Update metaSearchAgent.ts
This commit is contained in:
Lucas
GitHub
parent
3e53bab9b4
commit
19148eeba7
1 changed files with 366 additions and 433 deletions
|
|
@ -11,7 +11,11 @@ import {
|
|||
RunnableMap,
|
||||
RunnableSequence,
|
||||
} from '@langchain/core/runnables';
|
||||
import { BaseMessage } from '@langchain/core/messages';
|
||||
import {
|
||||
BaseMessage,
|
||||
AIMessage,
|
||||
HumanMessage,
|
||||
} from '@langchain/core/messages';
|
||||
import { StringOutputParser } from '@langchain/core/output_parsers';
|
||||
import LineListOutputParser from '../lib/outputParsers/listLineOutputParser';
|
||||
import LineOutputParser from '../lib/outputParsers/lineOutputParser';
|
||||
|
|
@ -33,10 +37,9 @@ import { SearxngSearchOptions } from '../lib/searxng';
|
|||
import { ChromaClient } from 'chromadb';
|
||||
import { OpenAIEmbeddings } from '@langchain/openai';
|
||||
import { EventEmitter } from 'events';
|
||||
import { MemoryVectorStore } from "langchain/vectorstores/memory";
|
||||
import { webSearchRetrieverPrompt, webSearchResponsePrompt } from '../prompts/webSearch';
|
||||
|
||||
export interface MetaSearchAgentType {
|
||||
initialize: (embeddings: Embeddings) => Promise<void>;
|
||||
searchAndAnswer: (
|
||||
message: string,
|
||||
history: BaseMessage[],
|
||||
|
|
@ -49,8 +52,8 @@ export interface MetaSearchAgentType {
|
|||
|
||||
interface Config {
|
||||
activeEngines: string[];
|
||||
queryGeneratorPrompt: string;
|
||||
responsePrompt: string;
|
||||
queryGeneratorPrompt?: string;
|
||||
responsePrompt?: string;
|
||||
rerank: boolean;
|
||||
rerankThreshold: number;
|
||||
searchWeb: boolean;
|
||||
|
|
@ -99,52 +102,26 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
private config: Config;
|
||||
private strParser = new StringOutputParser();
|
||||
private fileIds: string[];
|
||||
private memoryStore: MemoryVectorStore;
|
||||
private conversationHistory: BaseMessage[] = [];
|
||||
|
||||
constructor(config: Config) {
|
||||
this.config = config;
|
||||
this.fileIds = [];
|
||||
}
|
||||
|
||||
async initialize(embeddings: Embeddings) {
|
||||
try {
|
||||
if (!this.memoryStore) {
|
||||
console.log("🔄 Initialisation du memory store...");
|
||||
this.memoryStore = new MemoryVectorStore(embeddings);
|
||||
await this.memoryStore.addDocuments([{
|
||||
pageContent: "Initialisation du memory store",
|
||||
metadata: { timestamp: Date.now() }
|
||||
}]);
|
||||
console.log("✅ Memory store initialisé avec succès");
|
||||
}
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de l'initialisation du memory store:", error);
|
||||
throw error;
|
||||
}
|
||||
private updateMemory(message: BaseMessage) {
|
||||
this.conversationHistory.push(message);
|
||||
}
|
||||
|
||||
private async enrichWithMemory(message: string, embeddings: Embeddings): Promise<string> {
|
||||
try {
|
||||
if (!this.memoryStore) {
|
||||
await this.initialize(embeddings);
|
||||
}
|
||||
await this.memoryStore.addDocuments([{
|
||||
pageContent: message,
|
||||
metadata: { timestamp: Date.now() }
|
||||
}]);
|
||||
const results = await this.memoryStore.similaritySearch(message, 2);
|
||||
return results.map(doc => doc.pageContent).join('\n');
|
||||
} catch (error) {
|
||||
console.error("Erreur mémoire:", error);
|
||||
return '';
|
||||
}
|
||||
public getMemory(): BaseMessage[] {
|
||||
return this.conversationHistory;
|
||||
}
|
||||
|
||||
private async createSearchRetrieverChain(llm: BaseChatModel) {
|
||||
(llm as unknown as ChatOpenAI).temperature = 0;
|
||||
|
||||
return RunnableSequence.from([
|
||||
PromptTemplate.fromTemplate(this.config.queryGeneratorPrompt),
|
||||
PromptTemplate.fromTemplate(webSearchRetrieverPrompt),
|
||||
llm,
|
||||
this.strParser,
|
||||
RunnableLambda.from(async (input: string) => {
|
||||
|
|
@ -191,7 +168,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
// Recherche d'experts si activée
|
||||
if (this.config.searchDatabase) {
|
||||
try {
|
||||
console.log("🔍 Recherche d'experts...");
|
||||
console.log('🔍 Recherche d\'experts...');
|
||||
const expertResults = await handleExpertSearch(
|
||||
{
|
||||
query: question,
|
||||
|
|
@ -202,7 +179,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
llm
|
||||
);
|
||||
|
||||
console.log("🔍 Experts trouvés:", expertResults.experts.length);
|
||||
console.log('🔍 Experts trouvés:', expertResults.experts.length);
|
||||
const expertDocs = expertResults.experts.map(expert =>
|
||||
new Document({
|
||||
pageContent: `Expert: ${expert.prenom} ${expert.nom}
|
||||
|
|
@ -225,7 +202,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
|
||||
documents = [...expertDocs, ...documents];
|
||||
} catch (error) {
|
||||
console.error("Erreur lors de la recherche d'experts:", error);
|
||||
console.error('Erreur lors de la recherche d\'experts:', error);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -242,7 +219,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
|
||||
private async loadUploadedDocuments(fileIds: string[]): Promise<Document[]> {
|
||||
console.log("📂 Chargement des documents:", fileIds);
|
||||
console.log('📂 Chargement des documents:', fileIds);
|
