = (
+  item: Pick<AiSearchPipelineEvidence, 'conversationId' | 'messageId'>
+): string => `${item.conversationId}\u0000${item.messageId}`
+
+const compareEvidence = (left: AiSearchPipelineEvidence, right: AiSearchPipelineEvidence): number =>
+  (left.score ?? 0) - (right.score ?? 0) ||
+  right.timestamp - left.timestamp ||
+  evidenceIdentity(left).localeCompare(evidenceIdentity(right))
+
+const personIdentity = (item: AiSearchFinalEvidence): string =>
+  item.senderId
+    ? `sender:${item.senderId}`
+    : `conversation:${item.conversationId}:name:${item.sender}`
+
+export function buildEvidenceAggregation(evidence: AiSearchFinalEvidence[]): AiSearchAggregation {
+  const people = new Map<
+    string,
+    {
+      id: string
+      name: string
+      messageCount: number
+      conversationIds: Set<string>
+      lastMessageAt: number
+      evidenceIds: AiSearchFinalEvidence['id'][]
+    }
+  >()
+  const conversations = new Map<
+    string,
+    {
+      id: string
+      name: string
+      type: 'user' | 'group'
+      messageCount: number
+      people: Set<string>
+      lastMessageAt: number
+      evidenceIds: AiSearchFinalEvidence['id'][]
+    }
+  >()
+
+  for (const item of evidence) {
+    const personId = personIdentity(item)
+    const person = people.get(personId) || {
+      id: personId,
+      name: item.sender,
+      messageCount: 0,
+      conversationIds: new Set<string>(),
+      lastMessageAt: item.timestamp,
+      evidenceIds: []
+    }
+    person.messageCount += 1
+    person.conversationIds.add(item.conversationId)
+    person.lastMessageAt = Math.max(person.lastMessageAt, item.timestamp)
+    person.evidenceIds.push(item.id)
+    people.set(personId, person)
+
+    const conversation = conversations.get(item.conversationId) || {
+      id: item.conversationId,
+      name: item.conversationName,
+      type: item.conversationType,
+      messageCount: 0,
+      people: new Set<string>(),
+      lastMessageAt: item.timestamp,
+      evidenceIds: []
+    }
+    conversation.messageCount += 1
+    conversation.people.add(personId)
+    conversation.lastMessageAt = Math.max(conversation.lastMessageAt, item.timestamp)
+    conversation.evidenceIds.push(item.id)
+    conversations.set(item.conversationId, conversation)
+  }
+
+  return {
+    messageCount: evidence.length,
+    peopleCount: people.size,
+    conversationCount: conversations.size,
+    people: Array.from(people.values())
+      .map((person) => ({
+        id: person.id,
+        name: person.name,
+        messageCount: person.messageCount,
+        conversationCount: person.conversationIds.size,
+        lastMessageAt: person.lastMessageAt,
+        evidenceIds: person.evidenceIds
+      }))
+      .sort(
+        (left, right) =>
+          right.messageCount - left.messageCount || right.lastMessageAt - left.lastMessageAt
+      ),
+    conversations: Array.from(conversations.values())
+      .map((conversation) => ({
+        id: conversation.id,
+        name: conversation.name,
+        type: conversation.type,
+        messageCount: conversation.messageCount,
+        peopleCount: conversation.people.size,
+        lastMessageAt: conversation.lastMessageAt,
+        evidenceIds: conversation.evidenceIds
+      }))
+      .sort(
+        (left, right) =>
+          right.messageCount - left.messageCount || right.lastMessageAt - left.lastMessageAt
+      )
+  }
+}
+
+/**
+ * Performs all candidate ordering, identity de-duplication, final limiting and
+ * program-owned citation assignment in one place. Nothing downstream receives
+ * the candidate list as an AI context.
+ */
+export function buildFinalEvidence(
+  candidates: AiSearchPipelineEvidence[],
+  limit: number,
+  options?: { strategy?: 'ranked' | 'conversation_coverage' }
+): EvidenceBuildResult {
+  const rankingStartedAt = Date.now()
+  const ranked = [...candidates].sort(compareEvidence)
+  const candidateRankingMs = Date.now() - rankingStartedAt
+
+  const evidenceStartedAt = Date.now()
+  const unique = new Map<string, AiSearchPipelineEvidence>()
+  for (const item of ranked) {
+    const identity = evidenceIdentity(item)
+    if (!unique.has(identity)) unique.set(identity, item)
+  }
+  const uniqueEvidence = Array.from(unique.values())
+  const selected =
+    options?.strategy === 'conversation_coverage'
+      ? selectConversationCoverage(uniqueEvidence, limit)
+      : uniqueEvidence.slice(0, Math.max(1, limit))
+  const evidence = selected.map((item, index) => ({ ...item, id: `E${index + 1}` as const }))
+  const evidenceBuildMs = Date.now() - evidenceStartedAt
+
+  const aggregationStartedAt = Date.now()
+  const aggregation = buildEvidenceAggregation(evidence)
+  const aggregationMs = Date.now() - aggregationStartedAt
+
+  return {
+    evidence,
+    aggregation,
+    candidateCount: candidates.length,
+    deduplicatedCount: unique.size,
+    candidateRankingMs,
+    evidenceBuildMs,
+    aggregationMs
+  }
+}
+
+/**
+ * A recent-conversation answer should cover separate local conversation chunks,
+ * not merely pick eight adjacent newest messages from one exchange.
+ */
+function selectConversationCoverage(
+  evidence: AiSearchPipelineEvidence[],
+  limit: number
+): AiSearchPipelineEvidence[] {
+  const max = Math.max(1, limit)
+  const byChunk = new Map<string, AiSearchPipelineEvidence[]>()
+  for (const item of evidence) {
+    const chunk = byChunk.get(item.chunkId) || []
+    chunk.push(item)
+    byChunk.set(item.chunkId, chunk)
+  }
+  const representatives = Array.from(byChunk.values())
+    .map((items) => [...items].sort(compareEvidence)[0])
+    .sort((left, right) => left.timestamp - right.timestamp)
+  if (representatives.length <= max) return representatives
+  const selected: AiSearchPipelineEvidence[] = []
+  for (let index = 0; index < max; index += 1) {
+    const position = Math.round((index * (representatives.length - 1)) / (max - 1 || 1))
+    const item = representatives[position]
+    if (item && !selected.includes(item)) selected.push(item)
+  }
+  return selected
+}
+
+/** Do not expose citations that cannot resolve to program-owned Final Evidence. */
+export function sanitizeAnswerCitations(
+  answer: string,
+  evidence: Array<Pick<AiSearchFinalEvidence, 'id'>>
+): CitationValidationResult {
+  const allowed = new Set(evidence.map((item) => item.id))
+  const invalidCitationIds = new Set<string>()
+  const sanitized = answer.replace(/\[E(\d+)\]/g, (citation, number: string) => {
+    const id = `E${number}`
+    if (allowed.has(id as AiSearchFinalEvidence['id'])) return citation
+    invalidCitationIds.add(id)
+    return ''
+  })
+  return {
+    answer: sanitized,
+    invalidCitationIds: Array.from(invalidCitationIds),
+    status: invalidCitationIds.size ? 'sanitized' : 'valid'
+  }
+}

src/main/services/cache-service.ts

@@ -7,6 +7,11 @@ import type { CacheClearScope, CacheSummary, CacheSummaryItem } from '../../shar
 export type { CacheClearScope } from '../../shared/cache'
 
 const BOOTSTRAP_CACHE_DIR = path.join(app.getPath('userData'), 'cache', 'bootstrap')
+const KNOWLEDGE_CACHE_DIR = path.join(app.getPath('userData'), 'knowledge')
+
+export interface CacheClearOptions {
+  beforeClearKnowledge?: () => Promise<void>
+}
 
 function inspectDirectory(directory: string): { sizeBytes: number; fileCount: number } {
   if (!fs.existsSync(directory)) return { sizeBytes: 0, fileCount: 0 }
@@ -40,6 +45,7 @@ function inspectDirectory(directory: string): { sizeBytes: number; fileCount: nu
 export function getCacheSummary(): CacheSummary {
   const bootstrap = inspectDirectory(BOOTSTRAP_CACHE_DIR)
   const electron = inspectDirectory(path.join(app.getPath('userData'), 'Cache'))
+  const knowledge = inspectDirectory(KNOWLEDGE_CACHE_DIR)
   const items: CacheSummaryItem[] = [
     {
       id: 'bootstrap',
@@ -52,6 +58,12 @@ export function getCacheSummary(): CacheSummary {
       label: '应用临时缓存',
       description: 'Electron 页面资源缓存，清理后会自动重新生成。',
       ...electron
+    },
+    {
+      id: 'knowledge',
+      label: '本地知识库索引',
+      description: '为问问微信建立的所有账号本地检索索引。清理后需手动重新建立，不影响微信原始数据。',
+      ...knowledge
     }
   ]
   return {
@@ -61,7 +73,10 @@ export function getCacheSummary(): CacheSummary {
   }
 }
 
-export async function clearCache(scope: CacheClearScope): Promise<CacheSummary> {
+export async function clearCache(
+  scope: CacheClearScope,
+  options: CacheClearOptions = {}
+): Promise<CacheSummary> {
   if (scope === 'bootstrap' || scope === 'all') {
     clearBootstrapCache()
     await fs.remove(BOOTSTRAP_CACHE_DIR)
@@ -69,5 +84,9 @@ export async function clearCache(scope: CacheClearScope): Promise<CacheSummary>
   if (scope === 'electron' || scope === 'all') {
     await session.defaultSession.clearCache()
   }
+  if (scope === 'knowledge' || scope === 'all') {
+    await options.beforeClearKnowledge?.()
+    await fs.remove(KNOWLEDGE_CACHE_DIR)
+  }
   return getCacheSummary()
 }

src/main/services/contact-resolution-service.ts

@@ -0,0 +1,86 @@
+import type { Contact } from '../../shared/types'
+import {
+  emptyContactResolution,
+  normalizeContactName,
+  type ContactResolutionCandidate,
+  type ContactResolutionMatch,
+  type ContactResolutionResult
+} from '../../shared/contact-resolution'
+
+export type ContactResolutionScope = 'any' | 'person' | 'group'
+
+const displayName = (contact: Contact): string =>
+  contact.m_nsNickName || contact.remark || contact.wechatNickname || contact.m_nsUsrName
+
+const aliases = (contact: Contact): Array<{ value: string; primary: boolean }> =>
+  [
+    { value: contact.m_nsNickName, primary: true },
+    { value: contact.remark || '', primary: false },
+    { value: contact.wechatNickname || '', primary: false },
+    { value: contact.m_nsUsrName, primary: false }
+  ].filter((item) => Boolean(normalizeContactName(item.value)))
+
+/**
+ * The one main-process authority that converts a user/Agent supplied name to
+ * an existing conversation. It only auto-confirms an exact canonical alias.
+ * Fuzzy discovery intentionally returns candidates rather than a guessed ID.
+ */
+export function resolveContact(
+  query: string,
+  contacts: Contact[],
+  scope: ContactResolutionScope = 'any'
+): ContactResolutionResult {
+  const normalizedQuery = normalizeContactName(query)
+  if (!normalizedQuery) return emptyContactResolution()
+  const matches = new Map<string, { contact: Contact; matchedBy: ContactResolutionMatch }>()
+
+  for (const contact of contacts) {
+    if (!contact.md5) continue
+    if (scope === 'person' && contact.type !== 'user') continue
+    if (scope === 'group' && contact.type !== 'group') continue
+    for (const alias of aliases(contact)) {
+      if (normalizeContactName(alias.value) !== normalizedQuery) continue
+      const rawExact =
+        alias.value.trim().normalize('NFKC').toLocaleLowerCase() ===
+        query.trim().normalize('NFKC').toLocaleLowerCase()
+      const matchedBy: ContactResolutionMatch = rawExact
+        ? 'exact'
+        : alias.primary
+          ? 'normalized'
+          : 'alias'
+      const current = matches.get(contact.md5)
+      if (!current || (current.matchedBy === 'alias' && matchedBy !== 'alias')) {
+        matches.set(contact.md5, { contact, matchedBy })
+      }
+    }
+  }
+
+  const candidates: ContactResolutionCandidate[] = Array.from(matches.values())
+    .map(({ contact, matchedBy }) => ({
+      conversationId: contact.md5,
+      displayName: displayName(contact),
+      matchedBy,
+      confidence: 1
+    }))
+    .sort((left, right) => left.displayName.localeCompare(right.displayName, 'zh-CN'))
+  if (candidates.length !== 1) {
+    return {
+      ...emptyContactResolution(),
+      candidates,
+      ambiguous: candidates.length > 1
+    }
+  }
+  const candidate = candidates[0]
+  const contact = matches.get(candidate.conversationId)!.contact
+  return {
+    matched: true,
+    personId: contact.m_nsUsrName,
+    conversationId: contact.md5,
+    canonicalName: displayName(contact),
+    displayName: candidate.displayName,
+    matchedBy: candidate.matchedBy,
+    confidence: candidate.confidence,
+    candidates,
+    ambiguous: false
+  }
+}

src/preload/index.d.ts

@@ -52,6 +52,16 @@ import type {
   VoiceModelStatus,
   VoiceRecognitionResult
 } from '../shared/voice-recognition'
+import type {
+  AiSearchPipelineRequest,
+  AiSearchPipelineResult,
+  AiSearchProgressEvent
+} from '../shared/ai-search'
+import type {
+  KnowledgeRuntimeStatus,
+  KnowledgeSearchIpcRequest,
+  KnowledgeSearchIpcResult
+} from '../shared/knowledge'
 
 export type ParsedContent =
   | { type: 'text'; content: string }
@@ -121,7 +131,7 @@ declare global {
       installAppUpdate: () => Promise<{ success: boolean; error?: string }>
       onAppUpdateState: (callback: (state: AppUpdateState) => void) => () => void
       getCacheSummary: () => Promise<CacheSummary>
-      clearCache: (scope: 'bootstrap' | 'electron' | 'all') => Promise<CacheSummary>
+      clearCache: (scope: 'bootstrap' | 'electron' | 'knowledge' | 'all') => Promise<CacheSummary>
       initDb: (
         key: string,
         accountRoot: string
@@ -183,6 +193,12 @@ declare global {
         }[]
       } | null>
       search: (keyword: string) => Promise<string | null>
+      searchKnowledge: (request: KnowledgeSearchIpcRequest) => Promise<KnowledgeSearchIpcResult>
+      runAiSearch: (request: AiSearchPipelineRequest) => Promise<AiSearchPipelineResult>
+      onAiSearchProgress: (callback: (progress: AiSearchProgressEvent) => void) => () => void
+      getKnowledgeStatus: () => Promise<KnowledgeRuntimeStatus>
+      startKnowledgeIndex: () => Promise<KnowledgeRuntimeStatus>
+      onKnowledgeStatus: (callback: (status: KnowledgeRuntimeStatus) => void) => () => void
       aiChat: (
         messages: { role: string; content: string }[],
         options?: AIChatRequestOptions

src/preload/index.ts

@@ -29,6 +29,16 @@ import type {
   VoiceModelStatus,
   VoiceRecognitionResult
 } from '../shared/voice-recognition'
+import type {
+  AiSearchPipelineRequest,
+  AiSearchPipelineResult,
+  AiSearchProgressEvent
+} from '../shared/ai-search'
+import type {
+  KnowledgeRuntimeStatus,
+  KnowledgeSearchIpcRequest,
+  KnowledgeSearchIpcResult
+} from '../shared/knowledge'
 
 // 渲染器的自定义 API
 const api = {
@@ -46,7 +56,7 @@ const api = {
     return () => ipcRenderer.removeListener('app-update:state', listener)
   },
   getCacheSummary: (): Promise<CacheSummary> => ipcRenderer.invoke('cache:getSummary'),
-  clearCache: (scope: 'bootstrap' | 'electron' | 'all'): Promise<CacheSummary> =>
+  clearCache: (scope: 'bootstrap' | 'electron' | 'knowledge' | 'all'): Promise<CacheSummary> =>
     ipcRenderer.invoke('cache:clear', scope),
   initDb: (key: string, accountRoot: string) => ipcRenderer.invoke('db:init', key, accountRoot),
   discoverAccounts: (inputPath: string): Promise<AccountDiscoveryResult> =>
@@ -67,6 +77,26 @@ const api = {
   ) => ipcRenderer.invoke('db:getMessages', userMd5, startTime, endTime, options),
   getGroupSnapshot: (userMd5: string) => ipcRenderer.invoke('db:getGroupSnapshot', userMd5),
   search: (keyword: string) => ipcRenderer.invoke('db:search', keyword),
+  searchKnowledge: (request: KnowledgeSearchIpcRequest): Promise<KnowledgeSearchIpcResult> =>
+    ipcRenderer.invoke('knowledge:search', request),
+  runAiSearch: (request: AiSearchPipelineRequest): Promise<AiSearchPipelineResult> =>
+    ipcRenderer.invoke('ai-search:run', request),
+  onAiSearchProgress: (callback: (progress: AiSearchProgressEvent) => void) => {
+    const listener = (_event: Electron.IpcRendererEvent, progress: AiSearchProgressEvent): void =>
+      callback(progress)
+    ipcRenderer.on('ai-search:progress', listener)
+    return () => ipcRenderer.removeListener('ai-search:progress', listener)
+  },
+  getKnowledgeStatus: (): Promise<KnowledgeRuntimeStatus> =>
+    ipcRenderer.invoke('knowledge:getStatus'),
+  startKnowledgeIndex: (): Promise<KnowledgeRuntimeStatus> =>
+    ipcRenderer.invoke('knowledge:startIndex'),
+  onKnowledgeStatus: (callback: (status: KnowledgeRuntimeStatus) => void) => {
+    const listener = (_event: Electron.IpcRendererEvent, status: KnowledgeRuntimeStatus): void =>
+      callback(status)
+    ipcRenderer.on('knowledge:status', listener)
+    return () => ipcRenderer.removeListener('knowledge:status', listener)
+  },
   aiChat: (messages: { role: string; content: string }[], options?: AIChatRequestOptions) =>
     ipcRenderer.invoke('ai:chat', messages, options),
   listAIProviders: () => ipcRenderer.invoke('ai:listProviders'),

src/renderer/src/components/search/searchMarkdown.tsx

@@ -1,7 +1,16 @@
 import React from 'react'
 
-const inlineMarkdown = (value: string, keyPrefix: string): React.ReactNode[] =>
-  value.split(/(\*\*.*?\*\*|`.*?`|\*.*?\*)/g).map((part, index) => {
+type MarkdownOptions = {
+  evidenceCount?: number
+  onEvidenceClick?: (index: number) => void
+}
+
+const inlineMarkdown = (
+  value: string,
+  keyPrefix: string,
+  options: MarkdownOptions = {}
+): React.ReactNode[] =>
+  value.split(/(\*\*.*?\*\*|`.*?`|\*.*?\*|\[E\d+\])/g).map((part, index) => {
     const key = `${keyPrefix}-${index}`
     if (part.startsWith('**') && part.endsWith('**')) {
       return <strong key={key}>{part.slice(2, -2)}</strong>
@@ -12,24 +21,45 @@ const inlineMarkdown = (value: string, keyPrefix: string): React.ReactNode[] =>
     if (part.startsWith('*') && part.endsWith('*')) {
       return <em key={key}>{part.slice(1, -1)}</em>
     }
+    const evidence = /^\[E(\d+)\]$/.exec(part)
+    if (evidence) {
+      const evidenceIndex = Number(evidence[1]) - 1
+      if (
+        evidenceIndex >= 0 &&
+        evidenceIndex < (options.evidenceCount || 0) &&
+        options.onEvidenceClick
+      ) {
+        return (
+          <button
+            key={key}
+            type="button"
+            className="ai-search-inline-evidence"
+            onClick={() => options.onEvidenceClick?.(evidenceIndex)}
+            title={`查看证据 E${evidenceIndex + 1}`}
+          >
+            {part}
+          </button>
+        )
+      }
+    }
     return <React.Fragment key={key}>{part}</React.Fragment>
   })
 
-export const renderMarkdown = (value: string): React.ReactNode =>
+export const renderMarkdown = (value: string, options: MarkdownOptions = {}): React.ReactNode =>
   value.split(/\r?\n/).map((line, index) => {
     const key = `markdown-${index}`
     if (!line.trim()) return <div key={key} className="ai-search-markdown-spacer" />
     const heading = /^(#{1,3})\s+(.+)$/.exec(line)
     if (heading) {
       const Heading = `h${heading[1].length}` as 'h1' | 'h2' | 'h3'
-      return <Heading key={key}>{inlineMarkdown(heading[2], key)}</Heading>
+      return <Heading key={key}>{inlineMarkdown(heading[2], key, options)}</Heading>
     }
     const bullet = /^\s*[-*]\s+(.+)$/.exec(line)
     if (bullet) {
       return (
         <div key={key} className="ai-search-markdown-list-item">
           <span aria-hidden>•</span>
-          <span>{inlineMarkdown(bullet[1], key)}</span>
+          <span>{inlineMarkdown(bullet[1], key, options)}</span>
         </div>
       )
     }
@@ -38,9 +68,9 @@ export const renderMarkdown = (value: string): React.ReactNode =>
       return (
         <div key={key} className="ai-search-markdown-list-item">
           <span aria-hidden>{line.trim().match(/^\d+/)?.[0]}.</span>
-          <span>{inlineMarkdown(numbered[1], key)}</span>
+          <span>{inlineMarkdown(numbered[1], key, options)}</span>
         </div>
       )
     }
-    return <p key={key}>{inlineMarkdown(line, key)}</p>
+    return <p key={key}>{inlineMarkdown(line, key, options)}</p>
   })

src/renderer/src/components/search/searchTypes.ts

@@ -1,12 +1,14 @@
 import type { AIRuntimeModelConfig } from '../../../../shared/ai-provider'
 import type { Contact, Message } from '../../../../shared/types'
 
