app.py

import os
import streamlit as st
import faiss
import pickle
from datasets import load_dataset
from sentence_transformers import SentenceTransformer
from groq import Groq

# Constants
DATASET_NAME = "neural-bridge/rag-dataset-1200"
MODEL_NAME = "all-MiniLM-L6-v2"
INDEX_FILE = "faiss_index.pkl"
DOCS_FILE = "contexts.pkl"

# Groq API client
client = Groq(api_key=os.environ.get("MY_KEY"))

# Streamlit page setup
st.set_page_config(page_title="RAG App", layout="wide")
st.title("🧠 Retrieval-Augmented Generation (RAG) with Groq")

# Function to load or create database
@st.cache_resource
def setup_database():
    st.info("Setting up vector database...")
    progress = st.progress(0)

    # Step 1: Load dataset
    dataset = load_dataset(DATASET_NAME, split="train")
    contexts = [entry["context"] for entry in dataset]
    progress.progress(25)

    # Step 2: Compute embeddings
    embedder = SentenceTransformer(MODEL_NAME)
    embeddings = embedder.encode(contexts, show_progress_bar=True)
    progress.progress(50)

    # Step 3: Build FAISS index
    dimension = embeddings[0].shape[0]
    faiss_index = faiss.IndexFlatL2(dimension)
    faiss_index.add(embeddings)
    progress.progress(75)

    # Step 4: Save index and contexts for future use
    with open(INDEX_FILE, "wb") as f:
        pickle.dump(faiss_index, f)
    with open(DOCS_FILE, "wb") as f:
        pickle.dump(contexts, f)

    progress.progress(100)
    st.success("Database setup complete!")
    return faiss_index, contexts

# Check if the index and contexts are saved, otherwise set up
if os.path.exists(INDEX_FILE) and os.path.exists(DOCS_FILE):
    with open(INDEX_FILE, "rb") as f:
        faiss_index = pickle.load(f)
    with open(DOCS_FILE, "rb") as f:
        all_contexts = pickle.load(f)
    st.info("Loaded existing database.")
else:
    faiss_index, all_contexts = setup_database()

# UI for sample questions
sample_questions = [
    "What is the purpose of the RAG dataset?",
    "How does Falcon RefinedWeb contribute to this dataset?",
    "What are the benefits of using retrieval-augmented generation?",
    "Explain the structure of the RAG-1200 dataset.",
]

st.subheader("Ask a question based on the dataset:")
question = st.text_input("Enter your question:", value=sample_questions[0])

if st.button("Ask"):
    if question.strip() == "":
        st.warning("Please enter a question.")
    else:
        with st.spinner("Retrieving and generating answer..."):
            # Embed user query
            embedder = SentenceTransformer(MODEL_NAME)
            query_embedding = embedder.encode([question])
            D, I = faiss_index.search(query_embedding, k=1)

            # Get closest context
            context = all_contexts[I[0][0]]
            prompt = f"Context: {context}\n\nQuestion: {question}\n\nAnswer:"

            # Call Groq model
            response = client.chat.completions.create(
                messages=[{"role": "user", "content": prompt}],
                model="llama3-70b-8192"
            )

            answer = response.choices[0].message.content
            st.success("Answer:")
            st.markdown(answer)

            with st.expander("🔍 Retrieved Context"):
                st.markdown(context)