||||
const docs: Document[] = [];
|
||||
|
||||
for (const fileId of fileIds) {
|
||||
|
|
@ -310,56 +287,46 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
query: (input: BasicChainInput) => input.query,
|
||||
chat_history: (input: BasicChainInput) => input.chat_history,
|
||||
docs: RunnableLambda.from(async (input: BasicChainInput) => {
|
||||
console.log("Début de la recherche avec contexte enrichi...");
|
||||
console.log('Début de la recherche...');
|
||||
let docs: Document[] = [];
|
||||
|
||||
// Récupérer le contexte historique
|
||||
const memoryContext = await this.getRelevantContext(input.query);
|
||||
console.log("💭 Contexte historique pour la recherche:", memoryContext);
|
||||
|
||||
// Enrichir la requête avec le contexte
|
||||
const enrichedQuery = `
|
||||
Contexte précédent:
|
||||
${memoryContext}
|
||||
|
||||
Question actuelle:
|
||||
${input.query}
|
||||
`;
|
||||
|
||||
// 1. D'abord chercher dans les documents uploadés avec le contexte enrichi
|
||||
// 1. D'abord chercher dans les documents uploadés
|
||||
if (fileIds.length > 0) {
|
||||
try {
|
||||
const uploadedDocs = await this.loadUploadedDocuments(fileIds);
|
||||
console.log("📚 Documents uploadés chargés:", uploadedDocs.length);
|
||||
console.log('📚 Documents uploadés chargés:', uploadedDocs.length);
|
||||
|
||||
// Utiliser RAGDocumentChain pour la recherche dans les documents
|
||||
const ragChain = new RAGDocumentChain();
|
||||
await ragChain.initializeVectorStoreFromDocuments(uploadedDocs, embeddings);
|
||||
|
||||
// Utiliser le type 'specific' pour une recherche précise
|
||||
const searchChain = ragChain.createSearchChain(llm);
|
||||
const relevantDocs = await searchChain.invoke({
|
||||
query: enrichedQuery,
|
||||
query: input.query,
|
||||
chat_history: input.chat_history,
|
||||
type: 'specific'
|
||||
});
|
||||
|
||||
// Ajouter les documents pertinents avec un score élevé
|
||||
docs = uploadedDocs.map(doc => ({
|
||||
...doc,
|
||||
metadata: {
|
||||
...doc.metadata,
|
||||
score: 0.8
|
||||
score: 0.8 // Score élevé pour les documents uploadés
|
||||
}
|
||||
}));
|
||||
|
||||
console.log("📄 Documents pertinents trouvés:", docs.length);
|
||||
console.log('📄 Documents pertinents trouvés:', docs.length);
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la recherche dans les documents:", error);
|
||||
console.error('❌ Erreur lors de la recherche dans les documents:', error);
|
||||
}
|
||||
}
|
||||
|
||||
// 2. Ensuite chercher les experts si pertinent
|
||||
if (this.config.searchDatabase) {
|
||||
try {
|
||||
console.log("👥 Recherche d'experts...");
|
||||
console.log('👥 Recherche d\'experts...');
|
||||
const expertResults = await handleExpertSearch(
|
||||
{
|
||||
query: input.query,
|
||||
|
|
@ -375,7 +342,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
docs = [...docs, ...expertDocs];
|
||||
}
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la recherche d'experts:", error);
|
||||
console.error('❌ Erreur lors de la recherche d\'experts:', error);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -385,13 +352,13 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
const webResults = await this.performWebSearch(input.query);
|
||||
docs = [...docs, ...webResults];
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la recherche web:", error);
|
||||
console.error('❌ Erreur lors de la recherche web:', error);
|
||||
}
|
||||
}
|
||||
|
||||
console.log("🔍 DEBUG - Avant appel rerankDocs - Mode:", optimizationMode, "Query:", input.query);
|
||||
console.log('🔍 DEBUG - Avant appel rerankDocs - Mode:', optimizationMode, 'Query:', input.query);
|
||||
return this.rerankDocs(
|
||||
enrichedQuery,
|
||||
input.query,
|
||||
docs,
|
||||
fileIds,
|
||||
embeddings,
|
||||
|
|
@ -406,14 +373,14 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
chat_history: (input) => input.chat_history,
|
||||
date: () => new Date().toISOString(),
|
||||
context: (input) => {
|
||||
console.log("Préparation du contexte...");
|
||||
console.log('Préparation du contexte...');
|
||||
return this.processDocs(input.docs);
|
||||
},
|
||||
docs: (input) => input.docs,
|
||||
}),
|
||||
|
||||
ChatPromptTemplate.fromMessages([
|
||||
['system', this.config.responsePrompt],
|
||||
['system', webSearchResponsePrompt],
|
||||
new MessagesPlaceholder('chat_history'),
|
||||
['user', '{context}\n\n{query}'],
|
||||
]),
|
||||
|
|
@ -465,8 +432,8 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
|
||||
private processDocs(docs: Document[]) {
|
||||
// Trier les documents par score si disponible
|
||||
const sortedDocs = docs.sort((a, b) =>
|
||||
(b.metadata?.score || 0) - (a.metadata?.score || 0)
|
||||
const sortedDocs = docs.sort(
|
||||
(a, b) => (b.metadata?.score || 0) - (a.metadata?.score || 0)
|
||||
);
|
||||
|
||||
// Limiter à 5 documents maximum
|
||||
|
|
@ -475,8 +442,9 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
// Limiter la taille de chaque document à 1000 caractères
|
||||
return limitedDocs
|
||||
.map((doc, index) => {
|
||||
const content = doc.pageContent.length > 1000
|
||||
? doc.pageContent.substring(0, 1000) + "..."