-export type SearchStage = 'idle' | 'loading' | 'result' | 'insufficient'
-export type SearchScope = 'global' | 'conversation'
+export type SearchStage = 'idle' | 'loading' | 'result' | 'partial' | 'insufficient'
+export type SearchScope = 'global' | 'groups' | 'contacts' | 'conversation'
 export type SearchRange = 'today' | '7d' | '30d' | 'all'
 export type SearchIntent = 'general' | 'topic' | 'participants' | 'mixed'
 
 export interface EvidenceItem {
+  /** Program-owned Final Evidence ID. Cached legacy records may omit it. */
+  evidenceId?: string
   contact: Contact
   message: Message
 }

src/renderer/src/components/search/searchUtils.ts

@@ -8,7 +8,9 @@ export const RANGE_LABELS: Record<SearchRange, string> = {
   all: '全部历史'
 }
 
-export const SEARCH_CACHE_KEY = 'wxe_ai_search_cache_v8'
+// Final Evidence IDs are now program-owned; never replay answers cached under
+// the former candidate-context contract.
+export const SEARCH_CACHE_KEY = 'wxe_ai_search_cache_v9'
 export const SEARCH_HISTORY_KEY = 'wxe_ai_search_history_v1'
 export const SEARCH_CACHE_LIMIT = 20
 export const currentTimestamp = (): number => Date.now()
@@ -270,7 +272,8 @@ export const senderName = (message: Message, contact: Contact, names: Record<str
   return contact.type === 'user' ? contact.m_nsNickName || '联系人' : '群成员'
 }
 
-export const compactCacheItem = ({ contact, message }: EvidenceItem): EvidenceItem => ({
+export const compactCacheItem = ({ evidenceId, contact, message }: EvidenceItem): EvidenceItem => ({
+  evidenceId,
   contact: {
     md5: contact.md5,
     m_nsUsrName: contact.m_nsUsrName,

src/renderer/src/features/settings/pages/CacheCleanupPage.tsx

@@ -12,7 +12,9 @@ function formatBytes(value: number): string {
 
 export function CacheCleanupPage({ onNotice }: { onNotice: (message: string) => void }): React.ReactElement {
   const [summary, setSummary] = useState<CacheSummary | null>(null)
-  const [busyScope, setBusyScope] = useState<'bootstrap' | 'electron' | 'all' | 'local' | null>(null)
+  const [busyScope, setBusyScope] = useState<
+    'bootstrap' | 'electron' | 'knowledge' | 'all' | 'local' | null
+  >(null)
 
   const refresh = useCallback(async (): Promise<void> => {
     setSummary(await window.api.getCacheSummary())
@@ -29,14 +31,22 @@ export function CacheCleanupPage({ onNotice }: { onNotice: (message: string) =>
     onNotice('已清理检索和导出本地缓存')
   }
 
-  const clear = async (scope: 'bootstrap' | 'electron' | 'all'): Promise<void> => {
+  const clear = async (scope: 'bootstrap' | 'electron' | 'knowledge' | 'all'): Promise<void> => {
     setBusyScope(scope)
     try {
+      setSummary(await window.api.clearCache(scope))
       if (scope === 'all') {
         for (const key of SEARCH_CACHE_KEYS) localStorage.removeItem(key)
       }
-      setSummary(await window.api.clearCache(scope))
-      onNotice(scope === 'all' ? '已清理全部可恢复缓存和检索记录' : '缓存已清理')
+      onNotice(
+        scope === 'knowledge'
+          ? '已清理所有账号的本地知识库索引，需要时可在问问微信中重新建立'
+          : scope === 'all'
+            ? '已清理全部可恢复缓存和检索记录'
+            : '缓存已清理'
+      )
+    } catch (error) {
+      onNotice(error instanceof Error ? error.message : '清理缓存失败')
     } finally {
       setBusyScope(null)
     }

src/renderer/src/styles/search.scss

@@ -118,6 +118,9 @@
 }
 
 .ai-search-scope-panel {
+  display: flex;
+  flex-direction: column;
+  gap: 20px;
   padding: 16px 12px;
   border-right: 1px solid var(--wxex-border);
 }
@@ -152,7 +155,252 @@
     font-weight: 700;
     line-height: 20px;
   }
+}
 
+.ai-search-header-actions {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+}
+
+.ai-search-knowledge-pill {
+  display: inline-flex;
+  align-items: center;
+  gap: 7px;
+  padding: 5px 10px;
+  border: 1px solid var(--wxex-border);
+  border-radius: 999px;
+  background: var(--wxex-bg-elevated);
+  color: var(--wxex-text-secondary);
+  font-size: 11px;
+  white-space: nowrap;
+}
+
+.ai-search-knowledge-dot {
+  width: 7px;
+  height: 7px;
+  flex: 0 0 auto;
+  border-radius: 50%;
+  background: var(--wxex-text-muted);
+}
+
+.ai-search-knowledge-card.building .ai-search-knowledge-dot,
+.ai-search-knowledge-card.syncing .ai-search-knowledge-dot {
+  background: var(--wxex-warning);
+}
+
+.ai-search-knowledge-card.ready .ai-search-knowledge-dot {
+  background: var(--wxex-success);
+}
+
+.ai-search-knowledge-card.error .ai-search-knowledge-dot {
+  background: var(--wxex-danger, #c15d4d);
+}
+
+.ai-search-filter-section {
+  flex: 0 0 auto;
+}
+
+.ai-search-secondary-menu,
+.ai-search-time-menu {
+  display: flex;
+  flex-direction: column;
+  gap: 3px;
+}
+
+.ai-search-secondary-menu button,
+.ai-search-time-menu button {
+  display: flex;
+  align-items: center;
+  gap: 9px;
+  width: 100%;
+  min-height: 34px;
+  padding: 7px 9px;
+  border: 0;
+  border-radius: var(--wxex-radius-sm);
+  background: transparent;
+  color: var(--wxex-text-secondary);
+  cursor: pointer;
+  font: inherit;
+  font-size: 11px;
+  text-align: left;
+}
+
+.ai-search-secondary-menu button:hover,
+.ai-search-time-menu button:hover,
+.ai-search-secondary-menu button.active,
+.ai-search-time-menu button.active {
+  background: var(--wxex-brand-soft);
+  color: var(--wxex-brand);
+  font-weight: 700;
+}
+
+.ai-search-secondary-menu button:disabled {
+  color: var(--wxex-text-muted);
+  cursor: not-allowed;
+  opacity: 0.65;
+}
+
+.ai-search-secondary-menu button span,
+.ai-search-time-menu button span {
+  width: 18px;
+  color: currentColor;
+  font-size: 14px;
+  text-align: center;
+}
+
+.ai-search-time-section {
+  padding-bottom: 3px;
+  border-bottom: 1px solid var(--wxex-border);
+}
+
+.ai-search-knowledge-card {
+  display: grid;
+  gap: 10px;
+  margin-top: auto;
+  padding: 12px;
+  border: 1px solid var(--wxex-border);
+  border-radius: var(--wxex-radius-md);
+  background: var(--wxex-bg-elevated);
+}
+
+.ai-search-knowledge-card.building,
+.ai-search-knowledge-card.syncing {
+  border-color: color-mix(in srgb, var(--wxex-brand) 55%, var(--wxex-border));
+}
+
+.ai-search-knowledge-card.ready {
+  border-color: color-mix(in srgb, var(--wxex-success) 55%, var(--wxex-border));
+}
+
+.ai-search-knowledge-card.error {
+  border-color: color-mix(in srgb, var(--wxex-warning) 70%, var(--wxex-border));
+}
+
+.ai-search-knowledge-card-heading,
+.ai-search-sync-progress-top,
+.ai-search-knowledge-details > div {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+}
+
+.ai-search-knowledge-card-heading > div {
+  display: flex;
+  flex-direction: column;
+  gap: 2px;
+}
+
+.ai-search-knowledge-card-heading span:first-child {
+  color: var(--wxex-text-muted);
+  font-size: 9px;
+  font-weight: 700;
+  letter-spacing: 0.08em;
+}
+
+.ai-search-knowledge-card-heading strong {
+  color: var(--wxex-text-primary);
+  font-size: 12px;
+}
+
+.ai-search-knowledge-description,
+.ai-search-knowledge-error,
+.ai-search-knowledge-more p {
+  margin: 0;
+  color: var(--wxex-text-muted);
+  font-size: 10px;
+  line-height: 15px;
+}
+
+.ai-search-knowledge-error {
+  color: var(--wxex-warning);
+}
+
+.ai-search-sync-progress {
+  display: grid;
+  gap: 5px;
+}
+
+.ai-search-sync-progress-top {
+  color: var(--wxex-text-secondary);
+  font-size: 10px;
+}
+
+.ai-search-sync-progress-track {
+  height: 5px;
+  overflow: hidden;
+  border-radius: 999px;
+  background: var(--wxex-border);
+}
+
+.ai-search-sync-progress-track span {
+  display: block;
+  height: 100%;
+  border-radius: inherit;
+  background: var(--wxex-brand);
+  transition: width 250ms ease;
+}
+
+.ai-search-knowledge-card.building .ai-search-sync-progress-track span,
+.ai-search-knowledge-card.syncing .ai-search-sync-progress-track span {
+  animation: ai-search-sync-pulse 1.6s ease-in-out infinite;
+}
+
+@keyframes ai-search-sync-pulse {
+  0%,
+  100% {
+    opacity: 0.55;
+  }
+  50% {
+    opacity: 1;
+  }
+}
+
+.ai-search-knowledge-details {
+  display: grid;
+  gap: 5px;
+  padding-top: 2px;
+}
+
+.ai-search-knowledge-details span {
+  color: var(--wxex-text-muted);
+  font-size: 10px;
+}
+
+.ai-search-knowledge-details strong {
+  color: var(--wxex-text-secondary);
+  font-size: 10px;
+  font-weight: 600;
+}
+
+.ai-search-knowledge-action {
+  min-height: 30px;
+  padding: 6px 9px;
+  border: 1px solid var(--wxex-brand);
+  border-radius: var(--wxex-radius-sm);
+  background: var(--wxex-brand);
+  color: #fff;
+  cursor: pointer;
+  font: inherit;
+  font-size: 11px;
+  font-weight: 700;
+}
+
+.ai-search-knowledge-action:disabled {
+  cursor: wait;
+  opacity: 0.65;
+}
+
+.ai-search-knowledge-more {
+  color: var(--wxex-text-secondary);
+  font-size: 10px;
+}
+
+.ai-search-knowledge-more summary {
+  cursor: pointer;
+  color: var(--wxex-brand);
+  font-weight: 700;
 }
 
 .ai-search-local-badge,
@@ -566,10 +814,156 @@
   }
 }
 
+.ai-search-loading {
+  align-items: flex-start;
+  justify-content: flex-start;
+  padding-top: clamp(34px, 8vh, 88px);
+  text-align: left;
+
+  > h2,
+  > p {
+    width: 100%;
+    max-width: none;
+  }
+}
+
+.ai-search-pipeline {
+  width: 100%;
+  display: grid;
+  gap: 0;
+  margin-top: 24px;
+  border: 1px solid var(--wxex-border);
+  border-radius: var(--wxex-radius-md);
+  background: var(--wxex-bg-elevated);
+  overflow: hidden;
+}
+
+.ai-search-pipeline-step {
+  position: relative;
+  display: grid;
+  grid-template-columns: 28px minmax(0, 1fr);
+  gap: 10px;
+  padding: 14px 16px;
+  border-bottom: 1px solid var(--wxex-border);
+  color: var(--wxex-text-muted);
+
+  &:last-child {
+    border-bottom: 0;
+  }
+
+  > div {
+    min-width: 0;
+  }
+
+  strong {
+    display: block;
+    color: var(--wxex-text-secondary);
+    font-size: 12px;
+    line-height: 18px;
+  }
+
+  p {
+    max-width: none;
+    margin-top: 2px;
+    color: var(--wxex-text-muted);
+    font-size: 11px;
+    line-height: 17px;
+  }
+
+  &.done strong {
+    color: var(--wxex-success);
+  }
+
+  &.active {
+    background: color-mix(in srgb, var(--wxex-brand-soft) 58%, transparent);
+
+    strong,
+    p {
+      color: var(--wxex-brand);
+    }
+  }
+
+  &.error {
+    background: #fff8ef;
+
+    strong,
+    p {
+      color: var(--wxex-warning);
+    }
+  }
+}
+
+.ai-search-pipeline-mark {
+  width: 22px;
+  height: 22px;
+  display: grid;
+  place-items: center;
+  margin-top: 1px;
+  border: 1px solid var(--wxex-border);
+  border-radius: 50%;
+  background: var(--wxex-bg-main);
+  color: var(--wxex-text-muted);
+  font-size: 11px;
+  font-weight: 700;
+}
+
+.ai-search-pipeline-step.done .ai-search-pipeline-mark {
+  border-color: color-mix(in srgb, var(--wxex-success) 45%, transparent);
+  background: color-mix(in srgb, var(--wxex-success) 11%, transparent);
+  color: var(--wxex-success);
+}
+
+.ai-search-pipeline-step.active .ai-search-pipeline-mark {
+  border-color: var(--wxex-brand);
+  color: var(--wxex-brand);
+  animation: ai-search-pipeline-pulse 1.2s ease-in-out infinite;
+}
+
+.ai-search-pipeline-step.error .ai-search-pipeline-mark {
+  border-color: var(--wxex-warning);
+  background: #fff1df;
+  color: var(--wxex-warning);
+}
+
+@keyframes ai-search-pipeline-pulse {
+  50% {
+    transform: scale(1.08);
+    box-shadow: 0 0 0 5px color-mix(in srgb, var(--wxex-brand) 10%, transparent);
+  }
+}
+
+.ai-search-pipeline-details {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 5px;
+  margin-top: 7px;
+
+  span {
+    padding: 2px 6px;
+    border-radius: 4px;
+    background: var(--wxex-bg-sidebar);
+    color: var(--wxex-text-secondary);
+    font-size: 10px;
+    line-height: 15px;
+  }
+}
+
 .ai-search-result {
   width: min(820px, calc(100% - 48px));
   margin: 0 auto;
   padding: 28px 0 34px;
+  animation: ai-search-result-in 260ms ease-out both;
+}
+
+@keyframes ai-search-result-in {
+  from {
+    opacity: 0;
+    transform: translateY(8px);
+  }
+  to {
+    opacity: 1;
+    transform: translateY(0);
+  }
 }
 
 .ai-search-result-header {
@@ -617,6 +1011,41 @@
 
 .ai-search-summary-block {
   margin-top: 22px;
+  animation: ai-search-result-in 340ms 45ms ease-out both;
+}
+
+.ai-search-answer > * {
+  animation: ai-search-answer-line-in 260ms ease-out both;
+}
+
+.ai-search-answer > *:nth-child(1) {
+  animation-delay: 70ms;
+}
+.ai-search-answer > *:nth-child(2) {
+  animation-delay: 105ms;
+}
+.ai-search-answer > *:nth-child(3) {
+  animation-delay: 140ms;
+}
+.ai-search-answer > *:nth-child(4) {
+  animation-delay: 175ms;
+}
+.ai-search-answer > *:nth-child(5) {
+  animation-delay: 210ms;
+}
+.ai-search-answer > *:nth-child(n + 6) {
+  animation-delay: 245ms;
+}
+
+@keyframes ai-search-answer-line-in {
+  from {
+    opacity: 0;
+    transform: translateY(5px);
+  }
+  to {
+    opacity: 1;
+    transform: translateY(0);
+  }
 }
 
 .ai-search-section-heading {
@@ -647,6 +1076,176 @@
   white-space: pre-wrap;
 }
 
+.ai-search-answer-evidence,
+.ai-search-trace {
+  display: flex;
+  flex-wrap: wrap;
+  align-items: center;
+  gap: 6px;
+  color: var(--wxex-text-muted);
+  font-size: 10px;
+  line-height: 16px;
+}
+
+.ai-search-answer-evidence {
+  padding: 9px 12px;
+  border: 1px solid var(--wxex-border);
+  border-top: 0;
+  background: var(--wxex-bg-elevated);
+}
+
+.ai-search-answer-evidence button {
+  padding: 2px 6px;
+  border: 1px solid var(--wxex-ai);
+  border-radius: 4px;
+  background: transparent;
+  color: var(--wxex-ai);
+  cursor: pointer;
+  font: inherit;
+  font-weight: 700;
+}
+
+.ai-search-inline-evidence {
+  display: inline-flex;
+  align-items: center;
+  min-height: 22px;
+  margin: 0 2px;
+  padding: 1px 6px;
+  border: 1px solid color-mix(in srgb, var(--primary-color) 35%, transparent);
+  border-radius: 999px;
+  background: color-mix(in srgb, var(--primary-color) 10%, transparent);
+  color: var(--primary-color);
+  font: inherit;
+  font-size: 0.85em;
+  line-height: 1;
+  cursor: pointer;
+}
+
+.ai-search-inline-evidence:hover {
+  background: color-mix(in srgb, var(--primary-color) 18%, transparent);
+}
+
+.ai-search-trace {
+  margin-top: 9px;
+}
+
+.ai-search-trace span {
+  padding: 2px 5px;
+  border-radius: 4px;
+  background: var(--wxex-bg-sidebar);
+}
+
+.ai-search-details {
+  margin-top: 12px;
+  color: var(--wxex-text-secondary);
+  font-size: 11px;
+
+  summary {
+    width: fit-content;
+    color: var(--wxex-brand);
+    cursor: pointer;
+    list-style: none;
+  }
+
+  summary::-webkit-details-marker {
+    display: none;
+  }
+
+  summary::after {
+    content: ' ▾';
+  }
+
+  &[open] summary::after {
+    content: ' ▴';
+  }
+}
+
+.ai-search-details-grid {
+  display: grid;
+  grid-template-columns: repeat(3, minmax(0, 1fr));
+  gap: 8px;
+  margin-top: 10px;
+
+  section {
+    display: grid;
+    gap: 3px;
+    padding: 9px;
+    border: 1px solid var(--wxex-border);
+    border-radius: var(--wxex-radius-sm);
+    background: var(--wxex-bg-sidebar);
+  }
+
+  strong {
+    margin-bottom: 2px;
+    color: var(--wxex-text-primary);
+    font-size: 10px;
+  }
+
+  span {
+    color: var(--wxex-text-muted);
+    font-size: 10px;
+    line-height: 15px;
+  }
+}
+
+.ai-search-partial {
+  .ai-search-details {
+    width: 100%;
+    max-width: 640px;
+    margin-top: 18px;
+    text-align: left;
+  }
+}
+
+.ai-search-knowledge-status {
+  display: grid;
+  gap: 4px;
+  margin: 0 0 16px;
+  padding: 10px;
+  border: 1px solid var(--wxex-border);
+  border-radius: var(--wxex-radius-md);
+  background: var(--wxex-bg-elevated);
+}
+
+.ai-search-knowledge-status > div {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+}
+
+.ai-search-knowledge-status > div span,
+.ai-search-knowledge-status p,
+.ai-search-knowledge-status small {
+  margin: 0;
+  color: var(--wxex-text-muted);
+  font-size: 10px;
+  line-height: 15px;
+}
+
+.ai-search-knowledge-status > div span {
+  font-weight: 700;
+  letter-spacing: 0.04em;
+}
+
+.ai-search-knowledge-status strong {
+  color: var(--wxex-text-primary);
+  font-size: 11px;
+}
+
+.ai-search-knowledge-status.building,
+.ai-search-knowledge-status.syncing {
+  border-color: var(--wxex-ai);
+}
+
+.ai-search-knowledge-status.ready strong {
+  color: var(--wxex-success);
+}
+
+.ai-search-knowledge-status.error {
+  border-color: var(--wxex-warning);
+}
+
 .ai-search-insufficient-icon {
   width: 44px;
   height: 44px;
@@ -661,6 +1260,7 @@
 }
 
 .ai-search-composer {
+  position: relative;
   padding: 12px 18px 14px;
   border-top: 1px solid var(--wxex-border);
   background: var(--wxex-bg-elevated);
@@ -688,6 +1288,98 @@
   }
 }
 
