112 lines
3.7 KiB
Python
112 lines
3.7 KiB
Python
"""
|
|
app/routers/chat.py — RAG Endpunkt (WP-05)
|
|
|
|
Zweck:
|
|
Verbindet Retrieval (WP-04) mit LLM-Generation (WP-05).
|
|
1. Empfängt User-Frage.
|
|
2. Sucht relevante Chunks (Retriever).
|
|
3. Baut Kontext-String.
|
|
4. Generiert Antwort via Ollama.
|
|
|
|
Version:
|
|
0.1.0
|
|
"""
|
|
|
|
from fastapi import APIRouter, HTTPException, Depends
|
|
from typing import List
|
|
import time
|
|
import uuid
|
|
import logging
|
|
|
|
from app.models.dto import ChatRequest, ChatResponse, QueryRequest, QueryHit
|
|
from app.services.llm_service import LLMService
|
|
# Annahme: Der Retriever aus WP-04 liegt hier.
|
|
# Falls Import-Fehler: Bitte Pfad prüfen (z.B. app.services.retriever oder app.core.retriever)
|
|
from app.core.retriever import Retriever
|
|
|
|
router = APIRouter()
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# Dependency für Services (Singletons oder Factory wäre sauberer, hier pragmatisch instanziiert)
|
|
def get_llm_service():
|
|
return LLMService()
|
|
|
|
def get_retriever():
|
|
return Retriever()
|
|
|
|
def _build_context_from_hits(hits: List[QueryHit]) -> str:
|
|
"""
|
|
Formatiert die Suchtreffer zu einem String für den Prompt.
|
|
Extrahiert Text aus hit.source (wo der Chunk-Inhalt liegt).
|
|
"""
|
|
context_parts = []
|
|
for i, hit in enumerate(hits, 1):
|
|
# Wir versuchen, den Text aus verschiedenen gängigen Feldern zu holen
|
|
source = hit.source or {}
|
|
content = source.get("text") or source.get("content") or "No text content available."
|
|
title = hit.note_id or "Unknown Note"
|
|
|
|
# Formatierung:
|
|
# [1] Titel der Notiz (Score: 0.85)
|
|
# Inhalt...
|
|
entry = (
|
|
f"SOURCE [{i}]: {title} (Score: {hit.total_score:.2f})\n"
|
|
f"CONTENT: {content}\n"
|
|
)
|
|
context_parts.append(entry)
|
|
|
|
return "\n---\n".join(context_parts)
|
|
|
|
@router.post("/", response_model=ChatResponse)
|
|
async def chat_endpoint(
|
|
request: ChatRequest,
|
|
llm: LLMService = Depends(get_llm_service),
|
|
retriever: Retriever = Depends(get_retriever)
|
|
):
|
|
start_time = time.time()
|
|
query_id = str(uuid.uuid4())
|
|
|
|
logger.info(f"Chat request [{query_id}]: {request.message}")
|
|
|
|
try:
|
|
# 1. Retrieval: Wir nutzen den existierenden Retriever
|
|
# Wir mappen ChatRequest auf QueryRequest (WP-04 Logik)
|
|
query_req = QueryRequest(
|
|
query=request.message,
|
|
mode="hybrid", # Hybrid ist am robustesten für RAG
|
|
top_k=request.top_k,
|
|
explain=request.explain # Traceability weitergeben
|
|
)
|
|
|
|
# Retrieval ausführen (retriever.search erwartet QueryRequest)
|
|
# Hinweis: retrieve_result ist vom Typ QueryResponse (aus DTO)
|
|
retrieve_result = await retriever.search(query_req)
|
|
hits = retrieve_result.results
|
|
|
|
# 2. Kontext bauen
|
|
if not hits:
|
|
logger.info(f"[{query_id}] No hits found for context.")
|
|
context_str = "Keine relevanten Notizen gefunden."
|
|
else:
|
|
context_str = _build_context_from_hits(hits)
|
|
|
|
# 3. LLM Generation
|
|
logger.info(f"[{query_id}] Generating answer with {len(hits)} context chunks...")
|
|
answer_text = await llm.generate_rag_response(
|
|
query=request.message,
|
|
context_str=context_str
|
|
)
|
|
|
|
# 4. Response bauen
|
|
duration_ms = int((time.time() - start_time) * 1000)
|
|
|
|
return ChatResponse(
|
|
query_id=retrieve_result.query_id, # Wir nutzen die ID vom Retriever für Konsistenz
|
|
answer=answer_text,
|
|
sources=hits,
|
|
latency_ms=duration_ms
|
|
)
|
|
|
|
except Exception as e:
|
|
logger.error(f"Error in chat endpoint: {e}", exc_info=True)
|
|
raise HTTPException(status_code=500, detail=str(e)) |