mirror of
https://wget.la/https://github.com/Wxw-Gu/WechatExplorer
synced 2026-10-04 03:01:42 +08:00
328 lines
12 KiB
TypeScript
328 lines
12 KiB
TypeScript
/**
|
||
* 精确读消息（`query_messages`）必须能读到**图片里识别出的文字**。
|
||
*
|
||
* 真机回归：问「我今早给文件传输助手发的那张图片里写了什么」，Query Agent 准确找到了
|
||
* 原始图片消息（2026/9/16 07:30:15、sender=self、type=image），却回答
|
||
* 「查询只返回图片附件，没有取得 OCR 文字」，甚至反过来建议用户"建立图片文字索引后再查"。
|
||
*
|
||
* 真机派生库 + Knowledge 实测结论（CASE A）：
|
||
* L1 artifact state=indexed / char_count=17
|
||
* L2 binding state=indexed
|
||
* L3 Knowledge image_ocr_text 与 artifact 文本**逐字相同**，chunk 里也含该文本且指向原图 messageId
|
||
* —— 即"索引早就建好了，只是查询路径没把它接出来"。缺口在 L4，不在 L1/L2/L3。
|
||
*
|
||
* 这一组测试把 L4 的契约钉死：
|
||
* 1. 图片消息的 OCR 文本必须走 `imageOcrText` + `derivedSource=image_ocr` 独立字段；
|
||
* 2. 证据**永远是原始图片消息**，不许为了 OCR 文本编造一条文字消息；
|
||
* 3. `empty`（识别过没文字）与 `not_indexed`（还没索引）必须能被区分，
|
||
* 两者都不允许模型凭想象描述图片内容。
|
||
*/
|
||
import { beforeEach, describe, expect, it, vi } from 'vitest'
|
||
import type { ImageTextIndexCoverage } from '../../src/shared/image-text-index'
|
||
import { decodeMessageRef } from '../../src/shared/local-query-api'
|
||
|
||
process.env.TZ = 'Asia/Shanghai'
|
||
|
||
const fixture = vi.hoisted(() => {
|
||
const selfImageTime = Date.parse('2026-09-16T07:30:15+08:00')
|
||
const otherImageTime = Date.parse('2026-09-16T08:10:00+08:00')
|
||
return {
|
||
selfImageTime,
|
||
otherImageTime,
|
||
contacts: [
|
||
{
|
||
m_nsUsrName: 'filehelper',
|
||
m_nsNickName: '文件传输助手',
|
||
md5: 'md5-filehelper',
|
||
type: 'user' as const
|
||
}
|
||
],
|
||
messages: [
|
||
{
|
||
id: '9001',
|
||
localId: '9001',
|
||
from: 'assistant',
|
||
// 我发出的那张图：isSender = true（自我身份来自 mesDes，不是昵称）
|
||
type: '图片',
|
||
datetime: '2026/9/16 07:30:15',
|
||
content: '',
|
||
contentData: { type: 'image', md5: 'md5-self-image', datName: 'dat-self' },
|
||
isSender: true,
|
||
name: '我',
|
||
createTime: Math.floor(selfImageTime / 1000)
|
||
},
|
||
{
|
||
id: '9002',
|
||
localId: '9002',
|
||
from: 'user',
|
||
type: '图片',
|
||
datetime: '2026/9/16 08:10:00',
|
||
content: '',
|
||
contentData: { type: 'image', md5: 'md5-other-image', datName: 'dat-other' },
|
||
isSender: false,
|
||
name: '文件传输助手',
|
||
createTime: Math.floor(otherImageTime / 1000)
|
||
}
|
||
]
|
||
}
|
||
})
|
||
|
||
vi.mock('../../src/main/services/chat-service', () => ({
|
||
isReady: () => true,
|
||
listContactsAsync: vi.fn(async () => fixture.contacts),
|
||
listMessagesAsync: vi.fn(async () => fixture.messages)
|
||
}))
|
||
|
||
import { LocalQueryApiService } from '../../src/main/services/local-query-api-service'
|
||
import { createLocalQueryToolExecutor } from '../../src/main/services/local-query-tool-executor'
|
||
import {
|
||
QueryAgentService,
|
||
type QueryAgentProvider
|
||
} from '../../src/main/services/query-agent-service'
|
||
|
||
/** 与真实派生库同形：binding 主键 = `sourceMessageId(message)` = `local:<localId>`。 */