+.ai-search-history-trigger {
+  margin-left: auto;
+  padding: 3px 7px !important;
+  border: 1px solid var(--wxex-border) !important;
+  border-radius: 999px !important;
+  background: var(--wxex-bg-main) !important;
+  color: var(--wxex-brand) !important;
+  cursor: pointer;
+  font-size: 10px !important;
+}
+
+.ai-search-history-popover {
+  z-index: 30;
+  width: min(420px, calc(100% - 36px));
+  max-height: 236px;
+  overflow: auto;
+  padding: 9px;
+  border: 1px solid var(--wxex-border);
+  border-radius: var(--wxex-radius-md);
+  background: var(--wxex-bg-elevated);
+  box-shadow: 0 10px 28px rgba(31, 52, 45, 0.16);
+  transform-origin: var(--radix-popover-content-transform-origin);
+  animation: ai-search-history-popover-in 150ms ease-out both;
+}
+
+@keyframes ai-search-history-popover-in {
+  from {
+    opacity: 0;
+    transform: translateY(5px) scale(0.985);
+  }
+  to {
+    opacity: 1;
+    transform: translateY(0) scale(1);
+  }
+}
+
+.ai-search-history-popover-heading,
+.ai-search-history-popover-item {
+  display: flex;
+  align-items: center;
+  gap: 7px;
+}
+
+.ai-search-history-popover-heading {
+  justify-content: space-between;
+  padding: 2px 3px 7px;
+  color: var(--wxex-text-primary);
+  font-size: 11px;
+}
+
+.ai-search-history-popover-heading button,
+.ai-search-history-popover-item > button:last-child {
+  width: 22px;
+  height: 22px;
+  flex: 0 0 auto;
+  padding: 0 !important;
+  border: 0 !important;
+  background: transparent !important;
+  color: var(--wxex-text-muted) !important;
+  cursor: pointer;
+  font-size: 16px !important;
+  line-height: 20px;
+}
+
+.ai-search-history-popover-item > button:first-child {
+  flex: 1;
+  min-width: 0;
+  padding: 7px 6px !important;
+  overflow: hidden;
+  border: 0 !important;
+  border-radius: var(--wxex-radius-sm) !important;
+  background: transparent !important;
+  color: var(--wxex-text-secondary) !important;
+  cursor: pointer;
+  font: inherit;
+  text-align: left;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.ai-search-history-popover-item > button:hover {
+  background: var(--wxex-brand-soft) !important;
+  color: var(--wxex-brand) !important;
+}
+
+.ai-search-history-empty {
+  display: block;
+  padding: 8px 4px 3px;
+  color: var(--wxex-text-muted);
+  font-size: 10px;
+}
+
 .ai-search-composer-row {
   display: flex;
   align-items: flex-end;
@@ -743,24 +1435,6 @@
   margin-top: 6px;
 }
 
-.ai-search-evidence-meta {
-  display: grid;
-  grid-template-columns: 34px minmax(0, 1fr);
-  gap: 4px 8px;
-  margin-bottom: 12px;
-  color: var(--wxex-text-muted);
-  font-size: 10px;
-  line-height: 15px;
-}
-
-.ai-search-evidence-meta strong {
-  overflow: hidden;
-  color: var(--wxex-text-secondary);
-  font-weight: 600;
-  text-overflow: ellipsis;
-  white-space: nowrap;
-}
-
 .ai-search-evidence-card {
   display: block;
   width: 100%;
@@ -773,6 +1447,18 @@
   cursor: pointer;
   font: inherit;
   text-align: left;
+  animation: ai-search-evidence-in 240ms ease-out both;
+}
+
+@keyframes ai-search-evidence-in {
+  from {
+    opacity: 0;
+    transform: translateX(7px);
+  }
+  to {
+    opacity: 1;
+    transform: translateX(0);
+  }
 }
 
 .ai-search-evidence-card:hover,
@@ -781,6 +1467,20 @@
   background: #f4fbf8;
 }
 
+@media (prefers-reduced-motion: reduce) {
+  .ai-search-spinner,
+  .ai-search-result,
+  .ai-search-summary-block,
+  .ai-search-answer > *,
+  .ai-search-evidence-card,
+  .ai-search-history-popover,
+  .ai-search-pipeline-step.active .ai-search-pipeline-mark,
+  .ai-search-sync-progress-track span {
+    animation: none !important;
+    transition: none !important;
+  }
+}
+
 .ai-search-evidence-card-top,
 .ai-search-evidence-conversation,
 .ai-search-evidence-text,

src/shared/ai-search.ts

@@ -0,0 +1,625 @@
+import type { KnowledgeEvidence, KnowledgeSearchIpcResult } from './knowledge'
+
+export type AiSearchScope = 'global' | 'groups' | 'contacts' | 'conversation'
+export type AiSearchRange = 'today' | '7d' | '30d' | 'all'
+/**
+ * Retrieval semantics, not presentation labels. Each intent has a constrained
+ * execution path in the main process; a model must not be able to quietly turn
+ * an identity lookup into a generic message-keyword search.
+ */
+export type AiSearchIntent =
+  | 'conversation_recall'
+  | 'conversation_topic_search'
+  | 'global_topic_search'
+  | 'conversation_name_search'
+  | 'general'
+
+export interface AiSearchTimeRange {
+  /** Unix seconds. Undefined start means the user explicitly allowed all history. */
+  startTime?: number
+  endTime?: number
+  label: string
+  reason: string
+  source: 'ui' | 'query' | 'user_retry'
+}
+export type AiSearchProgressStage =
+  | 'query_understanding'
+  | 'agent_start'
+  | 'agent_tool'
+  | 'agent_decision'
+  | 'search_plan_ready'
+  | 'knowledge_searching'
+  | 'evidence_ranking'
+  | 'evidence_ready'
+  | 'aggregation'
+  | 'ai_generating'
+  | 'completed'
+  | 'error'
+export type AiSearchProgressStatus = 'running' | 'completed' | 'error'
+
+export interface AiSearchPlan {
+  intent: AiSearchIntent
+  keywords: string[]
+  variants: string[]
+  source: 'local' | 'ai' | 'hybrid'
+  scopeLabel: string
+  rangeLabel: string
+  timeRange: AiSearchTimeRange
+  contactNames: string[]
+  /** A user-supplied identity candidate. It must be resolved by Contact Resolution. */
+  contactQuery?: string
+  /** The message-content query, never a contact display name. */
+  topicQuery?: string
+}
+
+export interface AiSearchPipelineRequest {
+  requestId: string
+  text: string
+  scope: AiSearchScope
+  range: AiSearchRange
+  conversationId?: string
+  /** Explicit user retry takes precedence over natural-language inference. */
+  timeRangeOverride?: AiSearchTimeRange
+}
+
+export interface AiSearchProgressEvent {
+  requestId: string
+  stage: AiSearchProgressStage
+  status: AiSearchProgressStatus
+  message: string
+  plan?: AiSearchPlan
+  stats?: {
+    knowledgeMessageCount?: number
+    matchedMessages?: number
+    evidenceCount?: number
+    contextEvidenceCount?: number
+    tokenEstimate?: number
+    inputTokens?: number
+    inputTokensEstimated?: boolean
+    elapsedMs?: number
+    deduplicatedMessages?: number
+    peopleCount?: number
+    conversationCount?: number
+  }
+  timings?: AiSearchPipelineTimings
+  modelName?: string
+  agentTrace?: AiSearchAgentTraceItem
+  error?: string
+}
+
+export type AiSearchAgentToolName =
+  | 'search_conversations'
+  | 'search_people'
+  | 'search_messages'
+  | 'get_conversation_messages'
+  | 'get_messages_by_time'
+  | 'get_message_context'
+
+export type AiSearchAgentTraceEvent =
+  | 'agentStart'
+  | 'toolCallStart'
+  | 'toolCallEnd'
+  | 'agentDecision'
+  | 'evidenceBuild'
+  | 'summaryStart'
+  | 'summaryEnd'
+  | 'fallback'
+
+/** Public trace: deliberately contains no SQL, paths, raw IDs, or Worker details. */
+export interface AiSearchAgentTraceItem {
+  sequence: number
+  event: AiSearchAgentTraceEvent
+  label: string
+  toolName?: AiSearchAgentToolName
+  /** Sanitized, human-readable arguments only. */
+  arguments?: Record<string, string | number | boolean>
+  resultCount?: number
+  elapsedMs?: number
+  decision?: string
+  /** Bounded local snapshot of the exact decision prompt; never sent to analytics. */
+  decisionInput?: string
+}
+
+export interface AiSearchAgentRun {
+  mode: 'agent' | 'fallback'
+  toolCalls: number
+  trace: AiSearchAgentTraceItem[]
+  fallbackReason?: string
+}
+
+export interface AiSearchPipelineEvidence extends KnowledgeEvidence {
+  conversationName: string
+  conversationType: 'user' | 'group'
+}
+
+/** A program-generated, stable citation. This is the only Evidence shape sent to AI/UI. */
+export interface AiSearchFinalEvidence extends AiSearchPipelineEvidence {
+  id: `E${number}`
+}
+
+export interface AiSearchPersonAggregation {
+  id: string
+  name: string
+  messageCount: number
+  conversationCount: number
+  lastMessageAt: number
+  evidenceIds: Array<`E${number}`>
+}
+
+export interface AiSearchConversationAggregation {
+  id: string
+  name: string
+  type: 'user' | 'group'
+  messageCount: number
+  peopleCount: number
+  lastMessageAt: number
+  evidenceIds: Array<`E${number}`>
+}
+
+export interface AiSearchAggregation {
+  messageCount: number
+  peopleCount: number
+  conversationCount: number
+  people: AiSearchPersonAggregation[]
+  conversations: AiSearchConversationAggregation[]
+}
+
+/** All fields are directly measured around real work. */
+export interface AiSearchPipelineTimings {
+  queryUnderstandingMs: number
+  contactResolutionMs: number
+  knowledgeSearchMs: number
+  workerIpcMs: number
+  workerBootMs: number
+  dispatchMs: number
+  workerSqlMs: number
+  responseSerializeMs: number
+  responseTransferMs: number
+  ftsMs: number
+  chunkExpandMs: number
+  messageLoadMs: number
+  rankingMs: number
+  candidateRankingMs: number
+  evidenceBuildMs: number
+  aggregationMs: number
+  contextPreparationMs: number
+  agentDecisionMs: number
+  agentToolMs: number
+  aiGenerationMs: number
+  totalMs: number
+}
+
+export interface AiSearchCitationValidation {
+  status: 'valid' | 'sanitized'
+  invalidCitationIds: string[]
+}
+
+/** Truthful retrieval metadata shared by the AI, UI and diagnostics. */
+export interface AiSearchRetrievalContract {
+  intent: AiSearchIntent
+  conversationId?: string
+  timeRange: AiSearchTimeRange
+  retrievalMode:
+    | 'conversation_metadata'
+    | 'conversation_topic_fts'
+    | 'global_fts'
+    | 'conversation_name'
+    | 'unresolved_identity'
+  candidateCount: number
+  sourceMessageCount?: number
+  sourceCoverage: 'complete' | 'partial' | 'keyword_match' | 'unknown'
+  isComplete: boolean
+  fallbackUsed: boolean
+  fallbackReason?: string
+  suspicious: boolean
+}
+
+export interface AiSearchPipelineResult {
+  requestId: string
+  status: 'completed' | 'no_evidence' | 'retrieval_incomplete' | 'ai_failed' | 'failed'
+  plan: AiSearchPlan
+  knowledge: Pick<
+    KnowledgeSearchIpcResult,
+    | 'source'
+    | 'state'
+    | 'fallbackReason'
+    | 'indexedMessageCount'
+    | 'indexedChunkCount'
+    | 'totalMessages'
+  >
+  candidateEvidenceCount: number
+  retrieval: AiSearchRetrievalContract
+  evidence: AiSearchFinalEvidence[]
+  contextEvidenceCount: number
+  aggregation: AiSearchAggregation
+  agent: AiSearchAgentRun
+  citationValidation?: AiSearchCitationValidation
+  timings: AiSearchPipelineTimings
+  answer?: string
+  ai?: {
+    providerName: string
+    modelName: string
+    inputTokens?: number
+    inputTokensEstimated: boolean
+  }
+  error?: string
+  errorStage?: Exclude<AiSearchProgressStage, 'completed' | 'error'>
+  elapsedMs: number
+}
+
+const RANGE_LABELS: Record<AiSearchRange, string> = {
+  today: '今天',
+  '7d': '近 7 天',
+  '30d': '近 30 天',
+  all: '全部历史'
+}
+
+const SEARCH_INTENT_PHRASES = [
+  '全局搜一下',
+  '全局搜索',
+  '搜索一下',
+  '搜一下',
+  '查询一下',
+  '查一下',
+  '找一下',
+  '我和谁聊过',
+  '谁和我聊过',
+  '谁聊过',
+  '哪些人和我聊过',
+  '最近讨论了什么',
+  '最近聊了什么',
+  '最近说了什么',
+  '讨论了什么',
+  '讨论什么',
+  '聊了什么',
+  '聊些什么',
+  '说了什么',
+  '说些什么',
+  '最近讨论',
+  '最近聊天',
+  '这个话题',
+  '相关话题',
+  '的聊天',
+  '的内容',
+  '的记录',
+  '关于',
+  '聊天',
+  '记录',
+  '聊天记录',
+  '帮我',
+  '请问',
+  '最近'
+].sort((left, right) => right.length - left.length)
+
+const SEARCH_STOP_WORDS = new Set([
+  '我',
+  '谁',
+  '什么',
+  '哪些',
+  '哪个',
+  '人',
+  '和',
+  '聊过',
+  '说过',
+  '提到',
+  '讨论',
+  '聊天',
+  '记录',
+  '说',
+  '聊',
+  '话题',
+  '内容',
+  '相关',
+  '最近',
+  '一下'
+])
+
+export const aiSearchRangeLabel = (range: AiSearchRange): string => RANGE_LABELS[range]
+
+export const aiSearchRangeStart = (range: AiSearchRange): number | undefined => {
+  if (range === 'all') return undefined
+  if (range === 'today') {
+    const now = new Date()
+    return Math.floor(new Date(now.getFullYear(), now.getMonth(), now.getDate()).getTime() / 1000)
+  }
+  return Math.floor(Date.now() / 1000) - (range === '7d' ? 7 : 30) * 86400
+}
+
+const dayStart = (date: Date): number =>
+  Math.floor(new Date(date.getFullYear(), date.getMonth(), date.getDate()).getTime() / 1000)
+
+const currentYearStart = (date: Date): number =>
+  Math.floor(new Date(date.getFullYear(), 0, 1).getTime() / 1000)
+
+const currentMonthStart = (date: Date): number =>
+  Math.floor(new Date(date.getFullYear(), date.getMonth(), 1).getTime() / 1000)
+
+const CHINESE_NUMBERS: Record<string, number> = {
+  一: 1,
+  二: 2,
+  两: 2,
+  三: 3,
+  四: 4,
+  五: 5,
+  六: 6,
+  七: 7,
+  八: 8,
+  九: 9,
+  十: 10
+}
+
+const parseNaturalNumber = (value: string | undefined): number | undefined => {
+  if (!value) return undefined
+  const numeric = Number(value)
+  if (Number.isFinite(numeric)) return numeric
+  return CHINESE_NUMBERS[value]
+}
+
+/**
+ * Query time expressions are part of SearchPlan, never a renderer-only rule.
+ * A natural-language time constraint is more specific than the broad "all" UI scope.
+ */
+export const inferAiSearchTimeRange = (
+  query: string,
+  uiRange: AiSearchRange,
+  now = new Date(),
+  override?: AiSearchTimeRange
+): AiSearchTimeRange => {
+  if (override?.source === 'user_retry') return override
+  const nowSeconds = Math.floor(now.getTime() / 1000)
+  const fromQuery = (startTime: number, label: string, reason: string): AiSearchTimeRange => ({
+    startTime,
+    endTime: nowSeconds,
+    label,
+    reason,
+    source: 'query'
+  })
+  const recentDays = query.match(/最近\s*(\d{1,3}|[一二两三四五六七八九十])\s*天/)
+  if (recentDays) {
+    const days = Math.max(1, Math.min(365, parseNaturalNumber(recentDays[1]) || 30))
+    return fromQuery(nowSeconds - days * 86400, `近 ${days} 天`, `用户说“最近 ${days} 天”`)
+  }
+  const recentMonths = query.match(/最近\s*(\d{1,2}|[一二两三四五六七八九十])\s*个?月/)
+  if (recentMonths) {
+    const months = Math.max(1, Math.min(24, parseNaturalNumber(recentMonths[1]) || 1))
+    const start = new Date(now.getFullYear(), now.getMonth() - months, now.getDate()).getTime()
+    return fromQuery(Math.floor(start / 1000), `近 ${months} 个月`, `用户说“最近 ${months} 个月”`)
+  }
+  if (/刚刚|刚才/.test(query))
+    return fromQuery(nowSeconds - 24 * 3600, '近 24 小时', '用户说“刚刚”')
+  if (/这几天/.test(query)) return fromQuery(nowSeconds - 7 * 86400, '近 7 天', '用户说“这几天”')
+  if (/这周|本周/.test(query)) {
+    const weekday = now.getDay() || 7
+    return fromQuery(dayStart(now) - (weekday - 1) * 86400, '本周', '用户说“这周”')
+  }
+  if (/这个月|本月/.test(query)) return fromQuery(currentMonthStart(now), '本月', '用户说“这个月”')
+  if (/上个月/.test(query)) {
+    const start = Math.floor(new Date(now.getFullYear(), now.getMonth() - 1, 1).getTime() / 1000)
+    const end = Math.floor(new Date(now.getFullYear(), now.getMonth(), 1).getTime() / 1000) - 1
+    return {
+      startTime: start,
+      endTime: end,
+      label: '上个月',
+      reason: '用户说“上个月”',
+      source: 'query'
+    }
+  }
+  if (/今年/.test(query)) return fromQuery(currentYearStart(now), '今年', '用户说“今年”')
+  if (/最近/.test(query)) return fromQuery(nowSeconds - 30 * 86400, '近 30 天', '用户说“最近”')
+  return {
+    startTime: aiSearchRangeStart(uiRange),
+    endTime: undefined,
+    label: aiSearchRangeLabel(uiRange),
+    reason: '使用界面选择的时间范围',
+    source: 'ui'
+  }
+}
+
+export const aiSearchIntentLabel = (intent: AiSearchIntent): string => {
+  if (intent === 'conversation_recall') return '回顾最近聊天'
+  if (intent === 'conversation_topic_search') return '在指定聊天中查找话题'
+  if (intent === 'global_topic_search') return '按话题查找'
+  if (intent === 'conversation_name_search') return '查找聊天'
+  return '综合查找'
+}
+
+export const aiSearchScopeLabel = (scope: AiSearchScope, conversationName?: string): string => {
+  if (scope === 'groups') return '群聊'
+  if (scope === 'contacts') return '单聊'
+  if (scope === 'conversation') return conversationName || '当前会话'
+  return '所有聊天'
+}
+
+const normalizeTerms = (terms: unknown): string[] => {
+  if (!Array.isArray(terms)) return []
+  return Array.from(
+    new Set(
+      terms
+        .filter((term): term is string => typeof term === 'string')
+        .map((term) => term.trim())
+        .filter((term) => term.length >= 2 && term.length <= 32)
+    )
+  ).slice(0, 16)
+}
+
+const extractKeywords = (query: string): string[] => {
+  const cleaned = SEARCH_INTENT_PHRASES.reduce(
+    (value, phrase) => value.split(phrase).join(' '),
+    query.toLowerCase()
+  )
+  return Array.from(
+    new Set(
+      cleaned
+        .split(/[\s,，。！？!?、:：;；"“”‘’()（）[\]【】]+/)
+        .map((token) => token.trim())
+        .filter((token) => token.length >= 2 && !SEARCH_STOP_WORDS.has(token))
+    )
+  )
+}
+
+const keywordVariants = (keywords: string[]): string[] =>
+  Array.from(
+    new Set(
+      keywords.flatMap((keyword) => {
+        const variants = [keyword]
+        if (/^[\u4e00-\u9fff]+$/.test(keyword) && keyword.length > 2) {
+          variants.push(keyword.slice(-2))
+        }
+        return variants
+      })
+    )
+  )
+
+export const buildLocalAiSearchPlan = (
+  query: string
+): Pick<
+  AiSearchPlan,
+  'intent' | 'keywords' | 'variants' | 'source' | 'contactQuery' | 'topicQuery'
+> => {
+  const keywords = extractKeywords(query)
+  const normalized = query.replace(/[“”"'‘’「」『』]/g, '').trim()
+  const recall = normalized.match(
+    /(?:我和|我跟|我与)\s*(.+?)\s*(?:最近|这几天|本周|这个月|本月|今年|上个月|刚刚|刚才)?\s*(?:聊了什么|聊过什么|说了什么|谈了什么)/
+  )
+  const reverseRecall = normalized.match(
+    /^\s*(.+?)\s*(?:最近)?(?:跟我|和我|与我)\s*(?:聊了什么|聊过什么|说了什么|谈了什么)/
+  )
+  const namedConversationRecall = normalized.match(
+    /(?:我在|在)\s*(.+?)\s*(?:最近)?\s*(?:聊了什么|聊过什么|说了什么|谈了什么)/
+  )
+  const conversationTopic = normalized.match(
+    /(?:我和|我跟|我与)\s*(.+?)\s*(?:最近|这几天|本周|这个月|本月|今年|上个月)?\s*(?:聊过|提过|说过|讨论过)\s*(.+?)(?:吗|么|沒有|没有)?[？?。！!]*$/
+  )
+  const globalTopic = normalized.match(
+    /(?:最近|这几天|本周|这个月|本月|今年)?\s*(?:谁|哪些人|大家)\s*(?:聊过|提过|说过|讨论过)\s*(.+?)[？?。！!]*$/
+  )
+  const conversationName =
+    !recall &&
+    !reverseRecall &&
+    !conversationTopic &&
+    !namedConversationRecall &&
+    !globalTopic &&
+    /^[^，,。！？!?]{2,32}(?:群|群聊|交流群)$/.test(normalized)
+      ? normalized
+      : undefined
+  const contactQuery = (
+    conversationTopic?.[1] ||
+    recall?.[1] ||
+    reverseRecall?.[1] ||
+    namedConversationRecall?.[1]
+  )
+    ?.replace(/^(?:和|跟|与)\s*/, '')
+    .trim()
+  const topicQuery = (conversationTopic?.[2] || globalTopic?.[1])
+    ?.replace(/^(?:关于|一下|吗|么)\s*/, '')
+    .trim()
+  const intent: AiSearchIntent = conversationTopic
+    ? 'conversation_topic_search'
+    : recall || reverseRecall
+      ? 'conversation_recall'
+      : namedConversationRecall
+        ? 'conversation_name_search'
+        : globalTopic
+          ? 'global_topic_search'
+          : conversationName
+            ? 'conversation_name_search'
+            : keywords.length
+              ? 'global_topic_search'
+              : 'general'
+  const effectiveKeywords = topicQuery ? [topicQuery] : keywords
+  return {
+    intent,
+    keywords: effectiveKeywords,
+    variants: keywordVariants(effectiveKeywords),
+    source: 'local',
+    contactQuery: contactQuery || conversationName,
+    topicQuery
+  }
+}
+
+export const parseAiSearchPlan = (
+  value: string
+): Partial<Pick<AiSearchPlan, 'intent' | 'keywords' | 'variants' | 'topicQuery'>> | null => {
+  const jsonMatch = value.match(/\{[\s\S]*\}/)
+  if (!jsonMatch) return null
+  try {
+    const parsed = JSON.parse(jsonMatch[0]) as Record<string, unknown>
+    const intent = [
+      'general',
+      'conversation_recall',
+      'conversation_topic_search',
+      'global_topic_search',
+      'conversation_name_search'
+    ].includes(String(parsed.intent))
+      ? (parsed.intent as AiSearchIntent)
+      : undefined
+    return {
+      intent,
+      keywords: normalizeTerms(parsed.keywords),
+      variants: normalizeTerms(parsed.variants),
+      topicQuery:
+        typeof parsed.topicQuery === 'string' && parsed.topicQuery.trim().length >= 2
+          ? parsed.topicQuery.trim().slice(0, 64)
+          : undefined
+    }
+  } catch {
+    return null
+  }
+}
+
+export const mergeAiSearchPlans = (
+  local: Pick<
+    AiSearchPlan,
+    'intent' | 'keywords' | 'variants' | 'source' | 'contactQuery' | 'topicQuery'
+  >,
+  ai: Partial<Pick<AiSearchPlan, 'intent' | 'keywords' | 'variants' | 'topicQuery'>> | null
+): Pick<
+  AiSearchPlan,
+  'intent' | 'keywords' | 'variants' | 'source' | 'contactQuery' | 'topicQuery'
+> => {
+  if (!ai) return local
+  const keywords = normalizeTerms([...local.keywords, ...(ai.keywords || [])])
+  const variants = normalizeTerms([
+    ...keywordVariants(keywords),
+    ...local.variants,
+    ...(ai.variants || [])
+  ])
+  // Identity-bearing local intents are deterministic contracts. A planner may
+  // refine topic terms but may not weaken them into an unrelated FTS intent.
+  const lockedIntent =
+    local.intent === 'conversation_recall' ||
+    local.intent === 'conversation_topic_search' ||
+    local.intent === 'conversation_name_search'
+  return {
+    intent: lockedIntent ? local.intent : ai.intent || local.intent,
+    keywords,
+    variants,
+    source: 'hybrid',
+    contactQuery: local.contactQuery,
+    topicQuery: local.topicQuery || ai.topicQuery
+  }
+}
+
+export const includesExplicitAiSearchAlias = (query: string, alias: string): boolean => {
+  const name = alias.trim()
+  if (!name || name.length < 2) return false
+  const escaped = name.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+  const quoted = new RegExp(`[“"'‘「『]${escaped}[”"'’」』]`)
+  const relational = new RegExp(
+    `(?:我和|我跟|我与|和|跟|与|在|给|向|@)${escaped}(?=$|[\\s，,。！？!?、:：；;])`
+  )
+  if (quoted.test(query) || relational.test(query)) return true
+
+  // Users often omit a nickname's punctuation, for example typing
+  // “中田健身弘毅” for “中田健身-弘毅”. Keep this tolerant matching limited
+  // to an explicit relational query so a short alias cannot accidentally
+  // select a contact from unrelated prose.
+  const compact = (value: string): string =>
+    value.toLocaleLowerCase().replace(/[^\p{L}\p{N}]+/gu, '')
+  const compactName = compact(name)
+  return (
+    compactName.length >= 2 &&
+    /(?:我和|我跟|我与|和|跟|与|在|给|向|@)/.test(query) &&
+    compact(query).includes(compactName)
+  )
+}

