mirror of
https://github.com/hellodigua/ChatLab.git
synced 2026-04-30 16:52:45 +08:00
bfca7ff133
- llm:chatStream：正确处理 error/done 事件，防止终态重复发送 - llmComplete：检测 completeSimple 返回的 error stopReason 并抛出异常 - SessionTimeline：摘要生成失败时显示 toast 提示替代 console.log Made-with: Cursor
573 lines
16 KiB
TypeScript
573 lines
16 KiB
TypeScript
/**
|
||
* 会话摘要生成服务
|
||
*
|
||
* 利用 LLM 为会话生成摘要
|
||
* - 智能预处理：过滤无意义内容（纯表情、单字回复等）
|
||
* - 根据消息数量智能调整摘要长度
|
||
* - 超长会话采用 Map-Reduce 策略
|
||
*/
|
||
|
||
import Database from 'better-sqlite3'
|
||
import { completeSimple, type TextContent as PiTextContent } from '@mariozechner/pi-ai'
|
||
import { getActiveConfig, buildPiModel } from '../llm'
|
||
import { getDbPath, openDatabase } from '../../database/core'
|
||
import { aiLogger } from '../logger'
|
||
import { t } from '../../i18n'
|
||
|
||
/** 调用 LLM 生成文本（直接使用 pi-ai completeSimple） */
|
||
async function llmComplete(
|
||
systemPrompt: string,
|
||
userPrompt: string,
|
||
options?: { temperature?: number; maxTokens?: number }
|
||
): Promise<string> {
|
||
const activeConfig = getActiveConfig()
|
||
if (!activeConfig) {
|
||
throw new Error(t('llm.notConfigured'))
|
||
}
|
||
|
||
const piModel = buildPiModel(activeConfig)
|
||
const now = Date.now()
|
||
|
||
const result = await completeSimple(
|
||
piModel,
|
||
{
|
||
systemPrompt,
|
||
messages: [{ role: 'user', content: userPrompt, timestamp: now }],
|
||
},
|
||
{
|
||
apiKey: activeConfig.apiKey,
|
||
temperature: options?.temperature,
|
||
maxTokens: options?.maxTokens,
|
||
}
|
||
)
|
||
|
||
if (result.stopReason === 'error' || result.stopReason === 'aborted') {
|
||
throw new Error(result.errorMessage || t('llm.callFailed'))
|
||
}
|
||
|
||
return result.content
|
||
.filter((item): item is PiTextContent => item.type === 'text')
|
||
.map((item) => item.text)
|
||
.join('')
|
||
}
|
||
|
||
/** 最小消息数阈值（少于此数量不生成摘要） */
|
||
const MIN_MESSAGE_COUNT = 3
|
||
|
||
/** 单次 LLM 调用的最大内容字符数（约 2000 tokens，留安全余量） */
|
||
const MAX_CONTENT_PER_CALL = 8000
|
||
|
||
/** 需要分段处理的阈值 */
|
||
const SEGMENT_THRESHOLD = 8000
|
||
|
||
// ==================== 数据库操作函数（独立于 Worker） ====================
|
||
|
||
interface SessionMessagesResult {
|
||
messageCount: number
|
||
messages: Array<{
|
||
senderName: string
|
||
content: string | null
|
||
}>
|
||
}
|
||
|
||
/**
|
||
* 获取会话消息（主进程版本，使用 database/core）
|
||
*/
|
||
function getSessionMessagesForSummary(
|
||
dbSessionId: string,
|
||
chatSessionId: number,
|
||
limit: number = 500
|
||
): SessionMessagesResult | null {
|
||
const db = openDatabase(dbSessionId, true)
|
||
if (!db) {
|
||
aiLogger.error('Summary', `Failed to open database: ${dbSessionId}`)
|
||
return null
|
||
}
|
||
|
||
try {
|
||
// 获取会话消息
|
||
const messagesSql = `
|
||
SELECT
|
||
COALESCE(mb.group_nickname, mb.account_name, mb.platform_id) as senderName,
|
||
m.content
|
||
FROM message_context mc
|
||
JOIN message m ON m.id = mc.message_id
|
||
JOIN member mb ON mb.id = m.sender_id
|
||
WHERE mc.session_id = ?
|
||
ORDER BY m.ts ASC
|
||
LIMIT ?