|
||||
const content =
|
||||
doc.pageContent.length > 1000
|
||||
? doc.pageContent.substring(0, 1000) + '...'
|
||||
: doc.pageContent;
|
||||
|
||||
return `${content} [${index + 1}]`;
|
||||
|
|
@ -486,7 +454,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
|
||||
private async handleStream(
|
||||
stream: IterableReadableStream<StreamEvent>,
|
||||
emitter: eventEmitter,
|
||||
emitter: eventEmitter
|
||||
) {
|
||||
for await (const event of stream) {
|
||||
if (
|
||||
|
|
@ -496,7 +464,8 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
const sources = event.data.output;
|
||||
|
||||
// Normaliser les sources pour le frontend
|
||||
const normalizedSources = sources?.map(source => {
|
||||
const normalizedSources =
|
||||
sources?.map(source => {
|
||||
const isUploadedDoc = source.metadata?.type === 'uploaded';
|
||||
const isExpert = source.metadata?.type === 'expert';
|
||||
const pageNumber = source.metadata?.pageNumber || 1;
|
||||
|
|
@ -531,7 +500,9 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
url: url,
|
||||
source: sourceId,
|
||||
pageNumber: pageNumber,
|
||||
searchText: source.metadata?.searchText?.substring(0, 200) || limitedContent.substring(0, 200),
|
||||
searchText:
|
||||
source.metadata?.searchText?.substring(0, 200) ||
|
||||
limitedContent.substring(0, 200),
|
||||
expertData: source.metadata?.expertData,
|
||||
illustrationImage: source.metadata?.illustrationImage,
|
||||
imageTitle: source.metadata?.imageTitle,
|
||||
|
|
@ -542,7 +513,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
};
|
||||
}) || [];
|
||||
|
||||
console.log("🔍 Sources normalisées:", normalizedSources.length);
|
||||
console.log('🔍 Sources normalisées:', normalizedSources.length);
|
||||
|
||||
emitter.emit(
|
||||
'data',
|
||||
|
|
@ -572,13 +543,100 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
}
|
||||
|
||||
private async handleStreamWithMemory(
|
||||
stream: IterableReadableStream<StreamEvent>,
|
||||
emitter: eventEmitter
|
||||
) {
|
||||
let fullAssistantResponse = '';
|
||||
|
||||
for await (const event of stream) {
|
||||
if (event.event === 'on_chain_stream') {
|
||||
if (event.name === 'FinalResponseGenerator') {
|
||||
fullAssistantResponse += event.data.chunk;
|
||||
emitter.emit(
|
||||
'data',
|
||||
JSON.stringify({ type: 'response', data: event.data.chunk })
|
||||
);
|
||||
}
|
||||
} else if (event.event === 'on_chain_end') {
|
||||
if (event.name === 'FinalResponseGenerator') {
|
||||
this.updateMemory(new AIMessage(fullAssistantResponse.trim()));
|
||||
emitter.emit('end');
|
||||
}
|
||||
if (event.name === 'FinalSourceRetriever') {
|
||||
const sources = event.data.output;
|
||||
const normalizedSources =
|
||||
sources?.map(source => {
|
||||
const isUploadedDoc = source.metadata?.type === 'uploaded';
|
||||
const isExpert = source.metadata?.type === 'expert';
|
||||
const pageNumber = source.metadata?.pageNumber || 1;
|
||||
const sourceId = source.metadata?.source;
|
||||
|
||||
let url;
|
||||
if (isUploadedDoc && sourceId) {
|
||||
url = `/api/uploads/${sourceId}/content?page=${pageNumber}`;
|
||||
} else if (isExpert) {
|
||||
url = source.metadata?.url;
|
||||
} else if (source.metadata?.type === 'web') {
|
||||
url = source.metadata?.url;
|
||||
}
|
||||
|
||||
let title = source.metadata?.title || '';
|
||||
if (isUploadedDoc && title) {
|
||||
title = `${title} - Page ${pageNumber}`;
|
||||
} else if (isExpert) {
|
||||
title = source.metadata?.displayTitle || title;
|
||||
}
|
||||
|
||||
const limitedContent =
|
||||
source.pageContent?.substring(0, 1000) || '';
|
||||
|
||||
return {
|
||||
pageContent: limitedContent,
|
||||
metadata: {
|
||||
title: title,
|
||||
type: source.metadata?.type || 'web',
|
||||
url: url,
|
||||
source: sourceId,
|
||||
pageNumber: pageNumber,
|
||||
searchText:
|
||||
source.metadata?.searchText?.substring(0, 200) ||
|
||||
limitedContent.substring(0, 200),
|
||||
expertData: source.metadata?.expertData,
|
||||
illustrationImage: source.metadata?.illustrationImage,
|
||||
imageTitle: source.metadata?.imageTitle,
|
||||
favicon: source.metadata?.favicon,
|
||||
linkText: source.metadata?.linkText,
|
||||
expertName: source.metadata?.expertName
|
||||
}
|
||||
};
|
||||
}) || [];
|
||||
|
||||
console.log('🔍 Sources normalisées:', normalizedSources.length);