src/shared/cache.ts

@@ -1,7 +1,7 @@
-export type CacheClearScope = 'bootstrap' | 'electron' | 'all'
+export type CacheClearScope = 'bootstrap' | 'electron' | 'knowledge' | 'all'
 
 export interface CacheSummaryItem {
-  id: 'bootstrap' | 'electron'
+  id: 'bootstrap' | 'electron' | 'knowledge'
   label: string
   description: string
   sizeBytes: number

src/shared/contact-resolution.ts

@@ -0,0 +1,38 @@
+export type ContactResolutionMatch = 'exact' | 'normalized' | 'alias' | 'fuzzy'
+
+export interface ContactResolutionCandidate {
+  conversationId: string
+  displayName: string
+  matchedBy: ContactResolutionMatch
+  confidence: number
+}
+
+export interface ContactResolutionResult {
+  matched: boolean
+  personId?: string
+  conversationId?: string
+  canonicalName?: string
+  displayName?: string
+  matchedBy?: ContactResolutionMatch
+  confidence: number
+  candidates: ContactResolutionCandidate[]
+  ambiguous: boolean
+}
+
+/**
+ * Identity-only canonicalization. It deliberately does not use substring
+ * matching: callers may use a separate UI-filter policy for broad discovery,
+ * but identity resolution must never turn 张三 into 张三丰.
+ */
+export const normalizeContactName = (value: string): string =>
+  String(value || '')
+    .normalize('NFKC')
+    .toLocaleLowerCase()
+    .replace(/[\p{White_Space}\p{P}\p{S}_]+/gu, '')
+
+export const emptyContactResolution = (): ContactResolutionResult => ({
+  matched: false,
+  confidence: 0,
+  candidates: [],
+  ambiguous: false
+})

src/shared/knowledge.ts

@@ -0,0 +1,344 @@
+/**
+ * Contracts for the local, derived knowledge base. These values deliberately
+ * contain no WCDB handles, Electron objects, database keys, or UI state so the
+ * indexer can run in an isolated process.
+ */
+
+export const KNOWLEDGE_SCHEMA_VERSION = 1
+export const DEFAULT_CHUNKER_VERSION = 'conversation-v1'
+
+export type KnowledgeMessageKind = 'text' | 'voice' | 'file' | 'link' | 'system' | 'other'
+export type KnowledgeIndexPhase =
+  | 'idle'
+  | 'preflight'
+  | 'indexing'
+  | 'ready'
+  | 'cancelled'
+  | 'error'
+export type KnowledgeTemporalIntent = 'none' | 'current' | 'historical' | 'timeline'
+export type KnowledgeFtsTokenizer = 'unicode61' | 'trigram'
+export type KnowledgeFtsContentMode = 'external' | 'internal'
+export type KnowledgeFtsDetail = 'full' | 'column' | 'none'
+
+export interface KnowledgeAttachmentMetadata {
+  name: string
+  kind?: 'file' | 'link' | 'image' | 'video' | 'other'
+  url?: string
+  sizeBytes?: number
+}
+
+/** A read-only source record prepared by the future WCDB adapter. */
+export interface KnowledgeSourceMessage {
+  accountId: string
+  conversationId: string
+  messageId: string
+  /** Unix epoch milliseconds. Adapters must convert source-specific units. */
+  createTime: number
+  senderId?: string
+  senderName?: string
+  kind: KnowledgeMessageKind
+  text?: string
+  attachment?: KnowledgeAttachmentMetadata
+  voiceTranscript?: string
+}
+
+export interface KnowledgeNormalizedMessage extends KnowledgeSourceMessage {
+  searchableText: string
+  contentHash: string
+}
+
+export interface KnowledgeChunkerConfig {
+  version: string
+  maxGapMs: number
+  maxMessages: number
+  maxCharacters: number
+  overlapMessages: number
+}
+
+export interface KnowledgeChunk {
+  chunkId: string
+  accountId: string
+  conversationId: string
+  startTime: number
+  endTime: number
+  text: string
+  messageIds: string[]
+  participantIds: string[]
+  messageKinds: KnowledgeMessageKind[]
+  contentHash: string
+  chunkerVersion: string
+}
+
+/**
+ * Every FTS choice is explicit. The first production profile must be selected
+ * from the Task 0 report rather than being silently hard-coded in the UI.
+ */
+export interface KnowledgeFtsConfig {
+  profileId: string
+  tokenizer: KnowledgeFtsTokenizer
+  contentMode: KnowledgeFtsContentMode
+  detail: KnowledgeFtsDetail
+  columnsize: 0 | 1
+}
+
+/**
+ * Chosen after the realistic desensitized WeChat benchmark: trigram preserves
+ * Chinese-substring recall while external content avoids a second text copy.
+ */
+export const DEFAULT_KNOWLEDGE_FTS_CONFIG: KnowledgeFtsConfig = {
+  profileId: 'trigram-external-full-columnsize-v1',
+  tokenizer: 'trigram',
+  contentMode: 'external',
+  detail: 'full',
+  columnsize: 1
+}
+
+export interface KnowledgeConversationInput {
+  conversationId: string
+  /** true means this is a complete read-only snapshot of the conversation. */
+  completeSnapshot: boolean
+  messages: KnowledgeSourceMessage[]
+}
+
+export interface KnowledgeIndexRequest {
+  accountId: string
+  databaseRoot: string
+  conversations: KnowledgeConversationInput[]
+  chunker: KnowledgeChunkerConfig
+  fts: KnowledgeFtsConfig
+  /** Written only after a complete source pass; used for truthful coverage. */
+  sourceMessageCount?: number
+}
+
+export interface KnowledgeIndexProgress {
+  accountId: string
+  phase: KnowledgeIndexPhase
+  conversationId?: string
+  processedMessages: number
+  totalMessages: number
+  indexedChunks: number
+  error?: string
+}
+
+export interface KnowledgeIndexResult {
+  accountId: string
+  processedMessages: number
+  indexedChunks: number
+  updatedChunks: number
+  unchangedConversations: number
+  databaseBytes: number
+  walBytes: number
+  elapsedMs: number
+  cancelled: boolean
+}
+
+export interface KnowledgeCapacityPreflightRequest {
+  accountId: string
+  databaseRoot: string
+  conversations: KnowledgeConversationInput[]
+  chunker: KnowledgeChunkerConfig
+  /** Optional free space supplied by the platform layer; this module never probes WCDB paths. */
+  availableDiskBytes?: number
+}
+
+export interface KnowledgeCapacityPreflight {
+  accountId: string
+  sourceMessageCount: number
+  indexableMessageCount: number
+  indexableTextBytes: number
+  voiceTranscriptCount: number
+  attachmentMetadataCount: number
+  sampledChunkCount: number
+  estimatedChunkCount: number
+  estimatedDatabaseBytesLow: number
+  estimatedDatabaseBytesHigh: number
+  estimatedBuildPeakBytesLow: number
+  estimatedBuildPeakBytesHigh: number
+  availableDiskBytes?: number
+  hasSufficientDiskSpace?: boolean
+  warnings: string[]
+}
+
+export interface KnowledgeEvidence {
+  chunkId: string
+  conversationId: string
+  startTime: number
+  endTime: number
+  /** Stable source-message identity used by the archive jump action. */
+  messageId: string
+  senderId?: string
+  sender: string
+  /** Unix epoch milliseconds. */
+  timestamp: number
+  messageIds: string[]
+  text: string
+  score?: number
+}
+
+/** A bounded, local summary of a single conversation retrieval. */
+export interface KnowledgeConversationRetrieval {
+  conversationId: string
+  totalMessages: number
+  chunkCount: number
+  candidateMessages: number
+  systemMessagesDeprioritized: number
+  complete: boolean
+}
+
+export interface KnowledgeQuery {
+  accountId: string
+  text: string
+  /** Query-router terms. The raw question remains available for diagnostics. */
+  terms?: string[]
+  limit: number
+  conversationId?: string
+  conversationIds?: string[]
+  senderIds?: string[]
+  /** Unix epoch milliseconds. */
+  startTime?: number
+  /** Unix epoch milliseconds. */
+  endTime?: number
+  temporalIntent?: KnowledgeTemporalIntent
+}
+
+export interface KnowledgeSearchRequest extends KnowledgeQuery {
+  databaseRoot: string
+  fts: KnowledgeFtsConfig
+}
+
+export type KnowledgeSearchState = 'unavailable' | 'indexing' | 'ready'
+
+/** Measured in the Worker; never inferred from message counts or UI timers. */
+export interface KnowledgeSearchTimings {
+  /** Parent/child-process transport and host scheduling outside SQLite work. */
+  workerIpcMs: number
+  /** First request only: child process spawn and Node initialization until it received the request. */
+  workerBootMs: number
+  /** Parent send → Worker handler start. */
+  dispatchMs: number
+  /** Worker local SQLite/chunk work; equals the Worker-side search total. */
+  workerSqlMs: number
+  /** Worker response preparation → parent receipt; includes IPC serialization/transfer. */
+  responseTransferMs: number
+  /** Worker-side serialization preflight for the result payload. */
+  responseSerializeMs: number
+  /** FTS (or short-term database lookup) query time. */
+  ftsMs: number
+  /** Reading source message rows from matching chunks. */
+  messageLoadMs: number
+  /** Expanding chunk members, scoring terms and per-chunk de-duplication. */
+  chunkExpandMs: number
+  /** Final result ordering and limit application. */
+  rankingMs: number
+  /** Worker-side local search total. */
+  totalMs: number
+}
+
+export const emptyKnowledgeSearchTimings = (): KnowledgeSearchTimings => ({
+  workerIpcMs: 0,
+  workerBootMs: 0,
+  dispatchMs: 0,
+  workerSqlMs: 0,
+  responseTransferMs: 0,
+  responseSerializeMs: 0,
+  ftsMs: 0,
+  messageLoadMs: 0,
+  chunkExpandMs: 0,
+  rankingMs: 0,
+  totalMs: 0
+})
+
+export interface KnowledgeSearchResult {
+  state: KnowledgeSearchState
+  evidence: KnowledgeEvidence[]
+  indexedMessageCount: number
+  indexedChunkCount: number
+  timings: KnowledgeSearchTimings
+  conversationRetrieval?: KnowledgeConversationRetrieval
+}
+
+/** Renderer-facing request. Chat timestamps use Unix seconds in the existing UI. */
+export interface KnowledgeSearchIpcRequest {
+  text: string
+  terms: string[]
+  conversationIds?: string[]
+  senderIds?: string[]
+  startTime?: number
+  endTime?: number
+  limit?: number
+}
+
+export interface KnowledgeSearchIpcResult extends KnowledgeSearchResult {
+  source: 'knowledge' | 'fallback'
+  totalMessages: number
+  fallbackReason?: 'unavailable' | 'indexing' | 'error'
+}
+
+export type KnowledgeRuntimeState = 'unavailable' | 'building' | 'syncing' | 'ready' | 'error'
+
+export interface KnowledgeRuntimeStatus {
+  accountId: string
+  state: KnowledgeRuntimeState
+  indexedMessageCount: number
+  indexedChunkCount: number
+  /** Null means this source pass has not yet counted every source message. */
+  sourceMessageCount: number | null
+  processedMessages: number
+  totalMessages: number | null
+  currentConversationId?: string
+  /** Null is displayed as unavailable rather than a fabricated ETA. */
+  estimatedRemainingMs: number | null
+  databaseBytes: number
+  walBytes: number
+  shmBytes: number
+  lastError?: string
+}
+
+export interface KnowledgeStatusRequest {
+  accountId: string
+  databaseRoot: string
+  fts: KnowledgeFtsConfig
+}
+
+export interface KnowledgeWorkerRequest {
+  version: 1
+  type: 'index' | 'preflight' | 'search' | 'status' | 'remove' | 'cancel' | 'close'
+  requestId: string
+  /** Parent monotonic wall-clock used only for transport timing. */
+  sentAt?: number
+  payload:
+    | KnowledgeIndexRequest
+    | KnowledgeCapacityPreflightRequest
+    | KnowledgeSearchRequest
+    | KnowledgeStatusRequest
+    | { accountId: string; databaseRoot: string }
+    | { targetRequestId: string }
+    | Record<string, never>
+}
+
+export interface KnowledgeWorkerResponse {
+  version: 1
+  type: 'progress' | 'result' | 'error'
+  requestId: string
+  payload?:
+    | KnowledgeIndexProgress
+    | KnowledgeIndexResult
+    | KnowledgeCapacityPreflight
+    | KnowledgeSearchResult
+    | KnowledgeRuntimeStatus
+    | { removed: true }
+  error?: string
+  transport?: {
+    workerReceivedAt: number
+    workerCompletedAt: number
+    responseSerializeMs: number
+  }
+}
+
+export const DEFAULT_KNOWLEDGE_CHUNKER: KnowledgeChunkerConfig = {
+  version: DEFAULT_CHUNKER_VERSION,
+  maxGapMs: 10 * 60 * 1000,
+  maxMessages: 12,
+  maxCharacters: 1200,
+  overlapMessages: 3
+}

