app.py

import asyncio
import os

try:
    asyncio.get_running_loop()
except RuntimeError:
    loop = asyncio.new_event_loop()
    asyncio.set_event_loop(loop)

import streamlit as st
from langchain_community.vectorstores import FAISS
from langchain_community.document_loaders import PyMuPDFLoader, Docx2txtLoader
from langchain_huggingface import HuggingFaceEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_ollama import OllamaLLM
from langchain.docstore.document import Document
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from huggingface_hub import InferenceClient

os.environ["STREAMLIT_WATCHER_TYPE"] = "none"

@st.cache_resource
def load_llm():
    client = InferenceClient(model="microsoft/phi-3-mini-4k-instruct")
    return client

# 🧠 Cache embedder
@st.cache_resource
def load_embedder():
    return HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")

llm = load_llm()
embedder = load_embedder()

# Sidebar Upload
st.sidebar.title("📄 Upload Terms & Conditions")
input_mode = st.sidebar.radio("Choose Input Method", ["📋 Paste Text", "📁 Upload File"])

uploaded_text = ""
if input_mode == "📋 Paste Text":
    uploaded_text = st.sidebar.text_area("Paste your T&C text here")
elif input_mode == "📁 Upload File":
    uploaded_file = st.sidebar.file_uploader("Upload a .txt, .pdf, or .docx file", type=["txt", "pdf", "docx"])
    if uploaded_file:
        if uploaded_file.type == "text/plain":
            uploaded_text = uploaded_file.read().decode("utf-8")
        elif uploaded_file.type == "application/pdf":
            with open("temp.pdf", "wb") as f:
                f.write(uploaded_file.read())
            docs = PyMuPDFLoader("temp.pdf").load()
            uploaded_text = "\n".join([d.page_content for d in docs])
        elif uploaded_file.type == "application/vnd.openxmlformats-officedocument.wordprocessingml.document":
            with open("temp.docx", "wb") as f:
                f.write(uploaded_file.read())
            docs = Docx2txtLoader("temp.docx").load()
            uploaded_text = "\n".join([d.page_content for d in docs])

# ✅ Vectorstore setup
if uploaded_text:
    st.success("✅ Document loaded and processed!")

    if "db" not in st.session_state:
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
        documents = text_splitter.create_documents([uploaded_text])
        db = FAISS.from_documents(documents, embedder)
        st.session_state.db = db
        st.session_state.chat_history = []

# 💬 Chat section
if "db" in st.session_state:
    st.title("🧾 Legal Assistant Chat")
    st.markdown("Ask anything about the uploaded document.")

    user_input = st.chat_input("Type your question here...")

    if user_input:
        with st.spinner("🤖 Thinking..."):
            retriever = st.session_state.db.as_retriever(search_kwargs={"k": 3})
            docs = retriever.invoke(user_input)
            context = "\n\n".join([doc.page_content for doc in docs])

            prompt = f"""You are a helpful legal assistant.
Based on the following contract, answer the user's question, This application built by Vighnesh.

Context:
{context}

Question:
{user_input}

Answer:"""

            answer = llm.text_generation(prompt, max_new_tokens=200)

            # Save chat history
            st.session_state.chat_history.append(("user", user_input))
            st.session_state.chat_history.append(("assistant", answer))

    # Display chat history
    for role, message in st.session_state.chat_history:
        if role == "user":
            st.chat_message("user").write(message)
        else:
            st.chat_message("assistant").write(message)