|
||||
|
||||
emitter.emit(
|
||||
'data',
|
||||
JSON.stringify({
|
||||
type: 'sources',
|
||||
data: normalizedSources,
|
||||
illustrationImage: normalizedSources[0]?.metadata?.illustrationImage || null,
|
||||
imageTitle: normalizedSources[0]?.metadata?.imageTitle || null
|
||||
})
|
||||
);
|
||||
}
|
||||
} else {
|
||||
emitter.emit(event.event, event.data);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private async searchExperts(
|
||||
query: string,
|
||||
embeddings: Embeddings,
|
||||
llm: BaseChatModel
|
||||
): Promise<SearchResult[]> {
|
||||
try {
|
||||
console.log("👥 Recherche d'experts pour:", query);
|
||||
console.log('👥 Recherche d\'experts pour:', query);
|
||||
const expertResults = await handleExpertSearch(
|
||||
{
|
||||
query,
|
||||
|
|
@ -608,14 +666,14 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
}));
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la recherche d'experts:", error);
|
||||
console.error('❌ Erreur lors de la recherche d\'experts:', error);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
private async searchWeb(query: string): Promise<SearchResult[]> {
|
||||
try {
|
||||
console.log("🌐 Recherche web pour:", query);
|
||||
console.log('🌐 Recherche web pour:', query);
|
||||
const res = await searchSearxng(query, {
|
||||
language: 'fr',
|
||||
engines: this.config.activeEngines,
|
||||
|
|
@ -632,7 +690,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
}));
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la recherche web:", error);
|
||||
console.error('❌ Erreur lors de la recherche web:', error);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
|
@ -645,13 +703,13 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
optimizationMode: 'speed' | 'balanced' | 'quality',
|
||||
llm: BaseChatModel
|
||||
) {
|
||||
console.log("🔍 Mode d'optimisation:", optimizationMode);
|
||||
console.log("🔍 Query pour la recherche d'image:", query);
|
||||
console.log('🔍 Mode d\'optimisation:', optimizationMode);
|
||||
console.log('🔍 Query pour la recherche d\'image:', query);
|
||||
|
||||
if (optimizationMode === 'balanced' || optimizationMode === 'quality') {
|
||||
console.log("🔍 Démarrage de la recherche d'images...");
|
||||
console.log('🔍 Démarrage de la recherche d\'images...');
|
||||
try {
|
||||
console.log("🔍 Appel de handleImageSearch avec la query:", query);
|
||||
console.log('🔍 Appel de handleImageSearch avec la query:', query);
|
||||
const images = await handleImageSearch(
|
||||
{
|
||||
query,
|
||||
|
|
@ -659,11 +717,11 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
},
|
||||
llm
|
||||
);
|
||||
console.log("🔍 Résultat brut de handleImageSearch:", JSON.stringify(images, null, 2));
|
||||
console.log("🔍 Images trouvées:", images?.length);
|
||||
console.log('🔍 Résultat brut de handleImageSearch:', JSON.stringify(images, null, 2));
|
||||
console.log('🔍 Images trouvées:', images?.length);
|
||||
|
||||
if (images && images.length > 0) {
|
||||
console.log("🔍 Première image trouvée:", {
|
||||
console.log('🔍 Première image trouvée:', {
|
||||
src: images[0].img_src,
|
||||
title: images[0].title,
|
||||
url: images[0].url
|
||||
|
|
@ -673,199 +731,52 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
metadata: {
|
||||
...doc.metadata,
|
||||
illustrationImage: images[0].img_src,
|
||||
title: images[0].title
|
||||
imageTitle: images[0].title
|
||||
}
|
||||
}));
|
||||
} else {
|
||||
console.log("⚠️ Aucune image trouvée dans le résultat");
|
||||
console.log('⚠️ Aucune image trouvée dans le résultat');
|
||||
}
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur détaillée lors de la recherche d'image:", {
|
||||
console.error('❌ Erreur détaillée lors de la recherche d\'image:', {
|
||||
message: error.message,
|
||||
stack: error.stack
|
||||
});
|
||||
}
|
||||
} else {
|
||||
console.log("🔍 Mode speed: pas de recherche d'images");
|
||||
console.log('🔍 Mode speed: pas de recherche d\'images');
|
||||
}
|
||||
|
||||
return docs.slice(0, 15);
|
||||
}
|
||||
|
||||
private async updateMemoryStore(message: string, history: BaseMessage[]) {
|
||||
try {
|
||||
if (!this.memoryStore) {
|
||||
throw new Error("Memory store not initialized");
|
||||
}
|
||||
|
||||
// Créer un document avec le contexte actuel et sa structure
|
||||
const contextDoc = {