tests/benchmark/knowledge-benchmark.test.ts

@@ -0,0 +1,217 @@
+import { mkdtempSync } from 'fs'
+import { rm } from 'fs/promises'
+import { tmpdir } from 'os'
+import { join } from 'path'
+import { performance } from 'perf_hooks'
+import { afterAll, describe, expect, it } from 'vitest'
+import {
+  DEFAULT_KNOWLEDGE_CHUNKER,
+  type KnowledgeEvidence,
+  type KnowledgeFtsConfig,
+  type KnowledgeSourceMessage
+} from '../../src/shared/knowledge'
+import { KnowledgeStore } from '../../src/main/knowledge/knowledge-store'
+import {
+  createKnowledgeBenchmarkFixture,
+  type KnowledgeBenchmarkCase
+} from '../fixtures/knowledge-rag'
+
+const root = mkdtempSync(join(tmpdir(), 'wxe-knowledge-benchmark-'))
+
+afterAll(async () => {
+  await rm(root, { recursive: true, force: true })
+})
+
+type Retrieval = { messageIds: string[] }
+type Metrics = {
+  recallAt5: number
+  recallAt10: number
+  mrr: number
+  evidenceAccuracy: number
+  findSuccessAt10: number
+  p50LatencyMs: number
+  p95LatencyMs: number
+  averageInputTokens: number
+}
+
+function percentile(values: number[], percentileValue: number): number {
+  if (!values.length) return 0
+  const sorted = values.slice().sort((left, right) => left - right)
+  return sorted[Math.min(sorted.length - 1, Math.ceil(sorted.length * percentileValue) - 1)]
+}
+
+function estimateInputTokens(messages: KnowledgeSourceMessage[]): number {
+  const chars = messages.reduce((total, message) => total + (message.text || '').length, 0)
+  // Conservative Chinese-oriented baseline: question/system metadata plus the selected old-search context.
+  return 1_000 + Math.ceil(chars / 2)
+}
+
+function oldSearch(
+  messages: KnowledgeSourceMessage[],
+  testCase: KnowledgeBenchmarkCase
+): Retrieval[] {
+  const normalizedTerms = testCase.oldSearchTerms.map((term) => term.toLowerCase())
+  return messages
+    .map((message) => {
+      const text =
+        `${message.text || ''}\n${message.voiceTranscript || ''}\n${message.attachment?.name || ''}`.toLowerCase()
+      const score = normalizedTerms.reduce(
+        (total, term) => total + (text.includes(term) ? 1 : 0),
+        0
+      )
+      return { messageIds: [message.messageId], score }
+    })
+    .filter((item) => item.score > 0)
+    .sort((left, right) => right.score - left.score)
+}
+
+function scoreCases(
+  cases: KnowledgeBenchmarkCase[],
+  search: (testCase: KnowledgeBenchmarkCase) => Retrieval[],
+  inputTokens: (testCase: KnowledgeBenchmarkCase) => number
+): Metrics {
+  const latency: number[] = []
+  let recallAt5 = 0
+  let recallAt10 = 0
+  let reciprocalRank = 0
+  let evidenceAccuracy = 0
+  let findSuccessAt10 = 0
+  let totalInputTokens = 0
+  for (const testCase of cases) {
+    const started = performance.now()
+    const retrieved = search(testCase)
+    latency.push(performance.now() - started)
+    const flattened = retrieved.map((item) => item.messageIds)
+    const expected = new Set(testCase.expectedMessageIds)
+    const hitPosition = flattened.findIndex((ids) => ids.some((id) => expected.has(id)))
+    if (flattened.slice(0, 5).some((ids) => ids.some((id) => expected.has(id)))) recallAt5 += 1
+    if (hitPosition >= 0 && hitPosition < 10) {
+      recallAt10 += 1
+      findSuccessAt10 += 1
+      reciprocalRank += 1 / (hitPosition + 1)
+    }
+    const firstFive = flattened.slice(0, 5)
+    if (firstFive.length) {
+      evidenceAccuracy +=
+        firstFive.filter((ids) => ids.some((id) => expected.has(id))).length / firstFive.length
+    }
+    totalInputTokens += inputTokens(testCase)
+  }
+  return {
+    recallAt5: recallAt5 / cases.length,
+    recallAt10: recallAt10 / cases.length,
+    mrr: reciprocalRank / cases.length,
+    evidenceAccuracy: evidenceAccuracy / cases.length,
+    findSuccessAt10: findSuccessAt10 / cases.length,
+    p50LatencyMs: percentile(latency, 0.5),
+    p95LatencyMs: percentile(latency, 0.95),
+    averageInputTokens: totalInputTokens / cases.length
+  }
+}
+
+const profiles: KnowledgeFtsConfig[] = [
+  {
+    profileId: 'unicode61-external-full-columnsize',
+    tokenizer: 'unicode61',
+    contentMode: 'external',
+    detail: 'full',
+    columnsize: 1
+  },
+  {
+    profileId: 'trigram-external-full-columnsize',
+    tokenizer: 'trigram',
+    contentMode: 'external',
+    detail: 'full',
+    columnsize: 1
+  },
+  {
+    profileId: 'trigram-external-column-no-columnsize',
+    tokenizer: 'trigram',
+    contentMode: 'external',
+    detail: 'column',
+    columnsize: 0
+  },
+  {
+    profileId: 'trigram-internal-none-no-columnsize',
+    tokenizer: 'trigram',
+    contentMode: 'internal',
+    detail: 'none',
+    columnsize: 0
+  }
+]
+
+describe('desensitized local knowledge benchmark', () => {
+  it('records the 100-question Old Search baseline and FTS5 configuration comparisons', async () => {
+    const fixture = createKnowledgeBenchmarkFixture()
+    expect(fixture.cases).toHaveLength(100)
+    expect(new Set(fixture.cases.map((item) => item.category))).toEqual(
+      new Set(['fact', 'person', 'time', 'decision', 'semantic'])
+    )
+    const sourceMessages = fixture.conversations.flatMap((conversation) => conversation.messages)
+    const oldMetrics = scoreCases(
+      fixture.cases,
+      (testCase) => oldSearch(sourceMessages, testCase),
+      (testCase) => {
+        const selected = oldSearch(sourceMessages, testCase).slice(0, 240)
+        const ids = new Set(selected.flatMap((item) => item.messageIds))
+        return estimateInputTokens(sourceMessages.filter((message) => ids.has(message.messageId)))
+      }
+    )
+
+    const comparisons: Array<{
+      profile: KnowledgeFtsConfig
+      metrics: Metrics
+      databaseBytes: number
+    }> = []
+    for (const profile of profiles) {
+      const store = new KnowledgeStore(
+        join(root, profile.profileId),
+        sourceMessages[0].accountId,
+        profile
+      )
+      await store.index({
+        conversations: fixture.conversations,
+        chunker: DEFAULT_KNOWLEDGE_CHUNKER
+      })
+      const metrics = scoreCases(
+        fixture.cases,
+        (testCase) =>
+          store.search({
+            accountId: sourceMessages[0].accountId,
+            text: testCase.question,
+            terms: testCase.oldSearchTerms,
+            limit: 10
+          }),
+        (testCase) => {
+          const evidence: KnowledgeEvidence[] = store.search({
+            accountId: sourceMessages[0].accountId,
+            text: testCase.question,
+            terms: testCase.oldSearchTerms,
+            limit: 10
+          })
+          return (
+            1_000 + Math.ceil(evidence.reduce((total, item) => total + item.text.length, 0) / 2)
+          )
+        }
+      )
+      store.checkpoint()
+      comparisons.push({ profile, metrics, databaseBytes: store.getStorageStats().databaseBytes })
+      store.close()
+    }
+    console.log(
+      `KNOWLEDGE_BENCHMARK_REPORT=${JSON.stringify(
+        {
+          fixture: 'synthetic-desensitized-v1',
+          questions: fixture.cases.length,
+          categories: ['fact', 'person', 'time', 'decision', 'semantic'],
+          oldSearch: oldMetrics,
+          fts5Comparisons: comparisons
+        },
+        null,
+        2
+      )}`
+    )
+    expect(oldMetrics.averageInputTokens).toBeGreaterThan(1_000)
+    expect(comparisons).toHaveLength(profiles.length)
+  })
+})

tests/benchmark/knowledge-capacity.benchmark.test.ts

@@ -0,0 +1,94 @@
+import { mkdtempSync } from 'fs'
+import { rename, rm, writeFile } from 'fs/promises'
+import { tmpdir } from 'os'
+import { join } from 'path'
+import { performance } from 'perf_hooks'
+import { describe, expect, it } from 'vitest'
+import { DEFAULT_KNOWLEDGE_CHUNKER, type KnowledgeFtsConfig } from '../../src/shared/knowledge'
+import { KnowledgeStore } from '../../src/main/knowledge/knowledge-store'
+import { createSyntheticConversation, FIXTURE_ACCOUNT_A } from '../fixtures/knowledge-rag'
+
+const runCapacity = process.env.KNOWLEDGE_CAPACITY === '1'
+const capacityIt = runCapacity ? it : it.skip
+const scales = [100_000, 500_000, 1_000_000] as const
+const distributions = ['short', 'mixed', 'long'] as const
+const batchSize = 10_000
+const reportPath = process.env.KNOWLEDGE_CAPACITY_REPORT_PATH || join(tmpdir(), 'wechatexplorer-knowledge-capacity-report.json')
+const profile: KnowledgeFtsConfig = {
+  profileId: 'capacity-unicode-external-full-columnsize',
+  tokenizer: 'unicode61',
+  contentMode: 'external',
+  detail: 'full',
+  columnsize: 1
+}
+
+describe('knowledge capacity benchmark', () => {
+  capacityIt(
+    'measures 100k, 500k and 1m desensitized messages across text distributions',
+    async () => {
+      const reports: Array<Record<string, number | string>> = []
+      for (const distribution of distributions) {
+        for (const messageCount of scales) {
+          const root = mkdtempSync(join(tmpdir(), `wxe-knowledge-capacity-${distribution}-${messageCount}-`))
+          const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, profile)
+          let peakWalBytes = 0
+          let peakTemporaryBytes = 0
+          let peakRssBytes = process.memoryUsage().rss
+          const started = performance.now()
+          try {
+            for (let offset = 0; offset < messageCount; offset += batchSize) {
+              const count = Math.min(batchSize, messageCount - offset)
+              await store.index({
+                conversations: [
+                  createSyntheticConversation(
+                    FIXTURE_ACCOUNT_A,
+                    `capacity-${distribution}-${offset / batchSize}`,
+                    offset,
+                    count,
+                    distribution
+                  )
+                ],
+                chunker: DEFAULT_KNOWLEDGE_CHUNKER
+              })
+              const stats = store.getStorageStats()
+              peakWalBytes = Math.max(peakWalBytes, stats.walBytes)
+              peakTemporaryBytes = Math.max(peakTemporaryBytes, stats.walBytes + stats.shmBytes)
+              peakRssBytes = Math.max(peakRssBytes, process.memoryUsage().rss)
+            }
+            store.checkpoint()
+            const stats = store.getStorageStats()
+            reports.push({
+              profile: profile.profileId,
+              distribution,
+              messageCount,
+              finalDatabaseBytes: stats.databaseBytes,
+              perTenThousandMessagesBytes: Math.round(stats.databaseBytes / (messageCount / 10_000)),
+              peakWalBytes,
+              peakTemporaryBytes,
+              elapsedMs: Math.round(performance.now() - started),
+              workerPeakRssBytes: peakRssBytes,
+              pageSize: stats.pageSize,
+              pageCount: stats.pageCount,
+              freelistCount: stats.freelistCount
+            })
+          } finally {
+            store.close()
+            await rm(root, { recursive: true, force: true })
+          }
+        }
+      }
+      const report = {
+        generatedAt: new Date().toISOString(),
+        fixture: 'synthetic-desensitized-v1',
+        profile,
+        scenarios: reports
+      }
+      const temporaryReportPath = `${reportPath}.partial`
+      await writeFile(temporaryReportPath, JSON.stringify(report, null, 2), 'utf8')
+      await rename(temporaryReportPath, reportPath)
+      console.log(`KNOWLEDGE_CAPACITY_REPORT_PATH=${reportPath}`)
+      expect(reports).toHaveLength(scales.length * distributions.length)
+    },
+    20 * 60 * 1000
+  )
+})

tests/benchmark/knowledge-realistic-fts.benchmark.test.ts

@@ -0,0 +1,209 @@
+import { mkdtempSync } from 'fs'
+import { rm } from 'fs/promises'
+import { tmpdir } from 'os'
+import { join } from 'path'
+import { performance } from 'perf_hooks'
+import { afterAll, describe, expect, it } from 'vitest'
+import {
+  DEFAULT_KNOWLEDGE_CHUNKER,
+  type KnowledgeEvidence,
+  type KnowledgeFtsConfig,
+  type KnowledgeSourceMessage
+} from '../../src/shared/knowledge'
+import { KnowledgeStore } from '../../src/main/knowledge/knowledge-store'
+import {
+  createRealisticKnowledgeFixture,
+  type RealisticBenchmarkCase,
+  type RealisticBenchmarkCategory
+} from '../fixtures/knowledge-realistic'
+
+const root = mkdtempSync(join(tmpdir(), 'wxe-realistic-fts-'))
+
+afterAll(async () => {
+  await rm(root, { recursive: true, force: true })
+})
+
+type Retrieval = { messageIds: string[] }
+type Metrics = {
+  recallAt5: number
+  recallAt10: number
+  mrr: number
+  evidenceAccuracy: number
+  findSuccessAt10: number
+  p50LatencyMs: number
+  p95LatencyMs: number
+}
+
+function percentile(values: number[], ratio: number): number {
+  if (!values.length) return 0
+  const sorted = values.slice().sort((left, right) => left - right)
+  return sorted[Math.min(sorted.length - 1, Math.ceil(sorted.length * ratio) - 1)]
+}
+
+function oldSearch(
+  messages: KnowledgeSourceMessage[],
+  testCase: RealisticBenchmarkCase
+): Retrieval[] {
+  return messages
+    .map((message) => {
+      const text =
+        `${message.text || ''}\n${message.voiceTranscript || ''}\n${message.attachment?.name || ''}`.toLowerCase()
+      const score = testCase.searchTerms.reduce(
+        (total, term) => total + (text.includes(term.toLowerCase()) ? 1 : 0),
+        0
+      )
+      return { messageIds: [message.messageId], score }
+    })
+    .filter((item) => item.score > 0)
+    .sort((left, right) => right.score - left.score)
+}
+
+function score(
+  cases: RealisticBenchmarkCase[],
+  search: (testCase: RealisticBenchmarkCase) => Retrieval[]
+): Metrics {
+  const latency: number[] = []
+  let recallAt5 = 0
+  let recallAt10 = 0
+  let reciprocalRank = 0
+  let evidenceAccuracy = 0
+  let findSuccessAt10 = 0
+  for (const testCase of cases) {
+    const started = performance.now()
+    const results = search(testCase)
+    latency.push(performance.now() - started)
+    const expected = new Set(testCase.expectedMessageIds)
+    const firstTen = results.slice(0, 10)
+    const hitIndex = firstTen.findIndex((item) => item.messageIds.some((id) => expected.has(id)))
+    if (results.slice(0, 5).some((item) => item.messageIds.some((id) => expected.has(id)))) {
+      recallAt5 += 1
+    }
+    if (hitIndex >= 0) {
+      recallAt10 += 1
+      findSuccessAt10 += 1
+      reciprocalRank += 1 / (hitIndex + 1)
+    }
+    const firstFive = results.slice(0, 5)
+    if (firstFive.length) {
+      evidenceAccuracy +=
+        firstFive.filter((item) => item.messageIds.some((id) => expected.has(id))).length /
+        firstFive.length
+    }
+  }
+  return {
+    recallAt5: recallAt5 / cases.length,
+    recallAt10: recallAt10 / cases.length,
+    mrr: reciprocalRank / cases.length,
+    evidenceAccuracy: evidenceAccuracy / cases.length,
+    findSuccessAt10: findSuccessAt10 / cases.length,
+    p50LatencyMs: percentile(latency, 0.5),
+    p95LatencyMs: percentile(latency, 0.95)
+  }
+}
+
+function groupedByCategory<T>(
+  cases: RealisticBenchmarkCase[],
+  evaluate: (items: RealisticBenchmarkCase[]) => T
+): Record<RealisticBenchmarkCategory, T> {
+  const groups = new Map<RealisticBenchmarkCategory, RealisticBenchmarkCase[]>()
+  for (const item of cases) groups.set(item.category, [...(groups.get(item.category) || []), item])
+  return Object.fromEntries(
+    Array.from(groups.entries()).map(([category, items]) => [category, evaluate(items)])
+  ) as Record<RealisticBenchmarkCategory, T>
+}
+
+const profiles: KnowledgeFtsConfig[] = [
+  {
+    profileId: 'unicode61-external-full-columnsize',
+    tokenizer: 'unicode61',
+    contentMode: 'external',
+    detail: 'full',
+    columnsize: 1
+  },
+  {
+    profileId: 'trigram-external-full-columnsize',
+    tokenizer: 'trigram',
+    contentMode: 'external',
+    detail: 'full',
+    columnsize: 1
+  },
+  {
+    profileId: 'trigram-external-column-no-columnsize',
+    tokenizer: 'trigram',
+    contentMode: 'external',
+    detail: 'column',
+    columnsize: 0
+  }
+]
+
+describe('realistic desensitized WeChat FTS5 benchmark', () => {
+  it('compares unicode61 and trigram by recall quality before Task 3 chooses a runtime profile', async () => {
+    const fixture = createRealisticKnowledgeFixture()
+    const messages = fixture.conversations.flatMap((conversation) => conversation.messages)
+    const oldMetrics = score(fixture.cases, (testCase) => oldSearch(messages, testCase))
+    const comparisons: Array<{
+      profile: KnowledgeFtsConfig
+      metrics: Metrics
+      categoryMetrics: Record<RealisticBenchmarkCategory, Metrics>
+      databaseBytes: number
+    }> = []
+    for (const profile of profiles) {
+      const store = new KnowledgeStore(
+        join(root, profile.profileId),
+        messages[0].accountId,
+        profile
+      )
+      await store.index({
+        conversations: fixture.conversations,
+        chunker: DEFAULT_KNOWLEDGE_CHUNKER
+      })
+      const search = (testCase: RealisticBenchmarkCase): Retrieval[] =>
+        store
+          .search({
+            accountId: messages[0].accountId,
+            text: testCase.question,
+            terms: testCase.searchTerms,
+            limit: 10
+          })
+          .map((item: KnowledgeEvidence) => ({ messageIds: item.messageIds }))
+      const metrics = score(fixture.cases, search)
+      store.checkpoint()
+      comparisons.push({
+        profile,
+        metrics,
+        categoryMetrics: groupedByCategory(fixture.cases, (items) => score(items, search)),
+        databaseBytes: store.getStorageStats().databaseBytes
+      })
+      store.close()
+    }
+    console.log(
+      `KNOWLEDGE_REALISTIC_FTS_REPORT=${JSON.stringify(
+        {
+          fixture: 'realistic-desensitized-wechat-v1',
+          questionCount: fixture.cases.length,
+          categories: Array.from(new Set(fixture.cases.map((item) => item.category))),
+          oldSearch: oldMetrics,
+          fts5Comparisons: comparisons
+        },
+        null,
+        2
+      )}`
+    )
+    expect(fixture.cases).toHaveLength(14)
+    expect(new Set(fixture.cases.map((item) => item.category))).toEqual(
+      new Set([
+        'chinese-continuous',
+        'chinese-short',
+        'person-name',
+        'mixed-language',
+        'url',
+        'file-name',
+        'technical-term',
+        'number-email-path',
+        'short-message',
+        'long-voice'
+      ])
+    )
+    expect(comparisons).toHaveLength(profiles.length)
+  })
+})

tests/fixtures/knowledge-rag.ts

@@ -0,0 +1,158 @@
+import type { KnowledgeConversationInput, KnowledgeSourceMessage } from '../../src/shared/knowledge'
+
+export type KnowledgeBenchmarkCategory = 'fact' | 'person' | 'time' | 'decision' | 'semantic'
+
+export interface KnowledgeBenchmarkCase {
+  id: string
+  category: KnowledgeBenchmarkCategory
+  question: string
+  oldSearchTerms: string[]
+  expectedMessageIds: string[]
+}
+
+export const FIXTURE_ACCOUNT_A = 'fixture-account-alpha'
+export const FIXTURE_ACCOUNT_B = 'fixture-account-beta'
+
+function message(
+  id: string,
+  conversationId: string,
+  createTime: number,
+  text: string,
+  extra: Partial<KnowledgeSourceMessage> = {}
+): KnowledgeSourceMessage {
+  return {
+    accountId: extra.accountId || FIXTURE_ACCOUNT_A,
+    conversationId,
+    messageId: id,
+    createTime,
+    senderId: extra.senderId || 'fixture-sender',
+    senderName: extra.senderName || '脱敏成员',
+    kind: extra.kind || 'text',
+    text,
+    attachment: extra.attachment,
+    voiceTranscript: extra.voiceTranscript
+  }
+}
+
+/** Deterministic and fully synthetic: no wxid, file path, secret, or real chat text. */
+export function createKnowledgeBenchmarkFixture(): {
+  conversations: KnowledgeConversationInput[]
+  cases: KnowledgeBenchmarkCase[]
+} {
+  const messages: KnowledgeSourceMessage[] = []
+  const cases: KnowledgeBenchmarkCase[] = []
+  const base = Date.UTC(2025, 0, 1)
+  const add = (
+    category: KnowledgeBenchmarkCategory,
+    index: number,
+    text: string,
+    question: string,
+    oldSearchTerms: string[],
+    extra: Partial<KnowledgeSourceMessage> = {}
+  ): void => {
+    const conversationId = `fixture-${category}-${index % 5}`
+    const id = `fixture-${category}-${index}-evidence`
+    messages.push(message(id, conversationId, base + (cases.length + 1) * 60_000, text, extra))
+    messages.push(
+      message(
+        `fixture-${category}-${index}-context`,
+        conversationId,
+        base + (cases.length + 1) * 60_000 + 20_000,
+        `脱敏上下文 ${index}：确认后续会回到原始消息核对。`,
+        { senderId: 'fixture-context', senderName: '脱敏同事' }
+      )
+    )
+    cases.push({
+      id: `question-${category}-${index}`,
+      category,
+      question,
+      oldSearchTerms,
+      expectedMessageIds: [id]
+    })
+  }
+  for (let index = 1; index <= 20; index += 1) {
+    add(
+      'fact',
+      index,
+      `资料编号 FACT-${index} 的部署地址是 https://example.invalid/fact-${index}，附件名称是 runbook-${index}.pdf。`,
+      `第 ${index} 项部署资料在哪里？`,
+      [`FACT-${index}`, `runbook-${index}.pdf`],
+      { attachment: { name: `runbook-${index}.pdf`, kind: 'file' } }
+    )
+    add(
+      'person',
+      index,
+      `成员 代号成员${index} 负责发布检查，并说明本周会完成验证清单。`,
+      `代号成员${index} 最近负责什么？`,
+      [`代号成员${index}`, '发布检查']
+    )
+    add(
+      'time',
+      index,
+      `日期标记 TIME-${index}：在第 ${index} 次周会讨论了回归安排和验收顺序。`,
+      `TIME-${index} 当天讨论了什么？`,
+      [`TIME-${index}`, '回归安排']
+    )
+    add(
+      'decision',
+      index,
+      `决策 DECISION-${index}：最终选择方案蓝图${index}，原因是可追溯、可回滚且维护成本更低。`,
+      `为什么第 ${index} 个决策选择方案蓝图${index}？`,
+      [`DECISION-${index}`, `方案蓝图${index}`]
+    )
+    add(
+      'semantic',
+      index,
+      `语义样本 ${index}：把分散的讨论归档，方便以后重新查看和核对当时的上下文。`,
+      `哪里提到把内容收起来以后查看？第 ${index} 条。`,
+      [`内容收起来${index}`],
+      { kind: 'voice', voiceTranscript: `请将分散讨论集中保存，便于之后重新查看，第 ${index} 条。` }
+    )
+  }
+  const grouped = new Map<string, KnowledgeSourceMessage[]>()
+  for (const item of messages) {
+    const current = grouped.get(item.conversationId) || []
+    current.push(item)
+    grouped.set(item.conversationId, current)
+  }
+  return {
+    conversations: Array.from(grouped.entries()).map(([conversationId, source]) => ({
+      conversationId,
+      completeSnapshot: true,
+      messages: source
+    })),
+    cases
+  }
+}
+
+export function createSyntheticConversation(
+  accountId: string,
+  conversationId: string,
+  startIndex: number,
+  count: number,
+  distribution: 'short' | 'mixed' | 'long'
+): KnowledgeConversationInput {
+  const base = Date.UTC(2025, 0, 1) + startIndex * 1000
+  const messages: KnowledgeSourceMessage[] = []
+  const shortText = '脱敏短消息：已确认。'
+  const mixedText = '脱敏普通消息：讨论本地知识库、索引状态、证据回跳和增量恢复。'
+  const longText = `脱敏长文本/语音转写：${'用于容量测试的可检索上下文。'.repeat(12)}`
+  for (let offset = 0; offset < count; offset += 1) {
+    const index = startIndex + offset
+    const text = distribution === 'short' ? shortText : distribution === 'mixed' ? mixedText : longText
+    messages.push({
+      accountId,
+      conversationId,
+      messageId: `synthetic-${distribution}-${index}`,
+      createTime: base + offset * 60_000,
+      senderId: `fixture-member-${index % 8}`,
+      senderName: `脱敏成员${index % 8}`,
+      kind: distribution === 'long' && index % 4 === 0 ? 'voice' : 'text',
+      text,
+      voiceTranscript: distribution === 'long' && index % 4 === 0 ? longText : undefined,
+      attachment:
+        index % 97 === 0 ? { name: `fixture-${index}.txt`, kind: 'file', sizeBytes: 2048 } : undefined
+    })
+  }
+  return { conversationId, completeSnapshot: true, messages }
+}

