day3/rag_system.py

# day3/rag_system.py
from typing import List, Dict
import os
import chromadb

from pdf_loader import load_pdf
from optimal_chunker import OptimalChunker
from embeddings import embed_texts

# LLM-i sonradan yaradacağıq
from langchain_groq import ChatGroq
from dotenv import load_dotenv

load_dotenv()


class RAGPipeline:
    def __init__(self, persist_dir: str = "./chroma_db_space", collection_name: str = "pdf_docs"):
        # Vector DB (Chroma 1.x)
        self.client = chromadb.PersistentClient(path=persist_dir)
        self.col = self.client.get_or_create_collection(name=collection_name)

        # Chunker
        self.chunker = OptimalChunker()

        # LLM hələ YARADILMIR (lazy)
        self.llm = None

    def _ensure_llm(self):
        """GROQ_API_KEY varsa LLM-i gec (ilk sorğuda) yarat."""
        if self.llm is None:
            api_key = os.getenv("GROQ_API_KEY")
            if not api_key:
                # LLM olmadan da app işə düşsün deyə aydın mesaj veririk
                raise RuntimeError(
                    "GROQ_API_KEY tapılmadı. Space Settings → Variables and secrets bölməsində əlavə edin."
                )
            self.llm = ChatGroq(model="llama3-8b-8192", temperature=0.0, api_key=api_key)

    # 1) Load  2) Chunk  3) Embed  4) Upsert to Chroma
    def index_document(self, pdf_path: str, doc_id_prefix: str = "doc") -> Dict:
        docs = load_pdf(pdf_path)
        text = "\n\n".join(d.page_content for d in docs)

        summary = self.chunker.fit_on_text(text)
        chunks = self.chunker.transform()

        vectors = embed_texts(chunks)  # list[list[float]]
        ids = [f"{doc_id_prefix}-{i}" for i in range(len(chunks))]

        self.col.add(
            ids=ids,
            documents=chunks,
            embeddings=vectors,
            metadatas=[{"source": pdf_path, "chunk": i} for i in range(len(chunks))],
        )
        return {"chunks_indexed": len(chunks), "best_strategy": summary}

    # 5) Retrieve  6) Ask LLM
    def query(self, question: str, k: int = 4) -> Dict:
        # LLM-i bu zaman yaradacağıq (secret yoxdursa burda aydın xəta görünəcək)
        self._ensure_llm()

        results = self.col.query(query_texts=[question], n_results=k)
        chunks: List[str] = results["documents"][0] if results.get("documents") else []

        context = "\n\n".join(chunks)
        prompt = f"""You are an extraction assistant. Use ONLY the Context to answer.
Rules:
- If the answer is explicitly present in Context, return that substring EXACTLY.
- Do not paraphrase. Do not add words. Return a verbatim span from Context.
- If the answer is not in Context, reply exactly: I don't know

Question: {question}

Context:
{context}

Answer (verbatim from Context):"""
        resp = self.llm.invoke(prompt)
        answer = resp.content.strip() if hasattr(resp, "content") else str(resp)

        if (not answer or answer.lower().startswith("i don't know")) and context.strip():
            answer = chunks[0] if chunks else "I don't know"

        return {"answer": answer, "used_chunks": len(chunks), "context_preview": context[:500]}