|
||||
pageContent: `Question: ${message}\nContext: ${history.map(m =>
|
||||
`${m._getType()}: ${m.content}`).join('\n')}`,
|
||||
metadata: {
|
||||
timestamp: Date.now(),
|
||||
themes: this.extractThemes(message + ' ' + history.map(m => m.content).join(' ')),
|
||||
type: 'conversation'
|
||||
}
|
||||
};
|
||||
|
||||
console.log("💾 Ajout à la mémoire:", contextDoc);
|
||||
await this.memoryStore.addDocuments([contextDoc]);
|
||||
console.log("✅ Mémoire mise à jour avec succès");
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la mise à jour de la mémoire:", error);
|
||||
}
|
||||
}
|
||||
|
||||
private async getRelevantContext(message: string): Promise<string> {
|
||||
try {
|
||||
if (!this.memoryStore) {
|
||||
throw new Error("Memory store not initialized");
|
||||
}
|
||||
console.log("🔍 Recherche dans la mémoire pour:", message);
|
||||
|
||||
const results = await this.memoryStore.similaritySearch(message, 3);
|
||||
console.log("📚 Contexte trouvé dans la mémoire:", results);
|
||||
return results.map(doc => doc.pageContent).join('\n');
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de la récupération du contexte:", error);
|
||||
return '';
|
||||
}
|
||||
}
|
||||
|
||||
private extractThemes(context: string): string[] {
|
||||
const commonThemes = ['sas', 'entreprise', 'création', 'chomage', 'juridique', 'financement'];
|
||||
return commonThemes.filter(theme =>
|
||||
context.toLowerCase().includes(theme.toLowerCase())
|
||||
);
|
||||
}
|
||||
|
||||
private async analyzeConversationContext(
|
||||
message: string,
|
||||
history: BaseMessage[],
|
||||
llm: BaseChatModel
|
||||
): Promise<{ context: string; relevance: number }> {
|
||||
try {
|
||||
const formattedHistory = formatChatHistoryAsString(history);
|
||||
|
||||
const analysis = await llm.invoke(`
|
||||
Analysez cette conversation en profondeur en donnant une importance particulière aux 3 derniers échanges.
|
||||
Assurez-vous de maintenir la cohérence du contexte entre les questions.
|
||||
|
||||
Historique de la conversation:
|
||||
${formattedHistory}
|
||||
|
||||
Nouvelle question: "${message}"
|
||||
|
||||
Instructions spécifiques:
|
||||
1. Identifiez les thèmes récurrents des 3 derniers échanges
|
||||
2. Notez les contraintes ou conditions mentionnées précédemment qui restent pertinentes
|
||||
3. Évaluez comment la nouvelle question s'inscrit dans la continuité des échanges
|
||||
|
||||
Répondez au format JSON:
|
||||
{
|
||||
"mainTopic": "sujet principal actuel",
|
||||
"recentContext": {
|
||||
"lastThemes": ["thèmes des 3 derniers échanges"],
|
||||
"activeConstraints": ["contraintes toujours actives"],
|
||||
"continuityScore": <0.0 à 1.0>
|
||||
},
|
||||
"contextualFactors": {
|
||||
"financial": ["facteurs financiers"],
|
||||
"legal": ["aspects juridiques"],
|
||||
"administrative": ["aspects administratifs"]
|
||||
},
|
||||
"impactAnalysis": {
|
||||
"primary": "impact principal",
|
||||
"secondary": ["impacts secondaires"],
|
||||
"constraints": ["contraintes identifiées"]
|
||||
},
|
||||
"relevanceScore": <0.0 à 1.0>
|
||||
}
|
||||
`);
|
||||
|
||||
const result = JSON.parse(String(analysis.content));
|
||||
|
||||
// Construction d'un contexte enrichi qui met l'accent sur la continuité
|
||||
const enrichedContext = `
|
||||
Contexte principal: ${result.mainTopic}
|
||||
Thèmes récents: ${result.recentContext.lastThemes.join(', ')}
|
||||
Contraintes actives: ${result.recentContext.activeConstraints.join(', ')}
|
||||
Facteurs impactants: ${result.contextualFactors.financial.join(', ')}
|
||||
Implications légales: ${result.contextualFactors.legal.join(', ')}
|
||||
Impact global: ${result.impactAnalysis.primary}
|
||||
Contraintes à considérer: ${result.impactAnalysis.constraints.join(', ')}
|
||||
`.trim();
|
||||
|
||||
// Ajuster le score de pertinence en fonction de la continuité
|
||||
const adjustedRelevance = (result.relevanceScore + result.recentContext.continuityScore) / 2;
|
||||
|
||||
return {
|
||||
context: enrichedContext,
|
||||
relevance: adjustedRelevance
|
||||
};
|
||||
} catch (error) {
|
||||
console.error("Erreur lors de l'analyse du contexte:", error);
|
||||
return { context: '', relevance: 0 };
|
||||
}
|
||||
}
|
||||
|
||||
async searchAndAnswer(
|
||||
message: string,
|
||||
history: BaseMessage[],
|
