backend/services/qa.py

"""RAG QA chain with streaming and multi-doc modes."""

from typing import Dict, Iterator, List, Literal, Optional, Union

from db.faiss_client import FaissDB
from models.embedder import MiniCPMEmbedder
from models.llm import MiniCPMLLM
from services.retrieval import RetrievalService
from utils.formatters import build_context_block
from utils.response_cleaner import clean_model_response

SearchMode = Literal["all", "single"]


class QAService:
    def __init__(
        self,
        embedder: MiniCPMEmbedder,
        llm: MiniCPMLLM,
        db: FaissDB,
    ):
        self.embedder = embedder
        self.llm = llm
        self.db = db
        self.retrieval = RetrievalService(embedder, db)

    def stream_answer(
        self,
        query: str,
        chat_history: Optional[List[Dict]] = None,
        document_ids: Optional[List[str]] = None,
        mode: SearchMode = "all",
    ) -> Iterator[Union[str, Dict]]:
        if mode == "single":
            filter_ids = [document_ids[0]] if document_ids else None
        elif document_ids:
            filter_ids = document_ids
        else:
            filter_ids = None

        results = self.retrieval.search(query, document_ids=filter_ids)
        yield {"type": "sources", "data": results}
        if not results:
            yield "I couldn't find relevant information in the loaded documents."
            return

        context = build_context_block(results)
        full_answer = []
        for token in self.llm.stream_answer(query, context, chat_history):
            full_answer.append(token)
            yield token

        answer_text = clean_model_response("".join(full_answer))
        confidence = self.llm.evaluate_confidence(query, context, answer_text)
        yield {"type": "confidence", "data": confidence}