src/chain.py

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from src.retrieve import get_hybrid_retriever
from src import config

SYSTEM_PROMPT = """You are a Clinical Genomic Assistant.

You MUST answer strictly using the provided context.

RULES:
1. Do NOT use external knowledge.
2. If answer is not clearly present, say: "Insufficient evidence."
3. Always prioritize exact values (numbers, percentages, trial IDs, dates).
4. If multiple pieces of context exist, combine them carefully.
5. Be precise and clinical — no vague statements.

When relevant, include:
- Drug name
- Mutation
- Clinical outcome

Context:
{context}
"""

def format_docs(docs):
    return "\n\n".join(doc.page_content for doc in docs)

def get_rag_chain():
    retriever = get_hybrid_retriever()
    llm = ChatOpenAI(model=config.LLM_MODEL, temperature=0)

    prompt = ChatPromptTemplate.from_messages([
        ("system", SYSTEM_PROMPT),
        ("user", "{question}")
    ])

    def chain_with_source(query: str):
        docs = retriever.invoke(query)

        # LLM based relevance scoring
        scored_docs = []
        for doc in docs:
            score_prompt = f"""
            Query: {query}
            Document: {doc.page_content}

            Score relevance from 1 to 10 (only number).
            """
            
            score = llm.invoke(score_prompt).content.strip()
            
            try:
                score = int(score)
            except:
                score = 5
            
            scored_docs.append((score, doc))

        # Sort and keep top 5
        docs = [doc for _, doc in sorted(scored_docs, key=lambda x: x[0], reverse=True)[:5]]
        context = format_docs(docs)
        answer = (prompt | llm | StrOutputParser()).invoke({
            "context": context,
            "question": query
        })
        
        context_data = []
        for doc in docs:
            page = doc.metadata.get("page", "N/A")
            # Metadata key varies by loader- TextLoader uses 'source', PyMuPDFLoader uses 'file_path'
            source = doc.metadata.get("source", doc.metadata.get("file_path", "Unknown"))
            
            context_data.append({
                "content": doc.page_content,
                "metadata": f"Source: {source} | Page: {page}"
            })

        return {"answer": answer, "contexts": context_data}

    return chain_with_source