|
||
`
|
||
const messages = db.prepare(messagesSql).all(chatSessionId, limit) as Array<{
|
||
senderName: string
|
||
content: string | null
|
||
}>
|
||
|
||
return {
|
||
messageCount: messages.length,
|
||
messages,
|
||
}
|
||
} catch (error) {
|
||
aiLogger.error('Summary', `Failed to get session messages: ${error}`)
|
||
return null
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 保存会话摘要（主进程版本）
|
||
*/
|
||
function saveSessionSummaryToDb(dbSessionId: string, chatSessionId: number, summary: string): void {
|
||
const dbPath = getDbPath(dbSessionId)
|
||
const db = new Database(dbPath)
|
||
|
||
try {
|
||
db.prepare('UPDATE chat_session SET summary = ? WHERE id = ?').run(summary, chatSessionId)
|
||
} finally {
|
||
db.close()
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 获取会话摘要（主进程版本）
|
||
*/
|
||
function getSessionSummaryFromDb(dbSessionId: string, chatSessionId: number): string | null {
|
||
const db = openDatabase(dbSessionId, true)
|
||
if (!db) {
|
||
return null
|
||
}
|
||
|
||
try {
|
||
const result = db.prepare('SELECT summary FROM chat_session WHERE id = ?').get(chatSessionId) as
|
||
| { summary: string | null }
|
||
| undefined
|
||
return result?.summary || null
|
||
} catch {
|
||
return null
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 根据消息数量计算摘要长度限制
|
||
* - 3-10 条消息：50 字
|
||
* - 11-30 条消息：80 字
|
||
* - 31-100 条消息：120 字
|
||
* - 100+ 条消息：200 字
|
||
*/
|
||
function getSummaryLengthLimit(messageCount: number): number {
|
||
if (messageCount <= 10) return 50
|
||
if (messageCount <= 30) return 80
|
||
if (messageCount <= 100) return 120
|
||
return 200
|
||
}
|
||
|
||
/**
|
||
* 判断消息是否有意义（用于过滤）
|
||
* 支持中英文内容过滤
|
||
*/
|
||
function isValidMessage(content: string): boolean {
|
||
const trimmed = content.trim()
|
||
|
||
// 过滤空内容
|
||
if (!trimmed) return false
|
||
|
||
// 过滤单字/双字无意义回复（中文）
|
||
if (trimmed.length <= 2) {
|
||
const meaningfulShortZh = ['好的', '不是', '是的', '可以', '不行', '好吧', '明白', '知道', '同意']
|
||
if (!meaningfulShortZh.includes(trimmed)) return false
|
||
}
|
||
|
||
// 过滤短无意义回复（英文，不区分大小写）
|
||
const lowerTrimmed = trimmed.toLowerCase()
|
||
const meaninglessShortEn = [
|
||
'ok',
|
||
'k',
|
||
'yes',
|
||
'no',
|
||
'ya',
|
||
'yep',
|
||
'nope',
|
||
'lol',
|
||
'haha',
|
||
'hehe',
|
||
'hmm',
|
||
'ah',
|
||
'oh',
|
||
'wow',
|
||
'thx',
|
||
'ty',
|
||
'np',
|
||
'gg',
|
||
'brb',
|
||
'idk',
|
||
]
|
||
if (meaninglessShortEn.includes(lowerTrimmed)) return false
|
||
|
||
// 过滤纯表情消息
|
||
const emojiOnlyPattern = /^[\p{Emoji}\s[\]（）()]+$/u
|
||
if (emojiOnlyPattern.test(trimmed)) return false
|
||
|
||
// 过滤占位符文本（中文 + 英文）
|
||
const placeholders = [
|
||
// 中文占位符（QQ/微信导出格式）
|
||
'[图片]',
|
||
'[语音]',
|
||
'[视频]',
|
||
'[文件]',
|
||
'[表情]',
|
||
'[动画表情]',
|
||
'[位置]',
|
||
'[名片]',
|
||
'[红包]',
|
||
'[转账]',
|
||
'[撤回消息]',
|
||
// 英文占位符
|
||
'[image]',
|
||
'[voice]',
|
||
'[video]',
|
||
'[file]',
|
||
'[sticker]',
|
||
'[animated sticker]',
|
||
'[location]',
|
||
'[contact]',
|
||
'[red packet]',
|
||
'[transfer]',
|
||
'[recalled message]',
|
||
'[photo]',
|
||
'[audio]',
|
||
'[gif]',
|
||
]
|
||
if (placeholders.some((p) => lowerTrimmed === p.toLowerCase())) return false
|
||
|
||
// 过滤系统消息（中文：入群、退群等）
|
||
const systemPatternsZh = [/^.*邀请.*加入了群聊$/, /^.*退出了群聊$/, /^.*撤回了一条消息$/, /^你撤回了一条消息$/]
|
||