||||
llm: BaseChatModel,
|
||||
embeddings: Embeddings,
|
||||
optimizationMode: 'speed' | 'balanced' | 'quality',
|
||||
fileIds: string[],
|
||||
fileIds: string[]
|
||||
) {
|
||||
const effectiveMode = 'balanced';
|
||||
const emitter = new EventEmitter();
|
||||
const emitter = new eventEmitter();
|
||||
|
||||
try {
|
||||
// S'assurer que le memoryStore est initialisé
|
||||
if (!this.memoryStore) {
|
||||
console.log("🔄 Initialisation du memory store dans searchAndAnswer...");
|
||||
try {
|
||||
await this.initialize(embeddings);
|
||||
if (!this.memoryStore) {
|
||||
throw new Error("Memory store initialization failed");
|
||||
}
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur lors de l'initialisation du memory store:", error);
|
||||
throw error;
|
||||
}
|
||||
}
|
||||
// Ajouter le message utilisateur à la mémoire
|
||||
this.updateMemory(new HumanMessage(message));
|
||||
|
||||
// Analyser le contexte de la conversation
|
||||
const conversationContext = await this.analyzeConversationContext(message, history, llm);
|
||||
console.log("🧠 Analyse du contexte:", conversationContext);
|
||||
|
||||
// Mettre à jour la mémoire avec le contexte actuel
|
||||
await this.updateMemoryStore(message, history);
|
||||
|
||||
// Récupérer le contexte pertinent de la mémoire
|
||||
const memoryContext = await this.getRelevantContext(message);
|
||||
console.log("💭 Contexte mémoire récupéré:", memoryContext);
|
||||
|
||||
// Enrichir le message avec le contexte historique
|
||||
const enrichedMessage = `
|
||||
Contexte précédent:
|
||||
${memoryContext}
|
||||
|
||||
Contexte actuel:
|
||||
${conversationContext.context}
|
||||
|
||||
Question actuelle:
|
||||
${message}
|
||||
`;
|
||||
// Fusionner l'historique
|
||||
const mergedHistory: BaseMessage[] = [
|
||||
...this.conversationHistory,
|
||||
...history,
|
||||
];
|
||||
|
||||
// Analyse sophistiquée de la requête avec LLM
|
||||
const queryAnalysis = await llm.invoke(`
|
||||
En tant qu'expert en analyse de requêtes, examine cette demande dans son contexte complet.
|
||||
const queryAnalysis = await llm.invoke(`En tant qu'expert en analyse de requêtes, examine cette demande et détermine la stratégie de recherche optimale.
|
||||
|
||||
${enrichedMessage}
|
||||
Question/Requête: "${message}"
|
||||
|
||||
Documents disponibles: ${fileIds.length > 0 ? "Oui" : "Non"}
|
||||
Documents disponibles: ${fileIds.length > 0 ? 'Oui' : 'Non'}
|
||||
|
||||
Analyse et réponds au format JSON:
|
||||
{
|
||||
|
|
@ -874,37 +785,48 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
"requiresWebSearch": <boolean>,
|
||||
"requiresExpertSearch": <boolean>,
|
||||
"documentRelevance": <0.0 à 1.0>,
|
||||
"contextRelevance": ${conversationContext.relevance},
|
||||
"reasoning": "<courte explication>"
|
||||
}`);
|
||||
}
|
||||
|
||||
Critères d'analyse:
|
||||
- DOCUMENT_QUERY: La question porte spécifiquement sur le contenu des documents
|
||||
- WEB_SEARCH: Recherche d'informations générales ou actuelles
|
||||
- EXPERT_ADVICE: Demande nécessitant une expertise spécifique
|
||||
- HYBRID: Combinaison de plusieurs sources
|
||||
|
||||
Prends en compte:
|
||||
- La présence ou non de documents uploadés
|
||||
- La spécificité de la question
|
||||
- Le besoin d'expertise externe
|
||||
- L'actualité du sujet`);
|
||||
|
||||
const analysis = JSON.parse(String(queryAnalysis.content));
|
||||
console.log("🎯 Analyse de la requête enrichie:", analysis);
|
||||
console.log('🎯 Analyse de la requête:', analysis);
|
||||
|
||||
// 1. Analyse des documents uploadés avec RAG
|
||||
const uploadedDocs = await this.loadUploadedDocuments(fileIds);
|
||||
console.log("📚 Documents uploadés chargés:", uploadedDocs.length);
|
||||
console.log('📚 Documents uploadés chargés:', uploadedDocs.length);
|
||||
|
||||
if (uploadedDocs.length > 0) {
|
||||
// Création du vectorStore temporaire pour les documents
|
||||
const vectorStore = await Chroma.fromDocuments(uploadedDocs, embeddings, {
|
||||
collectionName: "temp_docs",
|
||||
url: "http://chroma:8000",
|
||||
collectionName: 'temp_docs',
|
||||
url: 'http://chroma:8000',
|
||||
numDimensions: 1536
|
||||
});
|
||||
|
||||
// Recherche sémantique sans filtre pour l'instant
|
||||
const relevantDocs = await vectorStore.similaritySearch(message, 5);
|
||||
|
||||
