app.py

# app.py
import os
import glob
import tempfile
from typing import List
import streamlit as st

# LangChain / loaders / vectorstore / embeddings / LLM
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain.embeddings import HuggingFaceEmbeddings
from langchain_groq import ChatGroq
from langchain.chains import RetrievalQA

st.set_page_config(page_title="RAG Papers Chat (Groq)", layout="wide")

# -----------------------
# Load custom CSS
# -----------------------
def load_css(path="style.css"):
    if os.path.exists(path):
        with open(path) as f:
            st.markdown(f"<style>{f.read()}</style>", unsafe_allow_html=True)

load_css()

# -----------------------
# Sidebar / settings
# -----------------------
st.sidebar.title("⚙️ Settings")
chunk_size = st.sidebar.number_input("Chunk size", min_value=256, max_value=5000, value=1000, step=100)
chunk_overlap = st.sidebar.number_input("Chunk overlap", min_value=0, max_value=1000, value=200, step=50)
top_k = st.sidebar.slider("Top-k chunks to retrieve", min_value=1, max_value=10, value=3)
model_choice = st.sidebar.selectbox(
    "Groq model",
    options=["llama-3.1-8b-instant", "llama-3.1-8b-8192", "mixtral-3b-16384"],
    index=0
)
st.sidebar.markdown("🔑 Your **Groq API key** must be set as a secret (`GROQ_API_KEY`) in Hugging Face Settings.")

# -----------------------
# Utility functions
# -----------------------
@st.cache_data(show_spinner=False)
def load_and_split_pdfs(file_paths: List[str], chunk_size: int, chunk_overlap: int):
    """Load PDFs and return list of split documents (LangChain docs)."""
    all_docs = []
    splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
    for path in file_paths:
        loader = PyPDFLoader(path)
        loaded = loader.load()
        splitted = splitter.split_documents(loaded)
        all_docs.extend(splitted)
    return all_docs

@st.cache_resource(show_spinner=False)
def build_vectorstore(docs):
    """Create HuggingFace embeddings + FAISS vectorstore."""
    embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
    vectorstore = FAISS.from_documents(docs, embeddings)
    return vectorstore

def initialize_llm(model_name: str):
    api_key = os.environ.get("GROQ_API_KEY")
    if not api_key:
        st.error("🚨 No `GROQ_API_KEY` found. Please add it in Hugging Face Space → Settings → Secrets.")
        st.stop()
    return ChatGroq(model=model_name, api_key=api_key, temperature=0)

# -----------------------
# Main UI
# -----------------------
st.title("📚 RAG Chat for Research Papers — Streamlit (Groq)")
st.write("Upload multiple PDFs and ask questions. Answers will include deduplicated file sources.")

uploaded_files = st.file_uploader("Upload PDF files", type="pdf", accept_multiple_files=True)
process_btn = st.button("Process uploaded PDFs")

if process_btn:
    if not uploaded_files:
        st.warning("Please upload one or more PDF files first.")
    else:
        tmp_paths = []
        for f in uploaded_files:
            tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".pdf")
            tmp.write(f.read())
            tmp.flush()
            tmp_paths.append(tmp.name)

        st.success("✅ PDFs saved. Processing...")

        with st.spinner("Splitting into chunks..."):
            docs = load_and_split_pdfs(tmp_paths, chunk_size, chunk_overlap)
            st.write(f"✅ Created {len(docs)} chunks.")

        with st.spinner("Building FAISS vectorstore..."):
            vectorstore = build_vectorstore(docs)

        st.session_state["vectorstore"] = vectorstore
        st.session_state["processed"] = True
        st.success("✅ Vectorstore ready! Ask questions below.")

# -----------------------
# Chat section
# -----------------------
st.markdown("---")
st.subheader("💬 Chat with your papers")

if "processed" not in st.session_state:
    st.info("Process PDFs first to build the index.")
else:
    if "llm" not in st.session_state:
        st.session_state["llm"] = initialize_llm(model_choice)

    if "qa_chain" not in st.session_state:
        retriever = st.session_state["vectorstore"].as_retriever(search_kwargs={"k": top_k})
        st.session_state["qa_chain"] = RetrievalQA.from_chain_type(
            llm=st.session_state["llm"],
            retriever=retriever,
            chain_type="stuff",
            return_source_documents=True,
        )

    if "history" not in st.session_state:
        st.session_state["history"] = []

    query = st.text_input("Enter your question")
    ask = st.button("Ask")

    if ask and query.strip():
        with st.spinner("Thinking..."):
            result = st.session_state["qa_chain"]({"query": query})
            answer = result.get("result", "")
            source_docs = result.get("source_documents", [])

            unique_sources = list({doc.metadata.get("source", "unknown") for doc in source_docs})
            sources_text = "\n".join([f"- {os.path.basename(s)}" for s in unique_sources])

            full_answer = f"{answer}\n\n📚 **Sources:**\n{sources_text}"
            st.session_state["history"].append((query, full_answer))

    st.markdown("### 📜 Conversation History")
    for user_msg, bot_msg in reversed(st.session_state["history"]):
        st.markdown(f"**You:** {user_msg}")
        st.markdown(f"**Bot:** {bot_msg}")