open-source/WechatExplorer
1
0
Fork 0
mirror of https://wget.la/https://github.com/Wxw-Gu/WechatExplorer synced 2026-08-18 20:19:09 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: 知识库更新，优化批量语音转写 - 合并索引 - 增加会话级批量转写处理 - 补充语音转写回归测试

Browse Source
This commit is contained in:
电摇小子
2026-08-06 20:31:08 +08:00
parent a0e8ab278f
commit 0ec2e6a0be
34 changed files with 2589 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/main/knowledge/knowledge-search-service.ts
+226 -7
View File
@@ -10,17 +10,31 @@ import type {
KnowledgeSearchResult,
KnowledgeSourceMessage
} from '../../shared/knowledge'
import type {
VoiceMessageReference,
VoiceTranscriptSnapshot,
VoiceTranscriptUpdate
} from '../../shared/voice-recognition'
import {
DEFAULT_KNOWLEDGE_CHUNKER,
DEFAULT_KNOWLEDGE_FTS_CONFIG,
emptyKnowledgeSearchTimings
} from '../../shared/knowledge'
import { KnowledgeService } from './knowledge-service'
import { voiceAccountIdentity, voiceMessageIdentity } from '../voice-pipeline/voice-message-identity'
const FALLBACK_LIMIT = 240
const MAX_SENDER_NAME_CONVERSATIONS = 8
const MAX_CONVERSATION_FILTERS_PER_WORKER_SEARCH = 700
type PendingVoiceTranscriptIndex = {
update: VoiceTranscriptUpdate
waiters: Array<{
resolve: () => void
reject: (error: unknown) => void
}>
}
function looksLikeOpaqueSenderId(value: string | undefined): boolean {
const normalized = value?.trim() || ''
return (
@@ -113,11 +127,12 @@ function sourceTextAndAttachment(message: chat.FormattedMessage): {
function toSourceMessage(
accountId: string,
conversationId: string,
message: chat.FormattedMessage
message: chat.FormattedMessage,
transcriptOverride?: string
): KnowledgeSourceMessage | null {
if (!message.createTime) return null
const extracted = sourceTextAndAttachment(message)
const voiceTranscript = message.voiceTranscript?.trim() || undefined
const voiceTranscript = transcriptOverride?.trim() || message.voiceTranscript?.trim() || undefined
if (!extracted.text && !extracted.attachment && !voiceTranscript) return null
return {
accountId,
@@ -160,6 +175,12 @@ export class KnowledgeSearchService {
private readonly statusByAccount = new Map<string, KnowledgeRuntimeStatus>()
private readonly statusListeners = new Set<(status: KnowledgeRuntimeStatus) => void>()
private wcdbReadTail: Promise<void> = Promise.resolve()
private voiceTranscriptResolver:
| ((reference: VoiceMessageReference) => VoiceTranscriptSnapshot)
| undefined
private voiceIndexTail: Promise<void> = Promise.resolve()
private voiceIndexFlushScheduled = false
private readonly pendingVoiceIndexes = new Map<string, PendingVoiceTranscriptIndex>()
constructor(userDataPath: string, workerPath: string) {
this.service = new KnowledgeService(userDataPath, workerPath)
@@ -200,6 +221,73 @@ export class KnowledgeSearchService {
return started
}
/**
* The voice cache remains owned by the voice pipeline. Knowledge only reads
* a current-account snapshot while constructing a derived local index.
*/
setVoiceTranscriptResolver(
resolver: (reference: VoiceMessageReference) => VoiceTranscriptSnapshot
): void {
this.voiceTranscriptResolver = resolver
}
/**
* A successful recognition updates its source conversation. Consecutive
* updates for the same conversation are coalesced because a complete
* snapshot already includes every finished transcript for that conversation.
*/
indexVoiceTranscript(update: VoiceTranscriptUpdate): Promise<void> {
const key = this.voiceIndexKey(update)
return new Promise<void>((resolve, reject) => {
const existing = this.pendingVoiceIndexes.get(key)
if (existing) {
existing.update = update
existing.waiters.push({ resolve, reject })
} else {
this.pendingVoiceIndexes.set(key, {
update,
waiters: [{ resolve, reject }]
})
}
this.scheduleVoiceIndexFlush()
})
}
private voiceIndexKey(update: VoiceTranscriptUpdate): string {
return `${update.accountIdentity}:${update.reference.sessionId}`
}
private scheduleVoiceIndexFlush(): void {
if (this.voiceIndexFlushScheduled) return
this.voiceIndexFlushScheduled = true
const task = this.voiceIndexTail.then(() => this.flushPendingVoiceIndexes())
this.voiceIndexTail = task.catch(() => undefined)
void task.then(
() => this.finishVoiceIndexFlush(),
() => this.finishVoiceIndexFlush()
)
}
private async flushPendingVoiceIndexes(): Promise<void> {
while (this.pendingVoiceIndexes.size) {
const pending = Array.from(this.pendingVoiceIndexes.values())
this.pendingVoiceIndexes.clear()
for (const entry of pending) {
try {
await this.indexVoiceTranscriptNow(entry.update)
entry.waiters.forEach((waiter) => waiter.resolve())
} catch (error) {
entry.waiters.forEach((waiter) => waiter.reject(error))
}
}
}
}
private finishVoiceIndexFlush(): void {
this.voiceIndexFlushScheduled = false
if (this.pendingVoiceIndexes.size) this.scheduleVoiceIndexFlush()
}
async search(request: KnowledgeSearchIpcRequest): Promise<KnowledgeSearchIpcResult> {
const accountId = this.currentAccountId()
if (!accountId) return this.searchFallback(request, 'unavailable')
@@ -282,7 +370,7 @@ export class KnowledgeSearchService {
// 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))
.map((message) => this.toSourceMessage(accountId, contact.md5, message))
.filter((message): message is KnowledgeSourceMessage => Boolean(message))
await this.service.index(
{
@@ -352,10 +440,11 @@ export class KnowledgeSearchService {
const messages = await this.listMessages(contact.md5, request.startTime, request.endTime)
totalMessages += messages.length
for (const message of messages) {
const hydrated = this.withVoiceTranscript(message)
matches.push({
contact,
message,
score: fallbackTermScore(message, terms)
message: hydrated,
score: fallbackTermScore(hydrated, terms)
})
}
}
@@ -393,7 +482,12 @@ export class KnowledgeSearchService {
sender: message.isSender ? '我' : message.name || '未知成员',
timestamp: (message.createTime || 0) * 1000,
messageIds: [sourceMessageId(message)],
text: sourceTextAndAttachment(message).text || message.content || `[${message.type}]`,
sourceKind: sourceKind(message),
text:
this.toSourceMessage('fallback', contact.md5, message)?.voiceTranscript ||
sourceTextAndAttachment(message).text ||
message.content ||
`[${message.type}]`,
score: -score
}))
}
@@ -466,6 +560,29 @@ export class KnowledgeSearchService {
const mergeRankingMs = Date.now() - mergeStartedAt
timings.rankingMs += mergeRankingMs
timings.totalMs += mergeRankingMs
const voiceCoverageParts = partialResults
.map((result) => result.voiceCoverage)
.filter((coverage): coverage is NonNullable<typeof coverage> => Boolean(coverage))
const voiceCoverage = voiceCoverageParts.length
? voiceCoverageParts.reduce(
(total, coverage) => ({
voiceMessageCount: total.voiceMessageCount + coverage.voiceMessageCount,
transcribedVoiceCount: total.transcribedVoiceCount + coverage.transcribedVoiceCount,
failedVoiceCount: total.failedVoiceCount + coverage.failedVoiceCount,
voiceCoverageComplete: false
}),
{
voiceMessageCount: 0,
transcribedVoiceCount: 0,
failedVoiceCount: 0,
voiceCoverageComplete: false
}
)
: undefined
if (voiceCoverage) {
voiceCoverage.voiceCoverageComplete =
voiceCoverage.voiceMessageCount === voiceCoverage.transcribedVoiceCount
}
return {
state: partialResults.some((result) => result.state === 'ready')
? 'ready'
@@ -475,7 +592,8 @@ export class KnowledgeSearchService {
indexedMessageCount: Math.max(...partialResults.map((result) => result.indexedMessageCount)),
indexedChunkCount: Math.max(...partialResults.map((result) => result.indexedChunkCount)),
evidence: mergedEvidence,
timings
timings,
voiceCoverage
}
}
@@ -507,6 +625,107 @@ export class KnowledgeSearchService {
return this.enqueueWcdbRead(() => chat.listMessagesAsync(conversationId, startTime, endTime))
}
private withVoiceTranscript(message: chat.FormattedMessage): chat.FormattedMessage {
const reference = this.voiceReferenceFromMessage(message)
if (!reference || !this.voiceTranscriptResolver) return message
const snapshot = this.voiceTranscriptResolver(reference)
if (snapshot.state !== 'transcribed' || !snapshot.transcript?.trim()) return message
return { ...message, voiceTranscript: snapshot.transcript.trim() }
}
private toSourceMessage(
accountId: string,
conversationId: string,
message: chat.FormattedMessage,