console.log("📄 Documents pertinents trouvés:", relevantDocs.length);
|
||||
console.log('📄 Documents pertinents trouvés:', relevantDocs.length);
|
||||
|
||||
// Extraction du contexte pour enrichir la recherche
|
||||
const documentContext = relevantDocs
|
||||
.map(doc => doc.pageContent)
|
||||
.join("\n")
|
||||
.join('\n')
|
||||
.substring(0, 500);
|
||||
|
||||
const documentTitle = uploadedDocs[0]?.metadata?.title || "";
|
||||
const documentTitle = uploadedDocs[0]?.metadata?.title || '';
|
||||
const enrichedQuery = `${message} ${documentTitle} ${documentContext}`;
|
||||
|
||||
// 2. Recherche d'experts en BDD
|
||||
|
|
@ -958,7 +880,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
|
||||
const stream = answeringChain.streamEvents(
|
||||
{
|
||||
chat_history: history,
|
||||
chat_history: mergedHistory,
|
||||
query: `${message}\n\nContexte pertinent:\n${finalResults.map(doc => doc.pageContent).join('\n\n')}`
|
||||
},
|
||||
{
|
||||
|
|
@ -966,7 +888,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
);
|
||||
|
||||
this.handleStream(stream, emitter);
|
||||
this.handleStreamWithMemory(stream, emitter);
|
||||
} else {
|
||||
// Fallback sans documents uploadés
|
||||
const answeringChain = await this.createAnsweringChain(
|
||||
|
|
@ -978,7 +900,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
|
||||
const stream = answeringChain.streamEvents(
|
||||
{
|
||||
chat_history: history,
|
||||
chat_history: mergedHistory,
|
||||
query: message
|
||||
},
|
||||
{
|
||||
|
|
@ -986,10 +908,10 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
);
|
||||
|
||||
this.handleStream(stream, emitter);
|
||||
this.handleStreamWithMemory(stream, emitter);
|
||||
}
|
||||
} catch (error) {
|
||||
console.error("❌ Erreur:", error);
|
||||
console.error('❌ Erreur:', error);
|
||||
// Fallback en mode standard
|
||||
const answeringChain = await this.createAnsweringChain(
|
||||
llm,
|
||||
|
|
@ -1000,7 +922,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
|
||||
const stream = answeringChain.streamEvents(
|
||||
{
|
||||
chat_history: history,
|
||||
chat_history: this.conversationHistory,
|
||||
query: message
|
||||
},
|
||||
{
|
||||
|
|
@ -1008,7 +930,7 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
}
|
||||
);
|
||||
|
||||
this.handleStream(stream, emitter);
|
||||
this.handleStreamWithMemory(stream, emitter);
|
||||
}
|
||||
|
||||
return emitter;
|
||||
|
|
@ -1018,25 +940,25 @@ export class MetaSearchAgent implements MetaSearchAgentType {
|
|||
export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
||||
// ... existing handlers ...
|
||||
legal: {
|
||||
initialize: async (embeddings: Embeddings) => {
|
||||
// Pas besoin d'initialisation spécifique pour le handler legal
|
||||
},
|
||||
searchAndAnswer: async (
|
||||
message,
|
||||
history,
|
||||
llm,
|
||||
embeddings,
|
||||
optimizationMode,
|
||||
fileIds,
|
||||
fileIds
|
||||
) => {
|
||||
const emitter = new eventEmitter();
|
||||
|
||||
try {
|
||||
const chain = new RAGDocumentChain();
|
||||
await chain.initializeVectorStoreFromDocuments(fileIds.map(fileId => new Document({
|
||||
await chain.initializeVectorStoreFromDocuments(
|
||||
fileIds.map(fileId => new Document({
|
||||
pageContent: '',
|
||||
metadata: { source: fileId }
|
||||
})), embeddings);
|
||||
})),
|
||||
embeddings
|
||||
);
|
||||
|
||||
const searchChain = chain.createSearchChain(llm);
|
||||
const results = await searchChain.invoke({
|
||||
|
|
@ -1056,7 +978,7 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
'data',
|
||||
JSON.stringify({
|
||||
type: 'response',
|
||||
data: response.text,
|
||||
data: response.text
|
||||
})
|
||||
);
|
||||
|
||||
|
|
@ -1066,25 +988,22 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
'error',
|
||||
JSON.stringify({
|
||||
type: 'error',
|
||||
data: error.message,
|
||||
data: error.message
|
||||
})
|
||||
);
|
||||
}
|
||||
|
||||
return emitter;
|
||||
},
|
||||
}
|
||||
},
|
||||
documents: {
|
||||
initialize: async (embeddings: Embeddings) => {
|
||||
// Pas besoin d'initialisation spécifique pour le handler documents
|
||||
},
|
||||
searchAndAnswer: async (
|
||||
message,
|
||||
history,
|
||||
llm,
|
||||
embeddings,
|
||||
optimizationMode,
|
||||
fileIds,
|
||||
fileIds
|
||||
) => {
|
||||
const emitter = new eventEmitter();