if (systemPatternsZh.some((p) => p.test(trimmed))) return false
|
||
|
||
// 过滤系统消息（英文）
|
||
const systemPatternsEn = [
|
||
/^.*invited.*to the group$/i,
|
||
/^.*left the group$/i,
|
||
/^.*recalled a message$/i,
|
||
/^you recalled a message$/i,
|
||
/^.*joined the group$/i,
|
||
/^.*has been removed$/i,
|
||
]
|
||
if (systemPatternsEn.some((p) => p.test(trimmed))) return false
|
||
|
||
return true
|
||
}
|
||
|
||
/**
|
||
* 预处理消息：过滤无意义内容
|
||
*/
|
||
function preprocessMessages(
|
||
messages: Array<{ senderName: string; content: string | null }>
|
||
): Array<{ senderName: string; content: string }> {
|
||
return messages
|
||
.filter((m) => m.content && isValidMessage(m.content))
|
||
.map((m) => ({ senderName: m.senderName, content: m.content!.trim() }))
|
||
}
|
||
|
||
/**
|
||
* 格式化消息为文本
|
||
*/
|
||
function formatMessages(messages: Array<{ senderName: string; content: string }>): string {
|
||
return messages.map((m) => `${m.senderName}: ${m.content}`).join('\n')
|
||
}
|
||
|
||
/**
|
||
* 将消息分成多个段落
|
||
*/
|
||
function splitIntoSegments(
|
||
messages: Array<{ senderName: string; content: string }>,
|
||
maxCharsPerSegment: number
|
||
): Array<Array<{ senderName: string; content: string }>> {
|
||
const segments: Array<Array<{ senderName: string; content: string }>> = []
|
||
let currentSegment: Array<{ senderName: string; content: string }> = []
|
||
let currentLength = 0
|
||
|
||
for (const msg of messages) {
|
||
const msgLength = msg.senderName.length + msg.content.length + 3 // "name: content\n"
|
||
|
||
if (currentLength + msgLength > maxCharsPerSegment && currentSegment.length > 0) {
|
||
segments.push(currentSegment)
|
||
currentSegment = []
|
||
currentLength = 0
|
||
}
|
||
|
||
currentSegment.push(msg)
|
||
currentLength += msgLength
|
||
}
|
||
|
||
if (currentSegment.length > 0) {
|
||
segments.push(currentSegment)
|
||
}
|
||
|
||
return segments
|
||
}
|
||
|
||
/**
|
||
* 生成摘要的 Prompt
|
||
*/
|
||
function buildSummaryPrompt(content: string, lengthLimit: number, locale: string): string {
|
||
if (locale.startsWith('zh')) {
|
||
return `请用简洁的语言（${lengthLimit}字以内）总结以下对话的主要内容或话题。只输出摘要内容，不要添加任何前缀、解释或引号。
|
||
|
||
${content}`
|
||
}
|
||
return `Summarize the following conversation concisely (max ${lengthLimit} characters). Output only the summary, no prefix, explanation, or quotes.
|
||
|
||
${content}`
|
||
}
|
||
|
||
/**
|
||
* 生成子摘要的 Prompt
|
||
*/
|
||
function buildSubSummaryPrompt(content: string, locale: string): string {
|
||
if (locale.startsWith('zh')) {
|
||
return `请用一句话（不超过50字）概括以下对话片段的主要内容。只输出摘要内容，不要添加任何前缀、解释或引号。
|
||
|
||
${content}`
|
||
}
|
||
return `Summarize this conversation segment in one sentence (max 50 characters). Output only the summary, no prefix or quotes.
|
||
|
||
${content}`
|
||
}
|
||
|
||
/**
|
||
* 合并子摘要的 Prompt
|
||
*/
|
||
function buildMergePrompt(subSummaries: string[], lengthLimit: number, locale: string): string {
|
||
const summaryList = subSummaries.map((s, i) => `${i + 1}. ${s}`).join('\n')
|
||
if (locale.startsWith('zh')) {
|
||
return `以下是一段对话的多个片段摘要，请将它们合并成一个完整的总结（${lengthLimit}字以内）。只输出摘要内容，不要添加任何前缀、解释或引号。
|
||
|
||
${summaryList}`
|
||
}
|
||
return `Below are summaries of different parts of a conversation. Merge them into one cohesive summary (max ${lengthLimit} characters). Output only the summary, no prefix or quotes.