transcriptOverride?: string,
stateOverride?: 'pending' | 'transcribed' | 'failed'
): KnowledgeSourceMessage | null {
const reference = this.voiceReferenceFromMessage(message)
const snapshot = reference ? this.voiceTranscriptResolver?.(reference) : undefined
const hydrated = this.withVoiceTranscript(message)
const source = toSourceMessage(
accountId,
conversationId,
hydrated,
transcriptOverride
)
if (!source || source.kind !== 'voice') return source
return {
...source,
voiceTranscriptState:
stateOverride ||
(transcriptOverride?.trim() ? 'transcribed' : undefined) ||
snapshot?.state ||
(source.voiceTranscript ? 'transcribed' : 'pending')
}
}
private voiceReferenceFromMessage(
message: chat.FormattedMessage
): VoiceMessageReference | undefined {
if (message.type !== '语音' || !message.sessionId || message.localId === undefined || !message.createTime) {
return undefined
}
return {
sessionId: message.sessionId,
localId: message.localId,
createTime: message.createTime,
svrId: message.serverId
}
}
private async indexVoiceTranscriptNow(update: VoiceTranscriptUpdate): Promise<void> {
if (!chat.isReady()) return
if (update.state === 'transcribed' && !update.transcript?.trim()) return
if (voiceAccountIdentity(chat.getCurrentAccountRoot()) !== update.accountIdentity) {
return
}
const accountId = this.currentAccountId()
if (!accountId) return
const activeIndex = this.indexing.get(accountId)
if (activeIndex) await activeIndex
if (voiceAccountIdentity(chat.getCurrentAccountRoot()) !== update.accountIdentity) {
return
}
const contacts = await this.listContacts()
const contact = contacts.find((item) => item.m_nsUsrName === update.reference.sessionId)
if (!contact) return
const messages = await this.listMessages(contact.md5)
const sourceMessages = messages
.map((message) => {
const reference = this.voiceReferenceFromMessage(message)
const transcriptOverride =
reference && voiceMessageIdentity(reference) === update.messageIdentity
? update.transcript
: undefined
const stateOverride =
reference && voiceMessageIdentity(reference) === update.messageIdentity
? update.state
: undefined
return this.toSourceMessage(
accountId,
contact.md5,
message,
transcriptOverride,
stateOverride
)
})
.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
})
await this.refreshStatus(accountId)
}
private async toKnowledgeResult(
result: KnowledgeSearchResult
): Promise<KnowledgeSearchIpcResult> {
src/main/knowledge/knowledge-store.ts
+61 -8
View File
@@ -8,6 +8,7 @@ import type {
KnowledgeChunk,
KnowledgeConversationRetrieval,
KnowledgeEvidence,
KnowledgeVoiceCoverage,
KnowledgeFtsConfig,
KnowledgeIndexProgress,
KnowledgeIndexRequest,
@@ -412,7 +413,7 @@ export class KnowledgeStore {
const messages = asRows(
this.database
.prepare(
`SELECT message_id, create_time, searchable_text, sender_id, sender_name
`SELECT message_id, create_time, searchable_text, kind, sender_id, sender_name
FROM knowledge_messages
WHERE conversation_id = ? AND message_id IN (${messageIds.map(() => '?').join(', ')})`
)
@@ -446,6 +447,7 @@ export class KnowledgeStore {
sender: String(row.sender_name || row.sender_id || '未知成员'),
timestamp: Number(row.create_time),
messageIds,
sourceKind: String(row.kind) as KnowledgeEvidence['sourceKind'],
text: String(row.searchable_text),
score: Number(chunk.score) - item.termScore / 1000
})
@@ -518,7 +520,7 @@ export class KnowledgeStore {
return asRows(
this.database
.prepare(
`SELECT m.conversation_id, m.message_id, m.create_time, m.searchable_text, m.sender_id, m.sender_name
`SELECT m.conversation_id, m.message_id, m.create_time, m.searchable_text, m.kind, m.sender_id, m.sender_name
FROM knowledge_messages m
WHERE ${clauses.join(' AND ')}
ORDER BY m.create_time DESC
@@ -547,6 +549,7 @@ export class KnowledgeStore {
sender: String(row.sender_name || row.sender_id || '未知成员'),
timestamp: Number(row.create_time),
messageIds: [messageId],
sourceKind: String(row.kind) as KnowledgeEvidence['sourceKind'],
text,
score: -termScore / 1000
}))
@@ -664,7 +667,7 @@ export class KnowledgeStore {
evidence: asRows(
this.database
.prepare(
`SELECT m.conversation_id, m.message_id, m.create_time, m.searchable_text, m.sender_id, m.sender_name
`SELECT m.conversation_id, m.message_id, m.create_time, m.searchable_text, m.kind, m.sender_id, m.sender_name
FROM knowledge_messages m
WHERE ${clauses.join(' AND ')}
ORDER BY m.create_time DESC
@@ -691,6 +694,7 @@ export class KnowledgeStore {
sender: String(row.sender_name || row.sender_id || '未知成员'),
timestamp: Number(row.create_time),
messageIds: chunk ? chunk.map((item) => String(item.message_id)) : [messageId],
sourceKind: String(row.kind) as KnowledgeEvidence['sourceKind'],
text: String(row.searchable_text),
score: String(row.kind) === 'system' ? 1 : 0
}
@@ -705,7 +709,45 @@ export class KnowledgeStore {
// safe to query and avoid falling back to a second scan of the source archive.
