feat: implement RAG Service MVP with PARSER + Memory integration
RAG Service Implementation: - Create rag-service/ with full structure (config, document_store, embedding, pipelines) - Document Store: PostgreSQL + pgvector via Haystack - Embedding: BAAI/bge-m3 (multilingual, 1024 dim) - Ingest Pipeline: Convert ParsedDocument to Haystack Documents, embed, index - Query Pipeline: Retrieve documents, generate answers via DAGI Router - FastAPI endpoints: /ingest, /query, /health Tests: - Unit tests for ingest and query pipelines - E2E test with example parsed JSON - Test fixtures with real PARSER output example Router Integration: - Add mode='rag_query' routing rule in router-config.yml - Priority 7, uses local_qwen3_8b for RAG queries Docker: - Add rag-service to docker-compose.yml - Configure dependencies (router, city-db) - Add model cache volume Documentation: - Complete README with API examples - Integration guides for PARSER and Router
This commit is contained in:
Apple
250
services/rag-service/app/query_pipeline.py
Normal file
250
services/rag-service/app/query_pipeline.py
Normal file
@@ -0,0 +1,250 @@
|
||||
"""
|
||||
Query Pipeline: RAG → LLM
|
||||
Retrieves relevant documents and generates answers
|
||||
"""
|
||||
|
||||
import logging
|
||||
from typing import List, Dict, Any, Optional
|
||||
import httpx
|
||||
|
||||
from haystack import Pipeline
|
||||
from haystack.components.embedders import SentenceTransformersTextEmbedder
|
||||
from haystack.components.retrievers import InMemoryEmbeddingRetriever
|
||||
from haystack.document_stores import PGVectorDocumentStore
|
||||
|
||||
from app.document_store import get_document_store
|
||||
from app.embedding import get_text_embedder
|
||||
from app.core.config import settings
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
async def answer_query(
|
||||
dao_id: str,
|
||||
question: str,
|
||||
top_k: Optional[int] = None,
|
||||
user_id: Optional[str] = None
|
||||
) -> Dict[str, Any]:
|
||||
"""
|
||||
Answer query using RAG pipeline
|
||||
|
||||
Args:
|
||||
dao_id: DAO identifier (for filtering)
|
||||
question: User question
|
||||
top_k: Number of documents to retrieve (default from settings)
|
||||
user_id: Optional user identifier
|
||||
|
||||
Returns:
|
||||
Dictionary with answer, citations, and retrieved documents
|
||||
"""
|
||||
logger.info(f"Answering query: dao_id={dao_id}, question={question[:50]}...")
|
||||
|
||||
top_k = top_k or settings.TOP_K
|
||||
|
||||
try:
|
||||
# Retrieve relevant documents
|
||||
documents = _retrieve_documents(dao_id, question, top_k)
|
||||
|
||||
if not documents:
|
||||
logger.warning(f"No documents found for dao_id={dao_id}")
|
||||
return {
|
||||
"answer": "На жаль, я не знайшов релевантної інформації в базі знань.",
|
||||
"citations": [],
|
||||
"documents": []
|
||||
}
|
||||
|
||||
logger.info(f"Retrieved {len(documents)} documents")
|
||||
|
||||
# Generate answer using LLM
|
||||
answer = await _generate_answer(question, documents, dao_id, user_id)
|
||||
|
||||
# Build citations
|
||||
citations = _build_citations(documents)
|
||||
|
||||
return {
|
||||
"answer": answer,
|
||||
"citations": citations,
|
||||
"documents": [
|
||||
{
|
||||
"content": doc.content[:200] + "..." if len(doc.content) > 200 else doc.content,
|
||||
"meta": doc.meta
|
||||
}
|
||||
for doc in documents
|
||||
]
|
||||
}
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Failed to answer query: {e}", exc_info=True)
|
||||
return {
|
||||
"answer": f"Помилка при обробці запиту: {str(e)}",
|
||||
"citations": [],
|
||||
"documents": []
|
||||
}
|
||||
|
||||
|
||||
def _retrieve_documents(
|
||||
dao_id: str,
|
||||
question: str,
|
||||
top_k: int
|
||||
) -> List[Any]:
|
||||
"""
|
||||
Retrieve relevant documents from DocumentStore
|
||||
|
||||
Args:
|
||||
dao_id: DAO identifier for filtering
|
||||
question: Query text
|
||||
top_k: Number of documents to retrieve
|
||||
|
||||
Returns:
|
||||
List of Haystack Document objects
|
||||
"""
|
||||
# Get components
|
||||
embedder = get_text_embedder()
|
||||
document_store = get_document_store()
|
||||
|
||||
# Embed query
|
||||
embedding_result = embedder.run(question)
|
||||
query_embedding = embedding_result["embedding"][0] if isinstance(embedding_result["embedding"], list) else embedding_result["embedding"]