|
||||
const ragChain = new RAGDocumentChain();
|
||||
|
|
@ -1098,7 +1017,7 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
pageContent: content.contents.join('\n'),
|
||||
metadata: {
|
||||
title: content.title,
|
||||
source: fileId,
|
||||
source: fileId
|
||||
}
|
||||
});
|
||||
});
|
||||
|
|
@ -1121,38 +1040,44 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
}))
|
||||
};
|
||||
|
||||
emitter.emit('data', JSON.stringify({
|
||||
emitter.emit(
|
||||
'data',
|
||||
JSON.stringify({
|
||||
type: 'response',
|
||||
data: response.text
|
||||
}));
|
||||
})
|
||||
);
|
||||
|
||||
emitter.emit('data', JSON.stringify({
|
||||
emitter.emit(
|
||||
'data',
|
||||
JSON.stringify({
|
||||
type: 'sources',
|
||||
data: response.sources
|
||||
}));
|
||||
})
|
||||
);
|
||||
|
||||
emitter.emit('end');
|
||||
} catch (error) {
|
||||
emitter.emit('error', JSON.stringify({
|
||||
emitter.emit(
|
||||
'error',
|
||||
JSON.stringify({
|
||||
type: 'error',
|
||||
data: error.message
|
||||
}));
|
||||
})
|
||||
);
|
||||
}
|
||||
|
||||
return emitter;
|
||||
}
|
||||
},
|
||||
uploads: {
|
||||
initialize: async (embeddings: Embeddings) => {
|
||||
// Pas besoin d'initialisation spécifique pour le handler uploads
|
||||
},
|
||||
searchAndAnswer: async (
|
||||
message,
|
||||
history,
|
||||
llm,
|
||||
embeddings,
|
||||
optimizationMode,
|
||||
fileIds,
|
||||
fileIds
|
||||
) => {
|
||||
const emitter = new eventEmitter();
|
||||
|
||||
|
|
@ -1171,10 +1096,11 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
`);
|
||||
|
||||
const intent = String(queryIntent.content).trim();
|
||||
console.log("🎯 Intention détectée:", intent);
|
||||
console.log('🎯 Intention détectée:', intent);
|
||||
|
||||
// Chargement optimisé des documents
|
||||
const docs = await Promise.all(fileIds.map(async fileId => {
|
||||
const docs = await Promise.all(
|
||||
fileIds.map(async fileId => {
|
||||
const filePath = path.join(process.cwd(), 'uploads', fileId);
|
||||
const contentPath = `${filePath}-extracted.json`;
|
||||
|
||||
|
|
@ -1188,7 +1114,7 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
const chunkSize = 1000; // Taille optimale pour le traitement
|
||||
const overlap = 100; // Chevauchement pour maintenir le contexte
|
||||
|
||||
const chunks = [];
|
||||
const chunks: string[] = [];
|
||||
let currentChunk = '';
|
||||
let currentSize = 0;
|
||||
|
||||
|
|
@ -1221,14 +1147,18 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
pageNumber: pageNumber,
|
||||
chunkIndex: index,
|
||||
totalChunks: chunks.length,
|
||||
searchText: chunk.substring(0, 100).replace(/[\n\r]+/g, ' ').trim()
|
||||
searchText: chunk
|
||||
.substring(0, 100)
|
||||
.replace(/[\n\r]+/g, ' ')
|
||||
.trim()
|
||||
}
|
||||
});
|
||||
});
|
||||
}));
|
||||
})
|
||||
);
|
||||
|
||||
const flatDocs = docs.flat();
|
||||
console.log("📚 Nombre total de chunks:", flatDocs.length);
|
||||
console.log('📚 Nombre total de chunks:', flatDocs.length);
|
||||
|
||||
const ragChain = new RAGDocumentChain();
|
||||
await ragChain.initializeVectorStoreFromDocuments(flatDocs, embeddings);
|
||||
|
|
@ -1238,7 +1168,8 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
let queryPrompt = message;
|
||||
switch (intent) {
|
||||
case 'SUMMARY':
|
||||
queryPrompt = "Fais un résumé complet et structuré de ce document en te concentrant sur les points clés";
|
||||
queryPrompt =
|
||||
'Fais un résumé complet et structuré de ce document en te concentrant sur les points clés';
|
||||
break;
|
||||
case 'ANALYSIS':
|
||||
queryPrompt = `Analyse en détail les aspects suivants du document concernant : ${message}. Fournis une analyse structurée avec des exemples du texte.`;
|
||||
|
|
@ -1300,13 +1231,15 @@ export const searchHandlers: Record<string, MetaSearchAgentType> = {
|
|||
emitter.emit('end');
|
||||
}
|
||||
}
|
||||
|
||||
} catch (error) {
|
||||
console.error("Erreur lors de la recherche dans les documents:", error);
|
||||
emitter.emit('error', JSON.stringify({
|
||||
console.error('Erreur lors de la recherche dans les documents:', error);
|
||||
emitter.emit(
|
||||
'error',
|
||||
JSON.stringify({
|
||||
type: 'error',
|
||||
data: error.message
|
||||
}));
|
||||
})
|
||||
);
|
||||
}
|
||||
|
||||
return emitter;
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue