main.py

import os
import zipfile
import tempfile
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel

from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_groq import ChatGroq
from langchain.chains import RetrievalQA
from langchain.prompts import PromptTemplate

app = FastAPI()

# === Globals ===
llm = None
embeddings = None
vectorstore = None
retriever = None
chain = None


class QueryRequest(BaseModel):
    question: str


def _unpack_faiss(src_path: str) -> str:
    """
    If src_path is a ZIP, unzip it into a temp dir and return the folder
    containing the .faiss files; if it’s already a folder, return it.
    """
    if zipfile.is_zipfile(src_path):
        tmp = tempfile.TemporaryDirectory()
        with zipfile.ZipFile(src_path, "r") as zf:
            zf.extractall(tmp.name)
        for root, _, files in os.walk(tmp.name):
            if any(f.endswith(".faiss") for f in files):
                return root
        raise RuntimeError(f"No .faiss index found inside ZIP: {src_path}")
    elif os.path.isdir(src_path):
        return src_path
    else:
        raise RuntimeError(f"Path is neither a valid ZIP nor a directory: {src_path}")


def load_and_merge_faiss(path1: str, path2: str, embeddings: HuggingFaceEmbeddings) -> FAISS:
    """
    Load two FAISS indexes (either zip files or folders), merge them,
    and return the combined FAISS vectorstore.
    """
    dir1 = _unpack_faiss(path1)
    dir2 = _unpack_faiss(path2)

    vs1 = FAISS.load_local(dir1, embeddings, allow_dangerous_deserialization=True)
    vs2 = FAISS.load_local(dir2, embeddings, allow_dangerous_deserialization=True)
    vs1.merge_from(vs2)
    return vs1


@app.on_event("startup")
def load_components():
    global llm, embeddings, vectorstore, retriever, chain

    # --- 1) Init LLM & Embeddings ---
    llm = ChatGroq(
        model="meta-llama/llama-4-scout-17b-16e-instruct",
        temperature=0,
        max_tokens=1024,
        api_key=os.getenv("API_KEY"),
    )
    embeddings = HuggingFaceEmbeddings(
        model_name="intfloat/multilingual-e5-large",
        model_kwargs={"device": "cpu"},
        encode_kwargs={"normalize_embeddings": True},
    )

    # --- 2) Load & merge two FAISS indexes ---
    src1 = os.getenv("FAISS_INDEX_PATH_1", "faiss_index.zip")
    src2 = os.getenv("FAISS_INDEX_PATH_2", "faiss_index_extra.zip")
    vectorstore = load_and_merge_faiss(src1, src2, embeddings)

    # --- 3) Build retriever & QA chain ---
    retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
    prompt = PromptTemplate(
        template="""
You are an expert assistant on Islamic knowledge.
Use **only** the information in the “Retrieved context” to answer the user’s question.
Do **not** add any outside information, personal opinions, or conjecture—if the answer is not contained in the context, reply with “لا أعلم”.
Be concise, accurate, and directly address the user’s question.

Retrieved context:
{context}

User’s question:
{question}

Your response:
""",
        input_variables=["context", "question"],
    )
    chain = RetrievalQA.from_chain_type(
        llm=llm,
        chain_type="stuff",
        retriever=retriever,
        return_source_documents=False,
        chain_type_kwargs={"prompt": prompt},
    )

    print("✅ Loaded & merged both FAISS indexes, QA chain ready.")


@app.get("/")
def root():
    return {"message": "Arabic Hadith Finder API is up and running!"}


@app.post("/query")
def query(request: QueryRequest):
    try:
        result = chain.invoke({"query": request.question})
        return {"answer": result["result"]}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))