|
||||
|
||||
# Retrieve with filters using vector similarity search
|
||||
filters = {"dao_id": [dao_id]}
|
||||
|
||||
try:
|
||||
documents = document_store.search(
|
||||
query_embedding=query_embedding,
|
||||
filters=filters,
|
||||
top_k=top_k,
|
||||
return_embedding=False
|
||||
)
|
||||
except Exception as e:
|
||||
logger.warning(f"Vector search failed: {e}, trying filter_documents")
|
||||
# Fallback to filter_documents
|
||||
documents = document_store.filter_documents(
|
||||
filters=filters,
|
||||
top_k=top_k,
|
||||
return_embedding=False
|
||||
)
|
||||
|
||||
# If no documents with filter, try without filter (fallback)
|
||||
if not documents:
|
||||
logger.warning(f"No documents found with dao_id={dao_id}, trying without filter")
|
||||
try:
|
||||
documents = document_store.search(
|
||||
query_embedding=query_embedding,
|
||||
filters=None,
|
||||
top_k=top_k,
|
||||
return_embedding=False
|
||||
)
|
||||
except Exception:
|
||||
documents = document_store.filter_documents(
|
||||
filters=None,
|
||||
top_k=top_k,
|
||||
return_embedding=False
|
||||
)
|
||||
|
||||
return documents
|
||||
|
||||
|
||||
async def _generate_answer(
|
||||
question: str,
|
||||
documents: List[Any],
|
||||
dao_id: str,
|
||||
user_id: Optional[str] = None
|
||||
) -> str:
|
||||
"""
|
||||
Generate answer using LLM (via DAGI Router or OpenAI)
|
||||
|
||||
Args:
|
||||
question: User question
|
||||
documents: Retrieved documents
|
||||
dao_id: DAO identifier
|
||||
user_id: Optional user identifier
|
||||
|
||||
Returns:
|
||||
Generated answer text
|
||||
"""
|
||||
# Build context from documents
|
||||
context = "\n\n".join([
|
||||
f"[Документ {idx+1}, сторінка {doc.meta.get('page', '?')}]: {doc.content[:500]}"
|
||||
for idx, doc in enumerate(documents[:3]) # Limit to first 3 documents
|
||||
])
|
||||
|
||||
# Build prompt
|
||||
prompt = (
|
||||
"Тобі надано контекст з бази знань та питання користувача.\n"
|
||||
"Відповідай на основі наданого контексту. Якщо в контексті немає відповіді, "
|
||||
"скажи що не знаєш.\n\n"
|
||||
f"Контекст:\n{context}\n\n"
|
||||
f"Питання: {question}\n\n"
|
||||
"Відповідь:"
|
||||
)
|
||||
|
||||
# Call LLM based on provider
|
||||
if settings.LLM_PROVIDER == "router":
|
||||
return await _call_router_llm(prompt, dao_id, user_id)
|
||||
elif settings.LLM_PROVIDER == "openai" and settings.OPENAI_API_KEY:
|
||||
return await _call_openai_llm(prompt)
|
||||
else:
|
||||
# Fallback: simple answer
|
||||
return f"Знайдено {len(documents)} релевантних документів. Перший фрагмент: {documents[0].content[:200]}..."
|
||||
|
||||
|
||||
async def _call_router_llm(
|
||||
prompt: str,
|
||||
dao_id: str,
|
||||
user_id: Optional[str] = None
|
||||
) -> str:
|
||||
"""Call DAGI Router LLM"""
|
||||
router_url = f"{settings.ROUTER_BASE_URL.rstrip('/')}/route"
|
||||
|
||||
payload = {
|
||||
"mode": "chat",
|
||||
"dao_id": dao_id,
|
||||
"user_id": user_id or "rag-service",
|
||||
"payload": {
|
||||
"message": prompt
|
||||
}
|
||||
}
|
||||
|
||||
try:
|
||||
async with httpx.AsyncClient(timeout=30.0) as client:
|
||||
resp = await client.post(router_url, json=payload)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
|
||||
return data.get("data", {}).get("text", "Не вдалося отримати відповідь")
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Router LLM call failed: {e}")
|
||||
return f"Помилка при виклику LLM: {str(e)}"
|
||||
|
||||
|
||||
async def _call_openai_llm(prompt: str) -> str:
|
||||
"""Call OpenAI LLM"""
|
||||
# TODO: Implement OpenAI client
|
||||
return "OpenAI integration not yet implemented"
|
||||
|
||||
|
||||
def _build_citations(documents: List[Any]) -> List[Dict[str, Any]]:
|
||||
"""
|
||||
Build citations from retrieved documents
|
||||
|
||||
Args:
|
||||
documents: List of Haystack Documents
|
||||
|
||||
Returns:
|
||||
List of citation dictionaries
|
||||
"""
|
||||
citations = []
|
||||
|
||||
for doc in documents:
|
||||
meta = doc.meta
|
||||
citations.append({
|
||||
"doc_id": meta.get("doc_id", "unknown"),
|
||||
"page": meta.get("page", 0),
|
||||
"section": meta.get("section"),
|
||||
"excerpt": doc.content[:200] + "..." if len(doc.content) > 200 else doc.content
|
||||
})
|
||||
|
||||
return citations
|
||||
|
||||
Reference in New Issue
Block a user