|
||
|
||
${summaryList}`
|
||
}
|
||
|
||
/**
|
||
* 生成会话摘要
|
||
*
|
||
* @param dbSessionId 数据库会话ID（用于访问数据库）
|
||
* @param chatSessionId 会话索引中的会话ID
|
||
* @param locale 语言设置
|
||
* @param forceRegenerate 是否强制重新生成（忽略缓存）
|
||
* @returns 摘要内容或错误
|
||
*/
|
||
export async function generateSessionSummary(
|
||
dbSessionId: string,
|
||
chatSessionId: number,
|
||
locale: string = 'zh-CN',
|
||
forceRegenerate: boolean = false
|
||
): Promise<{ success: boolean; summary?: string; error?: string }> {
|
||
try {
|
||
// 1. 检查是否已有摘要（除非强制重新生成）
|
||
if (!forceRegenerate) {
|
||
const existing = getSessionSummaryFromDb(dbSessionId, chatSessionId)
|
||
if (existing) {
|
||
return { success: true, summary: existing }
|
||
}
|
||
}
|
||
|
||
// 2. 获取会话消息
|
||
const sessionData = getSessionMessagesForSummary(dbSessionId, chatSessionId)
|
||
if (!sessionData) {
|
||
return { success: false, error: t('summary.sessionNotFound') }
|
||
}
|
||
|
||
// 3. 检查消息数量
|
||
if (sessionData.messageCount < MIN_MESSAGE_COUNT) {
|
||
return {
|
||
success: false,
|
||
error: t('summary.tooFewMessages', { count: MIN_MESSAGE_COUNT }),
|
||
}
|
||
}
|
||
|
||
// 4. 预处理：过滤无意义消息
|
||
const validMessages = preprocessMessages(sessionData.messages)
|
||
if (validMessages.length < MIN_MESSAGE_COUNT) {
|
||
return {
|
||
success: false,
|
||
error: t('summary.tooFewValidMessages', { count: MIN_MESSAGE_COUNT }),
|
||
}
|
||
}
|
||
|
||
// 5. 计算摘要长度限制
|
||
const lengthLimit = getSummaryLengthLimit(validMessages.length)
|
||
|
||
// 6. 格式化内容
|
||
const content = formatMessages(validMessages)
|
||
|
||
aiLogger.info(
|
||
'Summary',
|
||
`生成会话摘要: sessionId=${chatSessionId}, 原始消息=${sessionData.messageCount}, 有效消息=${validMessages.length}, 内容长度=${content.length}`
|
||
)
|
||
|
||
let summary: string
|
||
|
||
// 7. 根据内容长度决定处理策略
|
||
if (content.length <= SEGMENT_THRESHOLD) {
|
||
// 短会话：直接生成摘要
|
||
summary = await generateDirectSummary(content, lengthLimit, locale)
|
||
} else {
|
||
// 长会话：Map-Reduce 策略
|
||
summary = await generateMapReduceSummary(validMessages, lengthLimit, locale)
|
||
}
|
||
|
||
// 8. 后处理：移除引号
|
||
if ((summary.startsWith('"') && summary.endsWith('"')) || (summary.startsWith('「') && summary.endsWith('」'))) {
|
||
summary = summary.slice(1, -1)
|
||
}
|
||
|
||
// 如果摘要超过限制的 1.5 倍，进行截断
|
||
const hardLimit = Math.floor(lengthLimit * 1.5)
|
||
if (summary.length > hardLimit) {
|
||
summary = summary.slice(0, hardLimit - 3) + '...'