tests/fixtures/knowledge-realistic.ts

@@ -0,0 +1,304 @@
+import type { KnowledgeConversationInput, KnowledgeSourceMessage } from '../../src/shared/knowledge'
+
+export type RealisticBenchmarkCategory =
+  | 'chinese-continuous'
+  | 'chinese-short'
+  | 'person-name'
+  | 'mixed-language'
+  | 'url'
+  | 'file-name'
+  | 'technical-term'
+  | 'number-email-path'
+  | 'short-message'
+  | 'long-voice'
+
+export interface RealisticBenchmarkCase {
+  id: string
+  category: RealisticBenchmarkCategory
+  /** The question as a user would naturally phrase it. */
+  question: string
+  /** Deterministic local query-router output, not an LLM-generated answer. */
+  searchTerms: string[]
+  expectedMessageIds: string[]
+}
+
+export const REALISTIC_FIXTURE_ACCOUNT = 'fixture-realistic-account'
+
+function sourceMessage(
+  messageId: string,
+  conversationId: string,
+  createTime: number,
+  text: string,
+  extra: Partial<KnowledgeSourceMessage> = {}
+): KnowledgeSourceMessage {
+  return {
+    accountId: REALISTIC_FIXTURE_ACCOUNT,
+    conversationId,
+    messageId,
+    createTime,
+    senderId: extra.senderId || 'fixture-member-a',
+    senderName: extra.senderName || '脱敏成员甲',
+    kind: extra.kind || 'text',
+    text,
+    attachment: extra.attachment,
+    voiceTranscript: extra.voiceTranscript
+  }
+}
+
+/**
+ * Fully artificial messages written in the style of real WeChat conversations.
+ * Nicknames, domains, addresses, mailboxes, file names and paths are all fixtures;
+ * no user chat record, wxid, account directory, or source-database value is included.
+ */
+export function createRealisticKnowledgeFixture(): {
+  conversations: KnowledgeConversationInput[]
+  cases: RealisticBenchmarkCase[]
+} {
+  const base = Date.UTC(2026, 6, 12, 8, 0, 0)
+  const messages: KnowledgeSourceMessage[] = [
+    sourceMessage(
+      'msg-cn-continuous-1',
+      'conv-product-group',
+      base,
+      '刚确认：聊天档案导出失败时，先保留原图链接，再回退缩略图，避免用户以为图片丢了。',
+      { senderId: 'member-lan', senderName: '蓝图同学' }
+    ),
+    sourceMessage(
+      'msg-cn-continuous-noise',
+      'conv-product-group',
+      base + 60,
+      '导出完成后可以在任务中心查看文件夹。',
+      { senderId: 'member-river', senderName: '河岸' }
+    ),
+    sourceMessage('msg-cn-short-1', 'conv-family', base + 120, '周六见，咖啡我来带。', {
+      senderId: 'member-yu',
+      senderName: '小雨'
+    }),
+    sourceMessage(
+      'msg-person-1',
+      'conv-product-group',
+      base + 180,
+      '林澈把 Windows 安装包的签名检查补好了，今晚发测试包。',
+      { senderId: 'member-lin', senderName: '林澈' }
+    ),
+    sourceMessage(
+      'msg-person-noise',
+      'conv-product-group',
+      base + 240,
+      '小林晚点把截图发到群里。',
+      { senderId: 'member-lin', senderName: '林澈' }
+    ),
+    sourceMessage(
+      'msg-mixed-1',
+      'conv-engineering',
+      base + 300,
+      'Web 端的 dark mode 先跟随系统，Desktop 端继续保留手动切换。',
+      { senderId: 'member-echo', senderName: 'Echo' }
+    ),
+    sourceMessage(
+      'msg-url-1',
+      'conv-engineering',
+      base + 360,
+      '排障说明在 https://docs.example.invalid/guide/image-export?from=wechat ，不要把真实日志贴到公开 issue。',
+      { senderId: 'member-echo', senderName: 'Echo' }
+    ),
+    sourceMessage(
+      'msg-file-1',
+      'conv-filehelper',
+      base + 420,
+      '已上传 release-checklist-v2.1.9.xlsx，发布前把 macOS 和 Windows 两栏都勾完。',
+      {
+        senderId: 'self-fixture',
+        senderName: '我',
+        attachment: { name: 'release-checklist-v2.1.9.xlsx', kind: 'file', sizeBytes: 20480 }
+      }
+    ),
+    sourceMessage(
+      'msg-tech-mcp',
+      'conv-engineering',
+      base + 480,
+      'MCP Reader 只暴露只读查询；写入操作必须经过本地确认，不能让 Agent 直接改微信数据。',
+      { senderId: 'member-q', senderName: 'Q' }
+    ),
+    sourceMessage(
+      'msg-tech-react',
+      'conv-engineering',
+      base + 540,
+      'React 列表先做虚拟滚动，Electron 主进程不要把十万条消息一次性发给 renderer。',
+      { senderId: 'member-q', senderName: 'Q' }
+    ),
+    sourceMessage(
+      'msg-tech-sqlite',
+      'conv-engineering',
+      base + 600,
+      'SQLite FTS5 的 trigram 对中文子串更友好，但短词仍要有精确匹配补偿。',
+      { senderId: 'member-lan', senderName: '蓝图同学' }
+    ),
+    sourceMessage(
+      'msg-number-email-path',
+      'conv-operations',
+      base + 660,
+      '工单 48291 请发给 fixture@example.invalid；测试附件放到 /tmp/wechat-fixture/export-preview/，不要使用个人目录。',
+      { senderId: 'member-ops', senderName: '运营小组' }
+    ),
+    sourceMessage('msg-short-1', 'conv-family', base + 720, '收到，明早十点。', {
+      senderId: 'member-yu',
+      senderName: '小雨'
+    }),
+    sourceMessage('msg-short-noise', 'conv-family', base + 780, '好的，晚安。', {
+      senderId: 'member-yu',
+      senderName: '小雨'
+    }),
+    sourceMessage('msg-voice-long-1', 'conv-project-sync', base + 840, '[语音消息]', {
+      senderId: 'member-voice',
+      senderName: '语音同学',
+      kind: 'voice',
+      voiceTranscript:
+        '刚才同步一下长语音结论：本周不做向量检索，也不新增记忆页面。先把现有问问微信的关键词检索放进独立 Knowledge Worker，索引只读取原始数据库，结果必须保留 messageId、会话、发送人和时间，异常时继续使用旧搜索。'
+    }),
+    sourceMessage(
+      'msg-decision-1',
+      'conv-project-sync',
+      base + 900,
+      '决定先上 FTS，不接 Embedding：先验证中文、文件名和技术词的召回，再考虑下一阶段。',
+      { senderId: 'member-voice', senderName: '语音同学' }
+    ),
+    sourceMessage(
+      'msg-url-noise',
+      'conv-engineering',
+      base + 960,
+      '本周会议链接仍然走内部日历，不要混在发布文档里。',
+      { senderId: 'member-echo', senderName: 'Echo' }
+    ),
+    sourceMessage(
+      'msg-file-noise',
+      'conv-filehelper',
+      base + 1020,
+      '旧版 release-note.txt 仅供历史核对，不要再上传。',
+      { senderId: 'self-fixture', senderName: '我' }
+    ),
+    sourceMessage(
+      'msg-long-text-1',
+      'conv-project-sync',
+      base + 1080,
+      '补充记录：当索引仍在建立或 Worker 发生异常，界面行为不能中断。主进程需要保留旧关键词检索作为临时回退，但 renderer 不应重新批量加载全部会话消息。等索引完成后，Evidence 应统一由知识库返回，并能跳回原聊天。',
+      { senderId: 'member-lan', senderName: '蓝图同学' }
+    )
+  ]
+
+  const cases: RealisticBenchmarkCase[] = [
+    {
+      id: 'cn-continuous',
+      category: 'chinese-continuous',
+      question: '图片导出失败时应该怎样避免用户误以为图片丢失？',
+      searchTerms: ['原图链接', '缩略图'],
+      expectedMessageIds: ['msg-cn-continuous-1']
+    },
+    {
+      id: 'cn-short',
+      category: 'chinese-short',
+      question: '周六谁带咖啡？',
+      searchTerms: ['周六见', '咖啡'],
+      expectedMessageIds: ['msg-cn-short-1']
+    },
+    {
+      id: 'person-name',
+      category: 'person-name',
+      question: '林澈最近补了什么？',
+      searchTerms: ['林澈', '签名检查'],
+      expectedMessageIds: ['msg-person-1']
+    },
+    {
+      id: 'mixed-language',
+      category: 'mixed-language',
+      question: 'dark mode 在 Web 和 Desktop 分别怎么处理？',
+      searchTerms: ['dark mode', 'Desktop'],
+      expectedMessageIds: ['msg-mixed-1']
+    },
+    {
+      id: 'url',
+      category: 'url',
+      question: '图片导出排障文档的网址是什么？',
+      searchTerms: ['docs.example.invalid/guide/image-export'],
+      expectedMessageIds: ['msg-url-1']
+    },
+    {
+      id: 'file-name',
+      category: 'file-name',
+      question: '发布检查表文件叫什么？',
+      searchTerms: ['release-checklist-v2.1.9.xlsx'],
+      expectedMessageIds: ['msg-file-1']
+    },
+    {
+      id: 'technical-mcp',
+      category: 'technical-term',
+      question: 'MCP Reader 的写入限制是什么？',
+      searchTerms: ['MCP Reader', '只读查询'],
+      expectedMessageIds: ['msg-tech-mcp']
+    },
+    {
+      id: 'technical-react-electron',
+      category: 'technical-term',
+      question: 'React 和 Electron 的大量消息处理原则是什么？',
+      searchTerms: ['React', 'Electron'],
+      expectedMessageIds: ['msg-tech-react']
+    },
+    {
+      id: 'technical-sqlite',
+      category: 'technical-term',
+      question: 'SQLite 的中文全文检索要选什么？',
+      searchTerms: ['SQLite FTS5', 'trigram'],
+      expectedMessageIds: ['msg-tech-sqlite']
+    },
+    {
+      id: 'number-email-path',
+      category: 'number-email-path',
+      question: '工单 48291 的邮箱和测试附件目录在哪？',
+      searchTerms: ['48291', 'fixture@example.invalid', '/tmp/wechat-fixture/export-preview'],
+      expectedMessageIds: ['msg-number-email-path']
+    },
+    {
+      id: 'short-message',
+      category: 'short-message',
+      question: '约的是几点？',
+      searchTerms: ['十点'],
+      expectedMessageIds: ['msg-short-1']
+    },
+    {
+      id: 'long-voice',
+      category: 'long-voice',
+      question: '长语音里对 Knowledge Worker 和 fallback 的要求是什么？',
+      searchTerms: ['Knowledge Worker', '旧搜索'],
+      expectedMessageIds: ['msg-voice-long-1']
+    },
+    {
+      id: 'decision',
+      category: 'long-voice',
+      question: '为什么暂时不接 Embedding？',
+      searchTerms: ['不接 Embedding', 'FTS'],
+      expectedMessageIds: ['msg-decision-1']
+    },
+    {
+      id: 'fallback',
+      category: 'long-voice',
+      question: '索引未完成时搜索如何处理？',
+      searchTerms: ['Worker 发生异常', '旧关键词检索'],
+      expectedMessageIds: ['msg-long-text-1']
+    }
+  ]
+
+  const grouped = new Map<string, KnowledgeSourceMessage[]>()
+  for (const item of messages) {
+    const current = grouped.get(item.conversationId) || []
+    current.push(item)
+    grouped.set(item.conversationId, current)
+  }
+  return {
+    conversations: Array.from(grouped.entries()).map(([conversationId, source]) => ({
+      conversationId,
+      completeSnapshot: true,
+      messages: source
+    })),
+    cases
+  }
+}

tests/integration/preload-contract.test.ts

@@ -1,4 +1,6 @@
 import { beforeEach, describe, expect, it, vi } from 'vitest'
+import type { AiSearchPipelineRequest } from '../../src/shared/ai-search'
+import type { KnowledgeSearchIpcRequest } from '../../src/shared/knowledge'
 
 const invoke = vi.fn()
 const on = vi.fn()
@@ -37,6 +39,28 @@ describe('preload IPC contract', () => {
       limit: 50
     })
 
+    const knowledgeSearch: KnowledgeSearchIpcRequest = {
+      text: '测试 Knowledge Worker 检索',
+      terms: ['Knowledge Worker'],
+      conversationIds: ['fixture-user'],
+      startTime: 10,
+      limit: 20
+    }
+    await api.searchKnowledge(knowledgeSearch)
+    expect(invoke).toHaveBeenLastCalledWith('knowledge:search', knowledgeSearch)
+    const aiSearch: AiSearchPipelineRequest = {
+      requestId: 'fixture-search',
+      text: '最近谁聊过健身',
+      scope: 'global',
+      range: '7d'
+    }
+    await api.runAiSearch(aiSearch)
+    expect(invoke).toHaveBeenLastCalledWith('ai-search:run', aiSearch)
+    await api.startKnowledgeIndex()
+    expect(invoke).toHaveBeenLastCalledWith('knowledge:startIndex')
+    await api.clearCache('knowledge')
+    expect(invoke).toHaveBeenLastCalledWith('cache:clear', 'knowledge')
+
     await api.getImage('fixture-md5', 'fixture.dat', 'fixture-session', {
       force: true,
       priority: 0

tests/unit/ai-search-evidence.test.ts

@@ -0,0 +1,76 @@
+import { describe, expect, it } from 'vitest'
+import {
+  buildFinalEvidence,
+  evidenceIdentity,
+  sanitizeAnswerCitations
+} from '../../src/main/services/ai-search-evidence'
+import type { AiSearchPipelineEvidence } from '../../src/shared/ai-search'
+
+const candidate = (
+  index: number,
+  options: Partial<AiSearchPipelineEvidence> = {}
+): AiSearchPipelineEvidence => ({
+  chunkId: `chunk-${index}`,
+  conversationId: index % 2 ? 'fitness-group-a' : 'fitness-group-b',
+  conversationName: index % 2 ? '健身群 A' : '健身群 B',
+  conversationType: 'group',
+  messageId: `message-${index}`,
+  senderId: index % 3 ? 'member-yang' : 'member-dongfang',
+  sender: index % 3 ? '杨伟' : '东方小唠',
+  startTime: 1_785_895_200_000 + index,
+  endTime: 1_785_895_200_000 + index,
+  timestamp: 1_785_895_200_000 + index,
+  messageIds: [`message-${index}`],
+  text: `第 ${index} 条去健身相关消息`,
+  score: -index,
+  ...options
+})
+
+describe('Final Evidence builder', () => {
+  it('uses exactly the same program-owned E1-E8 collection for final context', () => {
+    const candidates = Array.from({ length: 16 }, (_, index) => candidate(index + 1))
+    const result = buildFinalEvidence(candidates, 8)
+
+    expect(result.candidateCount).toBe(16)
+    expect(result.evidence).toHaveLength(8)
+    expect(result.evidence.map((item) => item.id)).toEqual([
+      'E1',
+      'E2',
+      'E3',
+      'E4',
+      'E5',
+      'E6',
+      'E7',
+      'E8'
+    ])
+    expect(result.evidence.map(evidenceIdentity)).toEqual(
+      Array.from({ length: 8 }, (_, index) => evidenceIdentity(candidate(16 - index)))
+    )
+    expect(result.aggregation.messageCount).toBe(8)
+    expect(result.aggregation.peopleCount).toBe(2)
+    expect(result.aggregation.conversationCount).toBe(2)
+  })
+
+  it('does not merge same message ids from different conversations', () => {
+    const first = candidate(1, { conversationId: 'conversation-a', messageId: 'same-message-id' })
+    const second = candidate(2, { conversationId: 'conversation-b', messageId: 'same-message-id' })
+
+    const result = buildFinalEvidence([first, second], 8)
+
+    expect(result.evidence).toHaveLength(2)
+    expect(result.evidence.map(evidenceIdentity)).toEqual([
+      'conversation-b\u0000same-message-id',
+      'conversation-a\u0000same-message-id'
+    ])
+  })
+
+  it('removes citations which do not resolve to Final Evidence', () => {
+    const evidence = buildFinalEvidence([candidate(1), candidate(2)], 8).evidence
+    const result = sanitizeAnswerCitations('杨伟提到健身。[E1] 另有无效来源。[E10][E23]', evidence)
+
+    expect(result.status).toBe('sanitized')
+    expect(result.invalidCitationIds).toEqual(['E10', 'E23'])
+    expect(result.answer).toContain('[E1]')
+    expect(result.answer).not.toMatch(/\[E(?:10|23)\]/)
+  })
+})

