src/streamlit_app.py

import streamlit as st
import fitz  # PyMuPDF
import os
from sentence_transformers import SentenceTransformer
import faiss
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS
from langchain.docstore.document import Document
from langchain.chains import RetrievalQA
from langchain.llms import Groq
from dotenv import load_dotenv

load_dotenv()

# CONFIG
EMBEDDING_MODEL_NAME = "all-MiniLM-L6-v2"
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
GROQ_MODEL_NAME = "mixtral-8x7b-32768"

st.set_page_config(page_title="Lexicon - Policy Explainer", layout="centered")

st.markdown("<h1 style='text-align: center;'>📜 Lexicon: Policy Explainer Bot</h1>", unsafe_allow_html=True)
st.markdown("This app explains, summarizes, and highlights risks in large policy or T&C documents.")

uploaded_file = st.file_uploader("Upload PDF", type=["pdf"])
clipboard_input = st.text_area("Or paste policy text here", height=200)

if uploaded_file or clipboard_input:
    with st.spinner("Processing document..."):

        # Step 1: Extract Text
        def extract_text_from_pdf(file):
            doc = fitz.open(stream=file.read(), filetype="pdf")
            return " ".join(page.get_text() for page in doc)

        raw_text = extract_text_from_pdf(uploaded_file) if uploaded_file else clipboard_input

        # Step 2: Split and Embed
        sentences = raw_text.split(". ")
        model = SentenceTransformer(EMBEDDING_MODEL_NAME)
        embeddings = model.encode(sentences)

        # Step 3: Create Vector Store
        dimension = embeddings.shape[1]
        index = faiss.IndexFlatL2(dimension)
        index.add(embeddings)

        retriever = FAISS(embedding_function=lambda x: model.encode(x), index=index)
        documents = [Document(page_content=s) for s in sentences]
        retriever = FAISS.from_documents(documents, HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL_NAME)).as_retriever()

        # Step 4: LLM RAG
        llm = Groq(api_key=GROQ_API_KEY, model=GROQ_MODEL_NAME)
        qa_chain = RetrievalQA.from_chain_type(llm=llm, retriever=retriever)

        st.success("Document processed! You can now ask questions.")
        query = st.text_input("Ask a question about the document:")

        if query:
            with st.spinner("Generating answer..."):
                result = qa_chain.run(query)
                st.markdown("### 🧠 Answer:")
                st.markdown(result)

            if st.button("Suggest key risks"):
                risk_prompt = "List any risks or obligations a user should be aware of from this document."
                with st.spinner("Identifying risks..."):
                    risk_result = qa_chain.run(risk_prompt)
                    st.markdown("### ⚠️ Risks & Concerns:")
                    st.markdown(risk_result)