|
||
}
|
||
|
||
// 9. 保存到数据库
|
||
saveSessionSummaryToDb(dbSessionId, chatSessionId, summary)
|
||
|
||
aiLogger.info('Summary', `Summary generated: "${summary.slice(0, 50)}..."`)
|
||
|
||
return { success: true, summary }
|
||
} catch (error) {
|
||
aiLogger.error('Summary', 'Summary generation failed', error)
|
||
return {
|
||
success: false,
|
||
error: error instanceof Error ? error.message : String(error),
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 直接生成摘要（适用于短会话）
|
||
*/
|
||
async function generateDirectSummary(content: string, lengthLimit: number, locale: string): Promise<string> {
|
||
const result = await llmComplete(t('summary.systemPromptDirect'), buildSummaryPrompt(content, lengthLimit, locale), {
|
||
temperature: 0.3,
|
||
maxTokens: 300,
|
||
})
|
||
return result.trim()
|
||
}
|
||
|
||
/**
|
||
* Map-Reduce 策略生成摘要（适用于长会话）
|
||
*/
|
||
async function generateMapReduceSummary(
|
||
messages: Array<{ senderName: string; content: string }>,
|
||
lengthLimit: number,
|
||
locale: string
|
||
): Promise<string> {
|
||
// 1. Map：分段生成子摘要
|
||
const segments = splitIntoSegments(messages, MAX_CONTENT_PER_CALL)
|
||
aiLogger.info('Summary', `Long session segmented: ${segments.length} segments`)
|
||
|
||
const subSummaries: string[] = []
|
||
|
||
for (let i = 0; i < segments.length; i++) {
|
||
const segmentContent = formatMessages(segments[i])
|
||
const result = await llmComplete(t('summary.systemPromptDirect'), buildSubSummaryPrompt(segmentContent, locale), {
|
||
temperature: 0.3,
|
||
maxTokens: 100,
|
||
})
|
||
subSummaries.push(result.trim())
|
||
}
|
||
|
||
// 2. Reduce：合并子摘要
|
||
if (subSummaries.length === 1) {
|
||
return subSummaries[0]
|
||
}
|
||
|
||
const mergeResult = await llmComplete(
|
||
t('summary.systemPromptMerge'),
|
||
buildMergePrompt(subSummaries, lengthLimit, locale),
|
||
{ temperature: 0.3, maxTokens: 300 }
|
||
)
|
||
|
||
return mergeResult.trim()
|
||
}
|
||
|
||
/**
|
||
* 批量生成会话摘要
|
||
*
|
||
* @param dbSessionId 数据库会话ID
|
||
* @param chatSessionIds 会话ID列表
|
||
* @param locale 语言设置
|
||
* @param onProgress 进度回调
|
||
* @returns 生成结果
|
||
*/
|
||
export async function generateSessionSummaries(
|
||
dbSessionId: string,
|
||
chatSessionIds: number[],
|
||
locale: string = 'zh-CN',
|
||
onProgress?: (current: number, total: number) => void
|
||
): Promise<{ success: number; failed: number; skipped: number }> {
|
||
let success = 0
|
||
let failed = 0
|
||
let skipped = 0
|
||
|
||
for (let i = 0; i < chatSessionIds.length; i++) {
|
||
const chatSessionId = chatSessionIds[i]
|
||
|
||
const result = await generateSessionSummary(dbSessionId, chatSessionId, locale, false)
|
||
|
||
if (result.success) {
|
||
success++
|
||
} else if (result.error?.includes('少于') || result.error?.includes('less than') || result.error?.includes('few')) {
|
||
skipped++
|
||
} else {
|
||
failed++
|
||
}
|
||
|
||
if (onProgress) {
|
||
onProgress(i + 1, chatSessionIds.length)
|
||
}
|
||
}
|
||
|
||
return { success, failed, skipped }
|
||
}
|
||
|
||
/**
|
||
* 批量检查会话是否可以生成摘要
|
||
*
|
||
* @param dbSessionId 数据库会话ID
|
||
* @param chatSessionIds 会话ID列表
|
||
* @returns 每个会话的检查结果
|
||
*/
|
||
export function checkSessionsCanGenerateSummary(
|
||
dbSessionId: string,
|
||
chatSessionIds: number[]
|
||
): Map<number, { canGenerate: boolean; reason?: string }> {
|
||
const results = new Map<number, { canGenerate: boolean; reason?: string }>()
|
||
|
||
for (const chatSessionId of chatSessionIds) {
|
||
// 获取会话消息
|
||
const sessionData = getSessionMessagesForSummary(dbSessionId, chatSessionId)
|
||
|
||
if (!sessionData) {
|
||
results.set(chatSessionId, { canGenerate: false, reason: t('summary.sessionNotExist') })
|
||
continue
|
||
}
|
||
|
||
// 检查原始消息数量
|
||
if (sessionData.messageCount < MIN_MESSAGE_COUNT) {
|
||
results.set(chatSessionId, { canGenerate: false, reason: t('summary.messagesTooFew') })
|
||
continue
|
||
}
|
||
|
||
// 预处理：过滤无意义消息
|
||
const validMessages = preprocessMessages(sessionData.messages)
|
||
if (validMessages.length < MIN_MESSAGE_COUNT) {
|
||
results.set(chatSessionId, { canGenerate: false, reason: t('summary.validMessagesTooFew') })
|
||
continue
|
||
}
|
||
|
||
results.set(chatSessionId, { canGenerate: true })
|
||
}
|
||
|
||
return results
|
||
}
|