tests/unit/ai-search-pipeline-service.test.ts

@@ -0,0 +1,732 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+const { chatState, listContactsAsync } = vi.hoisted(() => ({
+  chatState: { ready: true },
+  listContactsAsync: vi.fn()
+}))
+
+vi.mock('../../src/main/services/chat-service', () => ({
+  isReady: () => chatState.ready,
+  listContactsAsync
+}))
+
+import { AiSearchPipelineService } from '../../src/main/services/ai-search-pipeline-service'
+import type { KnowledgeEvidence } from '../../src/shared/knowledge'
+
+const makeCandidate = (index: number): KnowledgeEvidence => ({
+  chunkId: `chunk-${index}`,
+  conversationId: index % 2 ? 'fitness-group-a' : 'fitness-group-b',
+  startTime: 1785900000000 + index,
+  endTime: 1785900000000 + index,
+  messageId: `message-${index}`,
+  sender: index % 2 ? '杨伟' : '东方小唠',
+  senderId: index % 2 ? 'member-yang' : 'member-dongfang',
+  timestamp: 1785900000000 + index,
+  messageIds: [`message-${index}`],
+  text: `candidate-${index} 去健身`,
+  score: -index
+})
+
+describe('AiSearchPipelineService', () => {
+  const knowledge = { search: vi.fn() }
+  const aiProvider = { getRuntimeConfig: vi.fn(), chat: vi.fn() }
+
+  beforeEach(() => {
+    chatState.ready = true
+    listContactsAsync.mockReset()
+    knowledge.search.mockReset()
+    aiProvider.getRuntimeConfig.mockReset()
+    aiProvider.chat.mockReset()
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'fitness-group',
+        m_nsUsrName: 'fitness-group@chatroom',
+        m_nsNickName: '健身交流组',
+        type: 'group'
+      }
+    ])
+    knowledge.search.mockResolvedValue({
+      source: 'knowledge',
+      state: 'ready',
+      indexedMessageCount: 2_000,
+      indexedChunkCount: 300,
+      totalMessages: 2_000,
+      evidence: [
+        {
+          chunkId: 'chunk-1',
+          conversationId: 'fitness-group',
+          startTime: 1785900000000,
+          endTime: 1785900000000,
+          messageId: 'message-1',
+          sender: '小明',
+          senderId: 'wxid_fixture',
+          timestamp: 1785900000000,
+          messageIds: ['message-1'],
+          text: '今天下班去健身。'
+        }
+      ]
+    })
+    aiProvider.getRuntimeConfig.mockReturnValue({
+      configured: true,
+      providerName: 'DeepSeek',
+      modelName: 'DeepSeek Chat'
+    })
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_messages","arguments":{"query":"健身"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"finalize","reason":"已找到足够的相关消息"}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '小明提到今天下班去健身。[E1]',
+        usage: { input: 120 }
+      })
+  })
+
+  it('emits actual planning, knowledge, evidence and AI completion states', async () => {
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+    const events: Array<{ stage: string; status: string; message: string }> = []
+    const result = await service.run(
+      {
+        requestId: 'fixture-request',
+        text: '最近谁聊过健身',
+        scope: 'global',
+        range: '7d'
+      },
+      (event) => events.push(event)
+    )
+
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({ text: '最近谁聊过健身', terms: ['健身'] })
+    )
+    expect(events).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({ stage: 'query_understanding', status: 'running' }),
+        expect.objectContaining({ stage: 'agent_start', status: 'completed' }),
+        expect.objectContaining({ stage: 'agent_tool', status: 'completed' }),
+        expect.objectContaining({ stage: 'search_plan_ready', status: 'completed' }),
+        expect.objectContaining({ stage: 'knowledge_searching', status: 'completed' }),
+        expect.objectContaining({ stage: 'evidence_ready', status: 'completed' }),
+        expect.objectContaining({ stage: 'aggregation', status: 'completed' }),
+        expect.objectContaining({
+          stage: 'ai_generating',
+          status: 'running',
+          modelName: 'DeepSeek Chat'
+        }),
+        expect.objectContaining({ stage: 'completed', status: 'completed' })
+      ])
+    )
+    expect(result).toMatchObject({
+      status: 'completed',
+      candidateEvidenceCount: 1,
+      contextEvidenceCount: 1,
+      answer: '小明提到今天下班去健身。[E1]',
+      ai: { inputTokens: 120, inputTokensEstimated: false }
+    })
+    expect(result.agent).toMatchObject({ mode: 'agent', toolCalls: 1 })
+  })
+
+  it('keeps real evidence when the answer model fails', async () => {
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_messages","arguments":{"query":"健身"}}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '{"action":"finalize","reason":"证据足够"}' })
+      .mockResolvedValueOnce({ success: false, error: '模型超时' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+    const events: Array<{ stage: string; status: string; message: string }> = []
+    const result = await service.run(
+      {
+        requestId: 'fixture-ai-error',
+        text: '最近聊过健身吗',
+        scope: 'global',
+        range: '7d'
+      },
+      (event) => events.push(event)
+    )
+
+    expect(result).toMatchObject({ status: 'ai_failed', evidence: [expect.any(Object)] })
+    expect(events).toContainEqual(
+      expect.objectContaining({ stage: 'ai_generating', status: 'error', error: '模型超时' })
+    )
+  })
+
+  it('uses Final Evidence only for AI context and strips invalid citations', async () => {
+    knowledge.search.mockResolvedValue({
+      source: 'knowledge',
+      state: 'ready',
+      indexedMessageCount: 2_000,
+      indexedChunkCount: 300,
+      totalMessages: 2_000,
+      evidence: Array.from({ length: 16 }, (_, index) => makeCandidate(index + 1)),
+      timings: {
+        workerIpcMs: 4,
+        ftsMs: 8,
+        messageLoadMs: 5,
+        chunkExpandMs: 6,
+        rankingMs: 2,
+        totalMs: 25
+      }
+    })
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_messages","arguments":{"query":"健身"}}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '{"action":"finalize","reason":"证据足够"}' })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '杨伟聊过去健身。[E1] 错误引用。[E10][E23]',
+        usage: { input: 160 }
+      })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'final-evidence-only',
+        text: '全局搜一下 谁聊过 去健身',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    const answerPrompt = aiProvider.chat.mock.calls[2][0][1].content as string
+    const contextIds = Array.from(answerPrompt.matchAll(/\[E(\d+)\]\nconversationId:/g)).map(
+      (match) => Number(match[1])
+    )
+    expect(contextIds).toEqual([1, 2, 3, 4, 5, 6, 7, 8])
+    expect(answerPrompt).not.toContain('candidate-1 去健身')
+    expect(result).toMatchObject({
+      status: 'completed',
+      candidateEvidenceCount: 16,
+      contextEvidenceCount: 8,
+      citationValidation: { status: 'sanitized', invalidCitationIds: ['E10', 'E23'] }
+    })
+    expect(result.evidence.map((item) => item.id)).toEqual([
+      'E1',
+      'E2',
+      'E3',
+      'E4',
+      'E5',
+      'E6',
+      'E7',
+      'E8'
+    ])
+    expect(result.answer).toContain('[E1]')
+    expect(result.answer).not.toMatch(/\[E(?:10|23)\]/)
+    expect(result.aggregation).toMatchObject({
+      messageCount: 8,
+      peopleCount: 2,
+      conversationCount: 2
+    })
+    expect(result.timings).toMatchObject({
+      queryUnderstandingMs: expect.any(Number),
+      contactResolutionMs: expect.any(Number),
+      knowledgeSearchMs: expect.any(Number),
+      ftsMs: 8,
+      totalMs: expect.any(Number)
+    })
+  })
+
+  it('retries a different conversation query after the first search returns zero results', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'technology-group',
+        m_nsUsrName: 'technology-group@chatroom',
+        m_nsNickName: '技术交流',
+        type: 'group'
+      }
+    ])
+    knowledge.search.mockResolvedValue({
+      source: 'knowledge',
+      state: 'ready',
+      indexedMessageCount: 2_000,
+      indexedChunkCount: 300,
+      totalMessages: 2_000,
+      evidence: [
+        {
+          chunkId: 'technology-chunk',
+          conversationId: 'technology-group',
+          startTime: 1785900000000,
+          endTime: 1785900000000,
+          messageId: 'technology-message',
+          sender: '小周',
+          timestamp: 1785900000000,
+          messageIds: ['technology-message'],
+          text: '今天讨论了 Electron 的打包问题。'
+        }
+      ],
+      timings: {
+        workerIpcMs: 1,
+        ftsMs: 2,
+        messageLoadMs: 1,
+        chunkExpandMs: 1,
+        rankingMs: 1,
+        totalMs: 6
+      }
+    })
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_conversations","arguments":{"query":"技术交流群"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_conversations","arguments":{"query":"技术交流"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"get_conversation_messages","arguments":{"conversationRef":"conversation-1","limit":50}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"finalize","reason":"已获得会话近期消息"}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '技术交流讨论了 Electron 打包问题。[E1]' })
+    const events: Array<Record<string, unknown>> = []
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      { requestId: 'retry-query', text: '我在技术交流群聊了什么？', scope: 'global', range: '30d' },
+      (event) => events.push(event as unknown as Record<string, unknown>)
+    )
+
+    expect(result).toMatchObject({ status: 'completed', agent: { mode: 'agent', toolCalls: 3 } })
+    expect(result.agent.trace).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({ toolName: 'search_conversations', resultCount: 0 }),
+        expect.objectContaining({ toolName: 'search_conversations', resultCount: 1 }),
+        expect.objectContaining({ toolName: 'get_conversation_messages', resultCount: 1 })
+      ])
+    )
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({ terms: [], conversationIds: ['technology-group'], limit: 50 })
+    )
+    expect(events).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({
+          stage: 'agent_tool',
+          agentTrace: expect.objectContaining({ resultCount: 0 })
+        })
+      ])
+    )
+  })
+
+  it('uses person lookup then metadata conversation retrieval for a contact summary', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    knowledge.search.mockResolvedValue({
+      source: 'knowledge',
+      state: 'ready',
+      indexedMessageCount: 2_000,
+      indexedChunkCount: 300,
+      totalMessages: 2_000,
+      evidence: Array.from({ length: 8 }, (_, index) => ({
+        ...makeCandidate(index + 1),
+        conversationId: 'zhongtian-contact'
+      })),
+      timings: {
+        workerIpcMs: 1,
+        ftsMs: 0,
+        messageLoadMs: 2,
+        chunkExpandMs: 0,
+        rankingMs: 1,
+        totalMs: 4
+      },
+      conversationRetrieval: {
+        conversationId: 'zhongtian-contact',
+        totalMessages: 327,
+        chunkCount: 10,
+        candidateMessages: 30,
+        systemMessagesDeprioritized: 2,
+        complete: true
+      }
+    })
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_people","arguments":{"query":"中田健身-弘毅"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"get_conversation_messages","arguments":{"conversationRef":"conversation-1"}}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '你们最近聊过健身安排。[E1]' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'contact-summary',
+        text: '我和中田健身-弘毅最近聊了什么？',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({ status: 'completed', agent: { mode: 'agent', toolCalls: 2 } })
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({
+        terms: [],
+        conversationIds: ['zhongtian-contact'],
+        startTime: expect.any(Number)
+      })
+    )
+    expect(knowledge.search).not.toHaveBeenCalledWith(
+      expect.objectContaining({ terms: expect.arrayContaining(['中田健身-弘毅']) })
+    )
+    expect(aiProvider.chat).toHaveBeenCalledTimes(3)
+    expect(result.agent.trace).toContainEqual(
+      expect.objectContaining({ label: '本地资料已覆盖所选时间范围，可直接整理回答' })
+    )
+    const decisions = result.agent.trace.filter((item) => item.event === 'agentDecision')
+    expect(decisions[0]?.decisionInput).toContain('上一次 Tool 结果：尚未执行 Tool。')
+    expect(decisions[1]?.decisionInput).toContain('中田健身-弘毅')
+  })
+
+  it('keeps a direct contact recap on metadata retrieval when the Agent JSON response is invalid', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    knowledge.search.mockResolvedValue({
+      source: 'knowledge',
+      state: 'ready',
+      indexedMessageCount: 2_000,
+      indexedChunkCount: 300,
+      totalMessages: 2_000,
+      evidence: Array.from({ length: 8 }, (_, index) => ({
+        ...makeCandidate(index + 1),
+        conversationId: 'zhongtian-contact',
+        text: `我肚子前面放盒肌酸，才是 ${118 + index}。`
+      }))
+    })
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({ success: true, data: '我建议先找到这位联系人。' })
+      .mockResolvedValueOnce({ success: true, data: '你们最近聊到了腰围和肌酸。[E1]' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'contact-summary-agent-recovery',
+        text: '我和中田健身弘毅最近聊了什么？',
+        scope: 'global',
+        range: 'all'
+      },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({
+      status: 'completed',
+      agent: {
+        mode: 'fallback',
+        fallbackReason: expect.stringContaining('相同检索意图的本地确定性策略')
+      }
+    })
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({
+        conversationIds: ['zhongtian-contact'],
+        terms: [],
+        startTime: expect.any(Number)
+      })
+    )
+    expect(knowledge.search).not.toHaveBeenCalledWith(
+      expect.objectContaining({ terms: expect.arrayContaining(['中田健身弘毅']) })
+    )
+  })
+
+  it('uses person lookup plus conversation-scoped topic search for a contact question', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_people","arguments":{"query":"中田健身-弘毅"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_messages","arguments":{"conversationRef":"conversation-1","query":"健身"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"finalize","reason":"已找到话题证据"}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '你们最近聊过健身。[E1]' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'contact-topic',
+        text: '我和中田健身-弘毅最近聊过健身吗？',
+        scope: 'global',
+        range: 'all'
+      },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({ status: 'completed', agent: { mode: 'agent', toolCalls: 2 } })
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({
+        terms: ['健身'],
+        conversationIds: ['zhongtian-contact'],
+        startTime: expect.any(Number)
+      })
+    )
+    expect(knowledge.search).not.toHaveBeenCalledWith(
+      expect.objectContaining({ terms: expect.arrayContaining(['中田健身-弘毅']) })
+    )
+  })
+
+  it('rejects a forbidden contact-recall FTS action and keeps the deterministic fallback semantic', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_messages","arguments":{"query":"中田健身弘毅"}}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '这不是有效 Agent JSON' })
+      .mockResolvedValueOnce({ success: true, data: '已从会话中整理出最近内容。[E1]' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'forbidden-contact-recall-fts',
+        text: '我和中田健身弘毅最近聊了什么？',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    expect(result.agent).toMatchObject({ mode: 'fallback' })
+    expect(result.agent.trace).toContainEqual(
+      expect.objectContaining({
+        toolName: 'search_messages',
+        decision: expect.stringContaining('联系人回顾只允许')
+      })
+    )
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({ conversationIds: ['zhongtian-contact'], terms: [] })
+    )
+    expect(knowledge.search).not.toHaveBeenCalledWith(
+      expect.objectContaining({ terms: expect.arrayContaining(['中田健身弘毅']) })
+    )
+  })
+
+  it('rejects an unscoped FTS action for a contact topic question', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_messages","arguments":{"query":"健身"}}'
+      })
+      .mockResolvedValueOnce({ success: true, data: '无效控制输出' })
+      .mockResolvedValueOnce({ success: true, data: '你们聊过健身。[E1]' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    await service.run(
+      {
+        requestId: 'forbidden-unscoped-contact-topic',
+        text: '我和中田健身弘毅最近聊过健身吗？',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    expect(knowledge.search).toHaveBeenCalledWith(
+      expect.objectContaining({
+        terms: ['健身'],
+        conversationIds: ['zhongtian-contact']
+      })
+    )
+    expect(knowledge.search).not.toHaveBeenCalledWith(
+      expect.objectContaining({ terms: ['健身'], conversationIds: undefined })
+    )
+  })
+
+  it('flags suspicious contact retrieval and refuses to summarize one message as a full conversation', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    knowledge.search.mockResolvedValue({
+      source: 'knowledge',
+      state: 'ready',
+      indexedMessageCount: 2_000,
+      indexedChunkCount: 300,
+      totalMessages: 2_000,
+      evidence: [{ ...makeCandidate(1), conversationId: 'zhongtian-contact' }],
+      conversationRetrieval: {
+        conversationId: 'zhongtian-contact',
+        totalMessages: 134,
+        chunkCount: 8,
+        candidateMessages: 1,
+        systemMessagesDeprioritized: 1,
+        complete: true
+      }
+    })
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_people","arguments":{"query":"中田健身弘毅"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"get_conversation_messages","arguments":{"conversationRef":"conversation-1"}}'
+      })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'suspicious-contact-retrieval',
+        text: '我和中田健身弘毅最近聊了什么？',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({
+      status: 'retrieval_incomplete',
+      retrieval: {
+        conversationId: 'zhongtian-contact',
+        sourceMessageCount: 134,
+        candidateCount: 1,
+        suspicious: true
+      }
+    })
+    expect(knowledge.search).toHaveBeenCalledTimes(2)
+    expect(aiProvider.chat).toHaveBeenCalledTimes(2)
+  })
+
+  it('does not turn a zero-result person lookup or early Agent finalize into contact-name FTS', async () => {
+    listContactsAsync.mockResolvedValue([
+      {
+        md5: 'zhongtian-contact',
+        m_nsUsrName: 'wxid_zhongtian',
+        m_nsNickName: '中田健身-弘毅',
+        type: 'user'
+      }
+    ])
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"tool","tool":"search_people","arguments":{"query":"不存在的人"}}'
+      })
+      .mockResolvedValueOnce({
+        success: true,
+        data: '{"action":"finalize","reason":"没有足够证据"}'
+      })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'zero-person-lookup-safe',
+        text: '我和中田健身弘毅最近聊了什么？',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({ status: 'no_evidence', agent: { mode: 'agent', toolCalls: 1 } })
+    expect(knowledge.search).not.toHaveBeenCalled()
+    expect(aiProvider.chat).toHaveBeenCalledTimes(2)
+  })
+
+  it('stops after five Tool calls instead of searching indefinitely', async () => {
+    aiProvider.chat.mockReset()
+    for (let index = 0; index < 5; index += 1) {
+      aiProvider.chat.mockResolvedValueOnce({
+        success: true,
+        data: `{"action":"tool","tool":"search_conversations","arguments":{"query":"不存在的群${index}"}}`
+      })
+    }
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      {
+        requestId: 'max-tool-calls',
+        text: '我在一个不存在的群聊了什么？',
+        scope: 'global',
+        range: '30d'
+      },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({ status: 'no_evidence', agent: { mode: 'agent', toolCalls: 5 } })
+    expect(result.agent.trace).toContainEqual(
+      expect.objectContaining({ label: '已达到本次检索上限' })
+    )
+    expect(aiProvider.chat).toHaveBeenCalledTimes(5)
+    expect(knowledge.search).not.toHaveBeenCalled()
+  })
+
+  it('falls back to the existing one-shot search when Agent output violates the control protocol', async () => {
+    aiProvider.chat.mockReset()
+    aiProvider.chat
+      .mockResolvedValueOnce({ success: true, data: '我来执行任意代码' })
+      .mockResolvedValueOnce({ success: true, data: '{"intent":"topic","keywords":["健身"]}' })
+      .mockResolvedValueOnce({ success: true, data: '小明聊到健身。[E1]' })
+    const service = new AiSearchPipelineService(knowledge as never, aiProvider as never)
+
+    const result = await service.run(
+      { requestId: 'agent-fallback', text: '最近聊过健身吗？', scope: 'global', range: '7d' },
+      () => undefined
+    )
+
+    expect(result).toMatchObject({ status: 'completed', agent: { mode: 'fallback', toolCalls: 0 } })
+    expect(result.agent.fallbackReason).toContain('受控搜索 Agent')
+  })
+})

tests/unit/ai-search-time-range.test.ts

@@ -0,0 +1,77 @@
+import { describe, expect, it } from 'vitest'
+import {
+  buildLocalAiSearchPlan,
+  includesExplicitAiSearchAlias,
+  inferAiSearchTimeRange
+} from '../../src/shared/ai-search'
+
+const NOW = new Date('2026-08-05T12:00:00+08:00')
+
+describe('AI search natural-language time ranges', () => {
+  it('tightens an all-history selection when the user says 最近', () => {
+    expect(inferAiSearchTimeRange('我和张三最近聊了什么？', 'all', NOW)).toMatchObject({
+      label: '近 30 天',
+      source: 'query',
+      startTime: Math.floor(NOW.getTime() / 1000) - 30 * 86400
+    })
+  })
+
+  it('recognizes explicit recent days and the current year', () => {
+    expect(inferAiSearchTimeRange('我和张三最近三天聊了什么？', 'all', NOW)).toMatchObject({
+      label: '近 3 天',
+      source: 'query'
+    })
+    expect(inferAiSearchTimeRange('我和张三今年聊了什么？', 'all', NOW)).toMatchObject({
+      label: '今年',
+      startTime: Math.floor(new Date(2026, 0, 1).getTime() / 1000)
+    })
+  })
+
+  it('keeps an explicit user retry override above the word 最近 in the original question', () => {
+    expect(
+      inferAiSearchTimeRange('我和张三最近聊了什么？', 'all', NOW, {
+        label: '全部历史',
+        reason: '用户主动扩大到全部历史',
+        source: 'user_retry'
+      })
+    ).toMatchObject({
+      label: '全部历史',
+      source: 'user_retry'
+    })
+  })
+
+  it('classifies a direct person recap as conversation_recall rather than a topic FTS query', () => {
+    expect(buildLocalAiSearchPlan('我和张三最近聊了什么？')).toMatchObject({
+      intent: 'conversation_recall',
+      contactQuery: '张三'
+    })
+  })
+
+  it('keeps identity and message topic separate for a contact topic search', () => {
+    expect(buildLocalAiSearchPlan('我和张三最近聊过健身吗？')).toMatchObject({
+      intent: 'conversation_topic_search',
+      contactQuery: '张三',
+      topicQuery: '健身',
+      keywords: ['健身']
+    })
+  })
+
+  it('classifies global topics and bare conversation names without turning names into FTS terms', () => {
+    expect(buildLocalAiSearchPlan('最近谁聊过 MCP？')).toMatchObject({
+      intent: 'global_topic_search',
+      topicQuery: 'MCP',
+      keywords: ['MCP']
+    })
+    expect(buildLocalAiSearchPlan('技术交流群')).toMatchObject({
+      intent: 'conversation_name_search',
+      contactQuery: '技术交流群',
+      topicQuery: undefined
+    })
+  })
+
+  it('matches an explicitly mentioned nickname when the user omits punctuation', () => {
+    expect(includesExplicitAiSearchAlias('我和中田健身弘毅最近聊了什么？', '中田健身-弘毅')).toBe(
+      true
+    )
+  })
+})

tests/unit/contact-resolution-service.test.ts

@@ -0,0 +1,69 @@
+import { describe, expect, it } from 'vitest'
+import { normalizeContactName } from '../../src/shared/contact-resolution'
+import { resolveContact } from '../../src/main/services/contact-resolution-service'
+
+const contacts = [
+  {
+    md5: 'coach',
+    m_nsUsrName: 'wxid_coach',
+    m_nsNickName: '中田健身-弘毅',
+    type: 'user' as const,
+    remark: '弘毅教练'
+  },
+  { md5: 'zhangsan', m_nsUsrName: 'wxid_zhangsan', m_nsNickName: '张三', type: 'user' as const },
+  {
+    md5: 'zhangsanfeng',
+    m_nsUsrName: 'wxid_zhangsanfeng',
+    m_nsNickName: '张三丰',
+    type: 'user' as const
+  }
+]
+
+describe('ContactResolutionService', () => {
+  it('canonicalizes whitespace, Unicode separators, punctuation and full-width variants', () => {
+    const forms = [
+      '中田健身-弘毅',
+      '中田健身弘毅',
+      '中田健身 弘毅',
+      '中田健身—弘毅',
+      '中田健身_弘毅'
+    ]
+    expect(new Set(forms.map(normalizeContactName))).toEqual(new Set(['中田健身弘毅']))
+  })
+
+  it('resolves every canonical name form to one conversation without substring guessing', () => {
+    for (const value of [
+      '中田健身-弘毅',
+      '中田健身弘毅',
+      '中田健身 弘毅',
+      '中田健身—弘毅',
+      '中田健身_弘毅'
+    ]) {
+      expect(resolveContact(value, contacts, 'person')).toMatchObject({
+        matched: true,
+        conversationId: 'coach',
+        ambiguous: false
+      })
+    }
+  })
+
+  it('does not treat a partial name as an identity match', () => {
+    expect(resolveContact('张三丰老师', contacts, 'person')).toMatchObject({
+      matched: false,
+      ambiguous: false,
+      candidates: []
+    })
+  })
+
+  it('does not auto-select duplicate canonical aliases', () => {
+    const duplicate = [
+      ...contacts,
+      { ...contacts[0], md5: 'coach-duplicate', m_nsUsrName: 'wxid_other' }
+    ]
+    expect(resolveContact('中田健身弘毅', duplicate, 'person')).toMatchObject({
+      matched: false,
+      ambiguous: true,
+      candidates: [expect.any(Object), expect.any(Object)]
+    })
+  })
+})