evidence: measured?.evidence || [],
timings: measured?.timings || emptyKnowledgeSearchTimings(),
conversationRetrieval: measured?.conversationRetrieval
conversationRetrieval: measured?.conversationRetrieval,
voiceCoverage: this.getVoiceCoverage(query)
}
}
private getVoiceCoverage(query: KnowledgeQuery): KnowledgeVoiceCoverage {
const clauses = ["kind = 'voice'"]
const values: (string | number)[] = []
const conversationIds = Array.from(
new Set([...(query.conversationIds || []), ...(query.conversationId ? [query.conversationId] : [])])
).filter(Boolean)
if (conversationIds.length) {
clauses.push(`conversation_id IN (${conversationIds.map(() => '?').join(', ')})`)
values.push(...conversationIds)
}
if (query.startTime !== undefined) {
clauses.push('create_time >= ?')
values.push(query.startTime)
}
if (query.endTime !== undefined) {
clauses.push('create_time <= ?')
values.push(query.endTime)
}
const row = this.database
.prepare(
`SELECT COUNT(*) AS total,
SUM(CASE WHEN voice_transcript IS NOT NULL AND trim(voice_transcript) <> '' THEN 1 ELSE 0 END) AS transcribed,
SUM(CASE WHEN voice_transcript_state = 'failed' THEN 1 ELSE 0 END) AS failed
FROM knowledge_messages WHERE ${clauses.join(' AND ')}`
)
.get(...values) as DbRow | undefined
const voiceMessageCount = Number(row?.total || 0)
const transcribedVoiceCount = Number(row?.transcribed || 0)
const failedVoiceCount = Number(row?.failed || 0)
return {
voiceMessageCount,
transcribedVoiceCount,
failedVoiceCount,
voiceCoverageComplete: voiceMessageCount === transcribedVoiceCount
}
}
@@ -731,6 +773,7 @@ export class KnowledgeStore {
sender_name TEXT,
attachment_json TEXT,
voice_transcript TEXT,
voice_transcript_state TEXT,
PRIMARY KEY (conversation_id, message_id)
) STRICT;
CREATE INDEX IF NOT EXISTS knowledge_messages_conversation_time
@@ -764,6 +807,14 @@ export class KnowledgeStore {
updated_at INTEGER NOT NULL
) STRICT;
`)
const messageColumns = new Set(
asRows(this.database.prepare('PRAGMA table_info(knowledge_messages)').all()).map((row) =>
String(row.name)
)
)
if (!messageColumns.has('voice_transcript_state')) {
this.database.exec('ALTER TABLE knowledge_messages ADD COLUMN voice_transcript_state TEXT')
}
this.writeMetaIfMissing('schema_version', String(KNOWLEDGE_SCHEMA_VERSION))
const storedAccount = this.readMeta('account_id')
if (storedAccount && storedAccount !== this.accountId) {
@@ -921,8 +972,8 @@ export class KnowledgeStore {
const upsert = this.database.prepare(
`INSERT INTO knowledge_messages (
account_id, conversation_id, message_id, create_time, content_hash, searchable_text,
kind, sender_id, sender_name, attachment_json, voice_transcript
) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
kind, sender_id, sender_name, attachment_json, voice_transcript, voice_transcript_state
) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
ON CONFLICT(conversation_id, message_id) DO UPDATE SET
create_time = excluded.create_time,
content_hash = excluded.content_hash,
@@ -931,7 +982,8 @@ export class KnowledgeStore {
sender_id = excluded.sender_id,
sender_name = excluded.sender_name,
attachment_json = excluded.attachment_json,
voice_transcript = excluded.voice_transcript`
voice_transcript = excluded.voice_transcript,
voice_transcript_state = excluded.voice_transcript_state`
)
for (let index = 0; index < messages.length; index += 1) {
this.assertNotAborted(signal)
@@ -947,7 +999,8 @@ export class KnowledgeStore {
message.senderId ?? null,
message.senderName ?? null,
message.attachment ? encodedJson(message.attachment) : null,
message.voiceTranscript ?? null
message.voiceTranscript ?? null,
message.voiceTranscriptState ?? null
)
if (index % YIELD_EVERY === 0) {
onProgress(index + 1, 0)
src/main/knowledge/normalizer.ts
+1
View File
@@ -44,6 +44,7 @@ export function normalizeKnowledgeMessage(
createTime: source.createTime,
senderId: source.senderId || '',
kind: source.kind,
voiceTranscriptState: source.voiceTranscriptState || '',
searchableText
})
)