|
||
const SELF_KEY = 'local:9001'
|
||
|
||
function coverage(overrides: Partial<ImageTextIndexCoverage> = {}): ImageTextIndexCoverage {
|
||
return {
|
||
totalImageMessages: 2,
|
||
processed: 2,
|
||
indexed: 2,
|
||
empty: 0,
|
||
missing: 0,
|
||
failed: 0,
|
||
runtimeUnavailable: 0,
|
||
pending: 0,
|
||
established: true,
|
||
complete: true,
|
||
systemicFailure: false,
|
||
countedAt: Date.parse('2026-09-16T09:00:00+08:00'),
|
||
tiers: [],
|
||
coveredToMs: null,
|
||
...overrides
|
||
}
|
||
}
|
||
|
||
type OcrFixture = Map<string, { state: string; text: string }>
|
||
|
||
function makeService(options: { ocr?: OcrFixture; coverage?: ImageTextIndexCoverage } = {}) {
|
||
const knowledge = {
|
||
search: vi.fn(async () => ({ state: 'ready', evidence: [] })),
|
||
requestCatchUp: vi.fn(() => ({ triggered: false, inProgress: false })),
|
||
waitForIndexingComplete: vi.fn(async () => false),
|
||
lastPassDurationMs: vi.fn(() => 0),
|
||
beginInteractiveQuery: vi.fn(),
|
||
endInteractiveQuery: vi.fn()
|
||
} as never
|
||
const service = new LocalQueryApiService(knowledge, () => new Date('2026-09-16T09:30:00+08:00'))
|
||
const ocr = options.ocr ?? new Map([[SELF_KEY, { state: 'indexed', text: 'ChatGPT Plus $20' }]])
|
||
service.setImageOcrEntryProvider((_conversationId, messageId) => ocr.get(messageId))
|
||
if (options.coverage !== undefined) {
|
||
service.setImageTextCoverageProvider(() => options.coverage!)
|
||
}
|
||
return service
|
||
}
|
||
|
||
const askSelfImages = {
|
||
target: { query: '文件传输助手' },
|
||
timeRange: { kind: 'all' },
|
||
direction: 'to_target',
|
||
messageTypes: ['image']
|
||
}
|
||
|
||
describe('query_messages：图片消息必须携带 OCR 派生文本', () => {
|
||
let service: LocalQueryApiService
|
||
beforeEach(() => {
|
||
service = makeService()
|
||
})
|
||
|
||
it('我发出的图片带 OCR 文本时，走 imageOcrText + derivedSource，不混进 text', async () => {
|
||
const result = await service.messages(askSelfImages as never)
|
||
|
||
expect(result.status).toBe('completed')
|
||
expect(result.returnedCount).toBe(1)
|
||
const message = result.messages![0]
|
||
|
||
// 派生文本必须单独一个字段：混进 `text` 就无法与"群友发的文字消息"区分。
|
||
expect(message.imageOcrText).toBe('ChatGPT Plus $20')
|
||
expect(message.derivedSource).toBe('image_ocr')
|
||
expect(message.imageTextState).toBe('indexed')
|
||
expect(message.text).toBeUndefined()
|
||
expect(message.attachment).toEqual({ kind: 'image' })
|
||
})
|
||
|
||
it('对方的图片不会被贴错 OCR 文本（键必须按消息身份匹配）', async () => {
|
||
const result = await service.messages({
|
||
...askSelfImages,
|
||
direction: 'from_target'
|
||
} as never)
|
||
|
||
expect(result.returnedCount).toBe(1)
|
||
// OTHER_KEY 在派生库里没有绑定 → 只能是 not_indexed，绝不能借用另一条消息的文本。
|
||
expect(result.messages![0].imageOcrText).toBeUndefined()
|
||
expect(result.messages![0].imageTextState).toBe('not_indexed')
|
||
})