tests/unit/knowledge-search-service.test.ts

@@ -0,0 +1,208 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+const { chatState, getGroupSnapshotAsync, listContactsAsync, listMessagesAsync, knowledgeService } =
+  vi.hoisted(() => ({
+    chatState: {
+      ready: false,
+      accountId: ''
+    },
+    getGroupSnapshotAsync: vi.fn(),
+    listContactsAsync: vi.fn(),
+    listMessagesAsync: vi.fn(),
+    knowledgeService: {
+      dispose: vi.fn().mockResolvedValue(undefined),
+      index: vi.fn().mockResolvedValue(undefined),
+      search: vi.fn()
+    }
+  }))
+
+vi.mock('../../src/main/services/chat-service', () => ({
+  isReady: () => chatState.ready,
+  getSelfAccountInfo: () => (chatState.accountId ? { wxid: chatState.accountId } : null),
+  getCurrentAccountRoot: () => chatState.accountId,
+  getGroupSnapshotAsync,
+  listContactsAsync,
+  listMessagesAsync
+}))
+
+vi.mock('../../src/main/knowledge/knowledge-service', () => ({
+  KnowledgeService: class {
+    dispose = knowledgeService.dispose
+    index = knowledgeService.index
+    search = knowledgeService.search
+  }
+}))
+
+import { KnowledgeSearchService } from '../../src/main/knowledge/knowledge-search-service'
+
+describe('KnowledgeSearchService legacy fallback', () => {
+  beforeEach(() => {
+    chatState.ready = false
+    chatState.accountId = ''
+    getGroupSnapshotAsync.mockReset()
+    listContactsAsync.mockReset()
+    listMessagesAsync.mockReset()
+    knowledgeService.dispose.mockClear()
+    knowledgeService.index.mockClear()
+    knowledgeService.search.mockReset()
+    listContactsAsync.mockResolvedValue([
+      {
+        m_nsUsrName: 'fixture-contact',
+        m_nsNickName: '脱敏会话',
+        md5: 'fixture-conversation',
+        type: 'user'
+      }
+    ])
+    listMessagesAsync.mockResolvedValue([
+      {
+        id: 'fixture-message',
+        localId: 42,
+        from: 'user',
+        type: '普通文本',
+        datetime: '2026/8/5 10:00:00',
+        content: '请把 Knowledge Worker 的 fallback 保留下来。',
+        isSender: false,
+        senderId: 'fixture-sender',
+        name: '脱敏成员',
+        createTime: 1785895200
+      }
+    ])
+    getGroupSnapshotAsync.mockResolvedValue(null)
+  })
+
+  it('keeps the old main-process search path when Knowledge is unavailable', async () => {
+    const service = new KnowledgeSearchService('/tmp/wxe-knowledge-fallback', '/missing-worker.js')
+    const result = await service.search({
+      text: 'Knowledge Worker fallback',
+      terms: ['Knowledge Worker', 'fallback'],
+      conversationIds: ['fixture-conversation'],
+      startTime: 1785800000,
+      limit: 10
+    })
+    expect(listMessagesAsync).toHaveBeenCalledWith('fixture-conversation', 1785800000, undefined)
+    expect(result).toMatchObject({
+      source: 'fallback',
+      fallbackReason: 'unavailable',
+      state: 'unavailable',
+      totalMessages: 1
+    })
+    expect(result.evidence).toEqual([
+      expect.objectContaining({
+        messageId: 'local:42',
+        conversationId: 'fixture-conversation',
+        sender: '脱敏成员',
+        senderId: 'fixture-sender',
+        timestamp: 1785895200000
+      })
+    ])
+    await service.dispose()
+  })
+
+  it('uses existing Knowledge evidence while a new incremental pass is running', async () => {
+    chatState.ready = true
+    chatState.accountId = 'fixture-account'
+    knowledgeService.search.mockResolvedValue({
+      state: 'indexing',
+      indexedMessageCount: 300,
+      indexedChunkCount: 60,
+      evidence: [
+        {
+          chunkId: 'chunk-1',
+          conversationId: 'fixture-conversation',
+          messageId: 'fixture-message',
+          senderId: 'fixture-sender',
+          sender: '脱敏成员',
+          timestamp: 1785895200000,
+          startTime: 1785895200000,
+          endTime: 1785895200000,
+          messageIds: ['fixture-message'],
+          text: 'Knowledge 已完成的部分可以立即检索。'
+        }
+      ]
+    })
+
+    const service = new KnowledgeSearchService('/tmp/wxe-knowledge-fallback', '/missing-worker.js')
+    const result = await service.search({
+      text: 'fallback',
+      terms: ['fallback'],
+      limit: 10
+    })
+
+    expect(result).toMatchObject({
+      source: 'knowledge',
+      state: 'indexing',
+      totalMessages: 300
+    })
+    expect(result.evidence).toHaveLength(1)
+    expect(listMessagesAsync).not.toHaveBeenCalled()
+    await service.dispose()
+  })
+
+  it('splits a large scope filter before sending it to the Knowledge Worker', async () => {
+    chatState.ready = true
+    chatState.accountId = 'fixture-account'
+    knowledgeService.search.mockResolvedValue({
+      state: 'ready',
+      indexedMessageCount: 1_500,
+      indexedChunkCount: 300,
+      evidence: []
+    })
+    const conversationIds = Array.from({ length: 1_401 }, (_, index) => `conversation-${index}`)
+    const service = new KnowledgeSearchService('/tmp/wxe-knowledge-fallback', '/missing-worker.js')
+
+    const result = await service.search({
+      text: '知识库',
+      terms: ['知识库'],
+      conversationIds,
+      limit: 10
+    })
+
+    expect(result).toMatchObject({ source: 'knowledge', totalMessages: 1_500 })
+    expect(knowledgeService.search).toHaveBeenCalledTimes(3)
+    for (const [request] of knowledgeService.search.mock.calls) {
+      expect(request.conversationIds.length).toBeLessThanOrEqual(700)
+    }
+    await service.dispose()
+  })
+
+  it('resolves a group member wxid to its group nickname in fallback evidence', async () => {
+    listContactsAsync.mockResolvedValue([
+      { md5: 'fixture-group', m_nsNickName: '脱敏群聊', type: 'group' }
+    ])
+    listMessagesAsync.mockResolvedValue([
+      {
+        id: 'group-message',
+        from: 'wxid_member',
+        type: '普通文本',
+        content: '今天继续健身。',
+        isSender: false,
+        senderId: 'wxid_member',
+        name: 'wxid_member',
+        createTime: 1785895200
+      }
+    ])
+    getGroupSnapshotAsync.mockResolvedValue({
+      roomId: 'fixture-group@chatroom',
+      memberCount: 1,
+      members: [
+        {
+          wxid: 'wxid_member',
+          nickname: '微信昵称',
+          groupNickname: '健身同学',
+          wechatNickname: '微信昵称',
+          remark: '',
+          avatar: ''
+        }
+      ]
+    })
+
+    const service = new KnowledgeSearchService('/tmp/wxe-knowledge-fallback', '/missing-worker.js')
+    const result = await service.search({ text: '健身', terms: ['健身'], limit: 10 })
+
+    expect(result.evidence).toEqual([
+      expect.objectContaining({ senderId: 'wxid_member', sender: '健身同学' })
+    ])
+    expect(getGroupSnapshotAsync).toHaveBeenCalledWith('fixture-group')
+    await service.dispose()
+  })
+})

tests/unit/knowledge-store.test.ts

@@ -0,0 +1,384 @@
+import { mkdtempSync, existsSync } from 'fs'
+import { rm } from 'fs/promises'
+import { tmpdir } from 'os'
+import { join } from 'path'
+import { afterEach, describe, expect, it } from 'vitest'
+import { DEFAULT_KNOWLEDGE_CHUNKER, type KnowledgeFtsConfig } from '../../src/shared/knowledge'
+import { chunkConversation } from '../../src/main/knowledge/chunker'
+import {
+  estimateKnowledgeCapacityPreflight,
+  getKnowledgeDatabasePath,
+  KnowledgeStore,
+  removeKnowledgeDatabase
+} from '../../src/main/knowledge/knowledge-store'
+import { normalizeKnowledgeMessage } from '../../src/main/knowledge/normalizer'
+import {
+  createSyntheticConversation,
+  FIXTURE_ACCOUNT_A,
+  FIXTURE_ACCOUNT_B
+} from '../fixtures/knowledge-rag'
+
+const roots: string[] = []
+const fts: KnowledgeFtsConfig = {
+  profileId: 'test-trigram-external-full',
+  tokenizer: 'trigram',
+  contentMode: 'external',
+  detail: 'full',
+  columnsize: 1
+}
+
+function makeRoot(): string {
+  const root = mkdtempSync(join(tmpdir(), 'wxe-knowledge-'))
+  roots.push(root)
+  return root
+}
+
+afterEach(async () => {
+  await Promise.all(roots.splice(0).map((root) => rm(root, { recursive: true, force: true })))
+})
+
+describe('knowledge normalizer and chunker', () => {
+  it('indexes text, attachment metadata and existing voice transcripts without paths or binary data', () => {
+    const normalized = normalizeKnowledgeMessage({
+      accountId: FIXTURE_ACCOUNT_A,
+      conversationId: 'conversation-a',
+      messageId: 'message-a',
+      createTime: 1,
+      kind: 'voice',
+      text: ' 原始说明 ',
+      attachment: { name: 'plan.txt', kind: 'file' },
+      voiceTranscript: ' 已完成语音转写 '
+    })
+    expect(normalized.searchableText).toContain('原始说明')
+    expect(normalized.searchableText).toContain('附件：plan.txt')
+    expect(normalized.searchableText).toContain('语音转写：已完成语音转写')
+  })
+
+  it('cuts on time gaps and preserves message evidence ids', () => {
+    const source = createSyntheticConversation(
+      FIXTURE_ACCOUNT_A,
+      'conversation-a',
+      0,
+      4,
+      'short'
+    ).messages
+    source[3].createTime += 20 * 60 * 1000
+    const chunks = chunkConversation(source.map(normalizeKnowledgeMessage), {
+      ...DEFAULT_KNOWLEDGE_CHUNKER,
+      maxMessages: 12
+    })
+    expect(chunks).toHaveLength(2)
+    expect(chunks.flatMap((chunk) => chunk.messageIds)).toEqual(
+      source.map((item) => item.messageId)
+    )
+  })
+})
+
+describe('knowledge sqlite', () => {
+  it('is idempotent, supports FTS evidence lookup, and does not mix accounts', async () => {
+    const root = makeRoot()
+    const source = createSyntheticConversation(FIXTURE_ACCOUNT_A, 'conversation-a', 0, 25, 'mixed')
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    const first = await store.index({ conversations: [source], chunker: DEFAULT_KNOWLEDGE_CHUNKER })
+    const second = await store.index({
+      conversations: [source],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+    expect(first.updatedChunks).toBeGreaterThan(0)
+    expect(second.updatedChunks).toBe(0)
+    expect(second.unchangedConversations).toBe(1)
+    const evidence = store.search({ accountId: FIXTURE_ACCOUNT_A, text: '本地知识库', limit: 10 })
+    expect(evidence).not.toHaveLength(0)
+    expect(evidence[0]).toMatchObject({
+      messageId: expect.stringMatching(/^synthetic-mixed-/),
+      conversationId: 'conversation-a',
+      sender: expect.any(String),
+      timestamp: expect.any(Number)
+    })
+    expect(
+      evidence.every((item) => item.messageIds.every((id) => id.startsWith('synthetic-mixed-')))
+    ).toBe(true)
+    expect(() =>
+      store.search({ accountId: FIXTURE_ACCOUNT_B, text: '本地知识库', limit: 10 })
+    ).toThrow(/account/)
+    store.close()
+  })
+
+  it('recovers safely after cancellation and only removes the derived database', async () => {
+    const root = makeRoot()
+    const source = createSyntheticConversation(
+      FIXTURE_ACCOUNT_A,
+      'conversation-a',
+      0,
+      2_000,
+      'mixed'
+    )
+    const controller = new AbortController()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    const cancelled = await store.index(
+      { conversations: [source], chunker: DEFAULT_KNOWLEDGE_CHUNKER },
+      controller.signal,
+      (progress) => {
+        if (progress.processedMessages >= 501) controller.abort()
+      }
+    )
+    expect(cancelled.cancelled).toBe(true)
+    const resumed = await store.index({
+      conversations: [source],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+    expect(resumed.cancelled).toBe(false)
+    const databasePath = getKnowledgeDatabasePath(root, FIXTURE_ACCOUNT_A)
+    store.close()
+    expect(existsSync(databasePath)).toBe(true)
+    removeKnowledgeDatabase(root, FIXTURE_ACCOUNT_A)
+    expect(existsSync(databasePath)).toBe(false)
+  })
+
+  it('uses a bounded exact fallback for two-character Chinese queries with the trigram profile', async () => {
+    const root = makeRoot()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    await store.index({
+      conversations: [
+        {
+          conversationId: 'short-query',
+          completeSnapshot: true,
+          messages: [
+            {
+              accountId: FIXTURE_ACCOUNT_A,
+              conversationId: 'short-query',
+              messageId: 'short-query-message',
+              createTime: Date.UTC(2026, 7, 5),
+              senderId: 'fixture-member',
+              senderName: '脱敏成员',
+              kind: 'text',
+              text: '收到，明早十点。'
+            }
+          ]
+        }
+      ],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+    expect(
+      store.search({ accountId: FIXTURE_ACCOUNT_A, text: '十点', terms: ['十点'], limit: 10 })
+    ).toEqual([
+      expect.objectContaining({ messageId: 'short-query-message', conversationId: 'short-query' })
+    ])
+    store.close()
+  })
+
+  it('keeps equal message ids from different conversations as separate Evidence', async () => {
+    const root = makeRoot()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    await store.index({
+      conversations: ['conversation-a', 'conversation-b'].map((conversationId) => ({
+        conversationId,
+        completeSnapshot: true,
+        messages: [
+          {
+            accountId: FIXTURE_ACCOUNT_A,
+            conversationId,
+            messageId: 'shared-message-id',
+            createTime: Date.UTC(2026, 7, 5),
+            senderId: `${conversationId}-sender`,
+            senderName: conversationId,
+            kind: 'text',
+            text: '今天去健身。'
+          }
+        ]
+      })),
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+
+    const result = store.searchWithStatus({
+      accountId: FIXTURE_ACCOUNT_A,
+      text: '去健身',
+      terms: ['去健身'],
+      limit: 10
+    })
+    const evidence = result.evidence
+
+    expect(evidence).toHaveLength(2)
+    expect(evidence.map((item) => `${item.conversationId}:${item.messageId}`).sort()).toEqual([
+      'conversation-a:shared-message-id',
+      'conversation-b:shared-message-id'
+    ])
+    expect(result.timings).toMatchObject({
+      workerIpcMs: 0,
+      ftsMs: expect.any(Number),
+      messageLoadMs: expect.any(Number),
+      chunkExpandMs: expect.any(Number),
+      rankingMs: expect.any(Number),
+      totalMs: expect.any(Number)
+    })
+    expect(result.timings.totalMs).toBeGreaterThanOrEqual(result.timings.ftsMs)
+    store.close()
+  })
+
+  it('keeps conversation, sender and time filters when a participant question has no topic terms', async () => {
+    const root = makeRoot()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    await store.index({
+      conversations: [
+        {
+          conversationId: 'participant-query',
+          completeSnapshot: true,
+          messages: [
+            {
+              accountId: FIXTURE_ACCOUNT_A,
+              conversationId: 'participant-query',
+              messageId: 'participant-a',
+              createTime: Date.UTC(2026, 7, 5, 9),
+              senderId: 'member-a',
+              senderName: '成员甲',
+              kind: 'text',
+              text: '第一条讨论。'
+            },
+            {
+              accountId: FIXTURE_ACCOUNT_A,
+              conversationId: 'participant-query',
+              messageId: 'participant-b',
+              createTime: Date.UTC(2026, 7, 5, 10),
+              senderId: 'member-b',
+              senderName: '成员乙',
+              kind: 'text',
+              text: '第二条讨论。'
+            }
+          ]
+        }
+      ],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+    expect(
+      store.search({
+        accountId: FIXTURE_ACCOUNT_A,
+        text: '成员甲最近聊了什么',
+        terms: [],
+        conversationIds: ['participant-query'],
+        senderIds: ['member-a'],
+        startTime: Date.UTC(2026, 7, 5, 8),
+        limit: 10
+      })
+    ).toEqual([expect.objectContaining({ messageId: 'participant-a', sender: '成员甲' })])
+    store.close()
+  })
+
+  it('compresses a single-conversation recap into time chunks and deprioritizes system messages', async () => {
+    const root = makeRoot()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    const base = Date.UTC(2026, 6, 1)
+    await store.index({
+      conversations: [
+        {
+          conversationId: 'recap-query',
+          completeSnapshot: true,
+          messages: Array.from({ length: 48 }, (_, index) => ({
+            accountId: FIXTURE_ACCOUNT_A,
+            conversationId: 'recap-query',
+            messageId: `recap-${index}`,
+            createTime: base + Math.floor(index / 12) * 3 * 3600 * 1000 + (index % 12) * 60_000,
+            senderId: 'fixture-member',
+            senderName: '脱敏成员',
+            kind: index % 11 === 0 ? ('system' as const) : ('text' as const),
+            text: index % 11 === 0 ? '对方撤回了一条消息' : `第 ${index} 条健身计划和饮食安排讨论。`
+          }))
+        }
+      ],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+    const result = store.searchWithStatus({
+      accountId: FIXTURE_ACCOUNT_A,
+      text: '我和张三最近聊了什么',
+      terms: [],
+      conversationIds: ['recap-query'],
+      startTime: base,
+      limit: 100
+    })
+
+    expect(result.conversationRetrieval).toMatchObject({
+      totalMessages: 48,
+      chunkCount: 4,
+      complete: true
+    })
+    expect(result.evidence.length).toBeLessThan(48)
+    expect(new Set(result.evidence.map((item) => item.chunkId)).size).toBeGreaterThan(1)
+    expect(result.evidence.filter((item) => item.text.includes('撤回')).length).toBeLessThan(5)
+    store.close()
+  })
+
+  it('keeps late conversation slices when the recap candidate budget is reached', async () => {
+    const root = makeRoot()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    const base = Date.UTC(2026, 6, 1)
+    await store.index({
+      conversations: [
+        {
+          conversationId: 'long-recap-query',
+          completeSnapshot: true,
+          messages: Array.from({ length: 90 }, (_, index) => ({
+            accountId: FIXTURE_ACCOUNT_A,
+            conversationId: 'long-recap-query',
+            messageId: `long-recap-${index}`,
+            createTime: base + Math.floor(index / 3) * 3 * 3600 * 1000 + (index % 3) * 60_000,
+            senderId: 'fixture-member',
+            senderName: '脱敏成员',
+            kind: 'text' as const,
+            text: `第 ${index} 条近期聊天内容。`
+          }))
+        }
+      ],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER
+    })
+
+    const result = store.searchWithStatus({
+      accountId: FIXTURE_ACCOUNT_A,
+      text: '我和张三最近聊了什么',
+      terms: [],
+      conversationIds: ['long-recap-query'],
+      startTime: base,
+      limit: 100
+    })
+
+    expect(result.conversationRetrieval).toMatchObject({ chunkCount: 30, candidateMessages: 60 })
+    expect(Math.max(...result.evidence.map((item) => item.timestamp))).toBeGreaterThan(
+      base + 28 * 3 * 3600 * 1000
+    )
+    store.close()
+  })
+
+  it('provides a read-only capacity preflight before a database exists', async () => {
+    const root = makeRoot()
+    const source = createSyntheticConversation(FIXTURE_ACCOUNT_A, 'conversation-a', 0, 20, 'long')
+    const result = await estimateKnowledgeCapacityPreflight({
+      accountId: FIXTURE_ACCOUNT_A,
+      databaseRoot: root,
+      conversations: [source],
+      chunker: DEFAULT_KNOWLEDGE_CHUNKER,
+      availableDiskBytes: 1
+    })
+    expect(result.sourceMessageCount).toBe(20)
+    expect(result.voiceTranscriptCount).toBeGreaterThan(0)
+    expect(result.hasSufficientDiskSpace).toBe(false)
+    expect(existsSync(getKnowledgeDatabasePath(root, FIXTURE_ACCOUNT_A))).toBe(false)
+  })
+
+  it('indexes 100,000 desensitized messages without touching the main process database', async () => {
+    const root = makeRoot()
+    const store = new KnowledgeStore(root, FIXTURE_ACCOUNT_A, fts)
+    const started = performance.now()
+    for (let batch = 0; batch < 10; batch += 1) {
+      const source = createSyntheticConversation(
+        FIXTURE_ACCOUNT_A,
+        `performance-${batch}`,
+        batch * 10_000,
+        10_000,
+        'mixed'
+      )
+      await store.index({ conversations: [source], chunker: DEFAULT_KNOWLEDGE_CHUNKER })
+    }
+    const stats = store.getStorageStats()
+    expect(stats.databaseBytes).toBeGreaterThan(0)
+    expect(performance.now() - started).toBeLessThan(60_000)
+    store.close()
+  }, 70_000)
+})

vitest.knowledge-benchmark.config.ts

@@ -0,0 +1,11 @@
+import { defineConfig } from 'vitest/config'
+
+export default defineConfig({
+  test: {
+    name: 'knowledge-benchmark',
+    environment: 'node',
+    include: ['tests/benchmark/**/*.test.ts'],
+    fileParallelism: false,
+    testTimeout: 20 * 60 * 1000
+  }
+})