app.py

import os
import streamlit as st
import pickle
import time
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain.chains import RetrievalQAWithSourcesChain
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.document_loaders import UnstructuredURLLoader
from langchain_groq import ChatGroq
from langchain.vectorstores import FAISS
import numpy as np

from dotenv import load_dotenv
load_dotenv()  # take environment variables from .env (especially openai api key)

st.title("RockyBot: News Research Tool 📈")
st.sidebar.title("News Article URLs")

urls = []
for i in range(3):
    url = st.sidebar.text_input(f"URL {i+1}")
    urls.append(url)

process_url_clicked = st.sidebar.button("Process URLs")
file_path = "faiss_store_openai.pkl"

main_placeholder = st.empty()
llm = ChatGroq(model_name="llama-3.3-70b-versatile", temperature=0.9, max_tokens=500)

if process_url_clicked:
    # Load data from URLs
    loader = UnstructuredURLLoader(urls=urls)
    main_placeholder.text("Data Loading...Started...✅✅✅")
    data = loader.load()
    
    # Split data into chunks
    text_splitter = RecursiveCharacterTextSplitter(
        separators=['\n\n', '\n', '.', ','],
        chunk_size=1000
    )
    main_placeholder.text("Text Splitter...Started...✅✅✅")
    docs = text_splitter.split_documents(data)
    
    # Debugging: Check if docs is empty
    if not docs:
        main_placeholder.text("No valid documents found! Please check the URLs.")
    
    # Debugging: Check the content of docs
    for doc in docs:
        main_placeholder.text(f"Document content: {doc.page_content[:200]}")  # Show first 200 chars of each document
    
    # Create embeddings using HuggingFaceEmbeddings
    embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
    main_placeholder.text("Embedding Vector Started Building...✅✅✅")
    
    # Generate embeddings
    embeddings = embedding_model.embed_documents([doc.page_content for doc in docs])
    
    # Debugging: Check if embeddings are generated
    if not embeddings:
        main_placeholder.text("No embeddings were generated! Check the embedding model or document content.")
    
    # Check the size of embeddings
    main_placeholder.text(f"Generated {len(embeddings)} embeddings.")

    # Convert embeddings to numpy array (needed by FAISS)
    embeddings_np = np.array(embeddings).astype(np.float32)
    
    # Check the shape of the embeddings
    main_placeholder.text(f"Shape of embeddings: {embeddings_np.shape}")
    
    # Create FAISS index
    if len(embeddings) > 0:
        dimension = len(embeddings[0])  # Embedding vector dimension
        index = FAISS(dimension)
        index.add(embeddings_np)  # Add embeddings to FAISS index
        
        # Wrap FAISS index using LangChain FAISS wrapper
        vectorstore_huggingface = FAISS(embedding_function=embedding_model, index=index)
        
        # Save the FAISS index to a pickle file
        with open(file_path, "wb") as f:
            pickle.dump(vectorstore_huggingface, f)
        
        time.sleep(2)
    else:
        main_placeholder.text("Embeddings could not be generated, skipping FAISS index creation.")

query = main_placeholder.text_input("Question: ")
if query:
    if os.path.exists(file_path):
        # Load the FAISS index from the pickle file
        with open(file_path, "rb") as f:
            vectorstore = pickle.load(f)
            chain = RetrievalQAWithSourcesChain.from_llm(llm=llm, retriever=vectorstore.as_retriever())
            result = chain({"question": query}, return_only_outputs=True)
            
            # Display the answer
            st.header("Answer")
            st.write(result["answer"])

            # Display sources, if available
            sources = result.get("sources", "")
            if sources:
                st.subheader("Sources:")
                sources_list = sources.split("\n")  # Split the sources by newline
                for source in sources_list:
                    st.write(source)