|
||
|
||
it('识别过但图里没文字 → empty（已知结论），不是 not_indexed', async () => {
|
||
const empty = makeService({ ocr: new Map([[SELF_KEY, { state: 'empty', text: '' }]]) })
|
||
const result = await empty.messages(askSelfImages as never)
|
||
|
||
// `empty` 与 `not_indexed` 必须能分辨：前者是"已经知道没文字"，
|
||
// 后者是"还不知道"。把两者混起来，模型就会在没索引时断言"图里没内容"。
|
||
expect(result.messages![0].imageTextState).toBe('empty')
|
||
expect(result.messages![0].imageOcrText).toBeUndefined()
|
||
})
|
||
|
||
it('图片文字索引未建立时，tool result 明确带上 not_built 覆盖度', async () => {
|
||
const notBuilt = makeService({
|
||
ocr: new Map(),
|
||
coverage: coverage({
|
||
processed: 0,
|
||
indexed: 0,
|
||
pending: 0,
|
||
established: false,
|
||
complete: false
|
||
})
|
||
})
|
||
const result = await notBuilt.messages(askSelfImages as never)
|
||
|
||
expect(result.messages![0].imageTextState).toBe('not_indexed')
|
||
expect(result.imageOcrCoverage?.state).toBe('not_built')
|
||
expect(result.imageOcrCoverage?.summary).toContain('尚未建立')
|
||
})
|
||
|
||
it('覆盖度部分完成时，summary 必须说明结果可能不完整（不许当 complete）', async () => {
|
||
const partial = makeService({
|
||
coverage: coverage({
|
||
totalImageMessages: 100,
|
||
processed: 30,
|
||
indexed: 30,
|
||
pending: 70,
|
||
complete: false
|
||
})
|
||
})
|
||
const result = await partial.messages(askSelfImages as never)
|
||
|
||
expect(result.imageOcrCoverage?.state).toBe('partial')
|
||
expect(result.imageOcrCoverage?.summary).toContain('30')
|
||
expect(result.imageOcrCoverage?.summary).toContain('100')
|
||
})
|
||
|
||
it('普通文字消息完全不受影响（对照组）', async () => {
|
||
const plain = makeService({ ocr: new Map() })
|
||
const result = await plain.messages({
|
||
target: { query: '文件传输助手' },
|
||
timeRange: { kind: 'all' },
|
||
direction: 'to_target',
|
||
messageTypes: ['text']
|
||
} as never)
|
||
|
||
// 图片那两条都是 image，文字查询必然是 0 条 —— 关键是**不能**因为接了 OCR 路径
|
||
// 就凭空多出消息。
|
||
expect(result.returnedCount).toBe(0)
|
||
})
|
||
})
|
||
|
||
describe('Query Agent：证据永远是原始图片消息', () => {
|
||
function provider(
|
||
responses: Array<Awaited<ReturnType<QueryAgentProvider['chatWithTools']>>>
|
||
): QueryAgentProvider {
|
||
return {
|
||
getRuntimeConfig: () => ({
|
||
configured: true,
|
||
providerName: 'Fixture Provider',
|
||
model: 'fixture-model',
|
||
modelName: 'Fixture Model'
|
||
}),
|
||
chatWithTools: vi.fn(async () => responses.shift() || { success: true, data: 'done' })
|
||
}
|
||
}
|
||
|
||
const selfImageArgs = JSON.stringify({
|
||
target: { query: '文件传输助手' },
|
||
timeRange: { kind: 'all' },
|
||
temporalBasis: { kind: 'none' },
|
||
direction: 'to_target',
|
||
messageTypes: ['image']
|
||
})
|
||
|
||
it('模型能在 Tool Result 里读到 imageOcrText，且证据仍指向原图 messageRef', async () => {
|
||
const configured = provider([
|
||
{
|
||
success: true,
|
||
toolCalls: [{ id: 'c1', name: 'query_messages', arguments: selfImageArgs }]
|
||
},
|
||
{ success: true, data: '那张图片里的文字是 ChatGPT Plus $20。' }
|
||
])
|
||
const service = makeService()
|
||
const result = await new QueryAgentService(
|
||
configured,
|
||
createLocalQueryToolExecutor(service)
|
||
).run('我今早给文件传输助手发的那张图片里写了什么')
|
||
|
||
const calls = vi.mocked(configured.chatWithTools).mock.calls
|
||
const toolResult = JSON.parse(
|
||
String(calls[1]?.[0].find((message) => message.role === 'tool')?.content)
|
||
) as Record<string, any>
|
||
|
||
// 1) 模型确实拿到了派生文本（这正是真机上缺的那一环）
|
||
expect(toolResult.messages?.[0].imageOcrText).toBe('ChatGPT Plus $20')
|
||
expect(toolResult.messages?.[0].derivedSource).toBe('image_ocr')
|
||
expect(toolResult.messages?.[0].imageTextState).toBe('indexed')
|
||
|
||
// 2) 证据只有一条，且解出来就是**原始图片消息**（不是虚构的 OCR 文字消息）
|
||
expect(result.evidence).toHaveLength(1)
|
||
expect(decodeMessageRef(result.evidence![0].messageRef)).toEqual({
|
||
conversationId: 'md5-filehelper',
|
||
messageId: '9001'
|
||
})
|
||
expect(result.evidence![0].messageType).toBe('image')
|
||
|
||
// 3) UI 拿得到来源语义（「图片文字」标记），且 snippet 不进模型上下文之外的重复字段
|
||
expect(result.evidence![0].derivedSource).toBe('image_ocr')
|
||
expect(result.evidence![0].imageOcrText).toBe('ChatGPT Plus $20')
|
||
})
|
||
|
||
it('系统提示词把图片文字的三态语义写死，并禁止凭空建议建立索引', async () => {
|
||
const scripted = provider([{ success: true, data: 'ok' }])
|
||
const service = makeService()
|
||
void new QueryAgentService(scripted, createLocalQueryToolExecutor(service)).run(
|
||
'我今早给文件传输助手发的那张图片里写了什么'
|
||
)
|
||
|
||
const systemPrompt = String(vi.mocked(scripted.chatWithTools).mock.calls[0]?.[0]?.[0]?.content)
|
||
expect(systemPrompt).toContain('imageOcrText')
|
||
// 三态必须分别说清楚
|
||
expect(systemPrompt).toContain('indexed')
|
||
expect(systemPrompt).toContain('empty')
|
||
expect(systemPrompt).toContain('not_indexed')
|
||
// OCR 不是看图：empty 时不许猜画面
|
||
expect(systemPrompt).toContain('OCR 不是看图')
|
||
// 不许无条件建议"先建立图片文字索引再查"
|
||
expect(systemPrompt).toContain('建立图片文字索引')
|
||
})
|
||
|
||
it('索引已建好的情况下，模型不会拿到任何"还没建立"的误导信号', async () => {
|
||
const configured = provider([
|
||
{
|
||
success: true,
|
||
toolCalls: [{ id: 'c1', name: 'query_messages', arguments: selfImageArgs }]
|
||
},
|
||
{ success: true, data: '那张图里有 ChatGPT Plus $20。' }
|
||
])
|
||
const service = makeService({ coverage: coverage() })
|
||
await new QueryAgentService(configured, createLocalQueryToolExecutor(service)).run(
|
||
'我今早给文件传输助手发的那张图片里写了什么'
|
||
)
|
||
|
||
const calls = vi.mocked(configured.chatWithTools).mock.calls
|
||
const toolResult = JSON.parse(
|
||
String(calls[1]?.[0].find((message) => message.role === 'tool')?.content)
|
||
) as Record<string, any>
|
||
|
||
// 覆盖度是 complete 且带了派生文本 → 模型没有任何理由说"没有取得 OCR 文字"。
|
||
expect(toolResult.imageOcrCoverage?.state).toBe('complete')
|
||
expect(toolResult.messages?.[0].imageOcrText).toBe('ChatGPT Plus $20')
|
||
})
|
||
})
|