app.py

import logging
import json
import pandas as pd
import streamlit as st
from pinecone import Pinecone
from llama_index.vector_stores.pinecone import PineconeVectorStore
from llama_index.core import (
    StorageContext, VectorStoreIndex, SimpleDirectoryReader, 
    get_response_synthesizer, Settings
)
from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.retrievers import (
    VectorIndexRetriever, RouterRetriever
)
from llama_index.retrievers.bm25 import BM25Retriever
from llama_index.core.tools import RetrieverTool
from llama_index.core.query_engine import (
    RetrieverQueryEngine, FLAREInstructQueryEngine, MultiStepQueryEngine
)
from llama_index.core.indices.query.query_transform import (
    StepDecomposeQueryTransform
)
from llama_index.llms.groq import Groq
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
from llama_index.llms.azure_openai import AzureOpenAI
from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.readers.file import PyMuPDFReader
import traceback
from oauth2client.service_account import ServiceAccountCredentials
import gspread
import uuid
from dotenv import load_dotenv
import os
from datetime import datetime

# Load environment variables
load_dotenv()

# Configure logging
logging.basicConfig(level=logging.INFO)

# Google Sheets setup
scope = ["https://spreadsheets.google.com/feeds", "https://www.googleapis.com/auth/drive"]
creds_dict = {
    "type": os.getenv("type"),
    "project_id": os.getenv("project_id"),
    "private_key_id": os.getenv("private_key_id"),
    "private_key": os.getenv("private_key"),
    "client_email": os.getenv("client_email"),
    "client_id": os.getenv("client_id"),
    "auth_uri": os.getenv("auth_uri"),
    "token_uri": os.getenv("token_uri"),
    "auth_provider_x509_cert_url": os.getenv("auth_provider_x509_cert_url"),
    "client_x509_cert_url": os.getenv("client_x509_cert_url")
}
creds_dict['private_key'] = creds_dict['private_key'].replace('\\n', '\n')
creds = ServiceAccountCredentials.from_json_keyfile_dict(creds_dict, scope)
client = gspread.authorize(creds)
sheet = client.open("RAG").sheet1

# Fixed variables
AZURE_DEPLOYMENT_NAME = os.getenv("AZURE_DEPLOYMENT_NAME")
AZURE_API_VERSION = os.getenv("AZURE_API_VERSION")
AZURE_OPENAI_ENDPOINT = os.getenv("AZURE_OPENAI_ENDPOINT")

# Global variables for lazy loading
llm = None
pinecone_index = None

def log_and_exit(message):
    logging.error(message)
    raise SystemExit(message)

def initialize_apis(api, model, pinecone_api_key, groq_api_key, azure_api_key):
    global llm, pinecone_index
    try:
        if llm is None:
            llm = initialize_llm(api, model, groq_api_key, azure_api_key)
        if pinecone_index is None:
            pinecone_client = Pinecone(pinecone_api_key)
            pinecone_index = pinecone_client.Index("ll144")
        logging.info("Initialized LLM and Pinecone.")
    except Exception as e:
        log_and_exit(f"Error initializing APIs: {e}")

def initialize_llm(api, model, groq_api_key, azure_api_key):
    if api == 'groq':
        model_mappings = {
            'mixtral-8x7b': "mixtral-8x7b-32768",
            'llama3-8b': "llama3-8b-8192",
            'llama3-70b': "llama3-70b-8192",
            'gemma-7b': "gemma-7b-it"
        }
        return Groq(model=model_mappings[model], api_key=groq_api_key)
    elif api == 'azure':
        if model == 'gpt35':
            return AzureOpenAI(
                deployment_name=AZURE_DEPLOYMENT_NAME,
                temperature=0,
                api_key=azure_api_key,
                azure_endpoint=AZURE_OPENAI_ENDPOINT,
                api_version=AZURE_API_VERSION
            )

def load_pdf_data(chunk_size):
    reader = PyMuPDFReader()
    file_extractor = {".pdf": reader}
    documents = SimpleDirectoryReader(input_files=['LL144.pdf', 'LL144_Definitions.pdf'], file_extractor=file_extractor).load_data()
    return documents

def create_index(documents, embedding_model_type="HF", embedding_model="BAAI/bge-large-en-v1.5", retriever_method="BM25", chunk_size=512):
    global llm, pinecone_index
    try:
        embed_model = select_embedding_model(embedding_model_type, embedding_model)

        Settings.llm = llm
        Settings.embed_model = embed_model
        Settings.chunk_size = chunk_size

        if retriever_method in ["BM25", "BM25+Vector"]:
            nodes = create_bm25_nodes(documents, chunk_size)
            logging.info("Created BM25 nodes from documents.")
            if retriever_method == "BM25+Vector":
                vector_store = PineconeVectorStore(pinecone_index=pinecone_index)
                storage_context = StorageContext.from_defaults(vector_store=vector_store)
                index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)
                logging.info("Created index for BM25+Vector from documents.")
                return index, nodes
            return None, nodes
        else:
            vector_store = PineconeVectorStore(pinecone_index=pinecone_index)
            storage_context = StorageContext.from_defaults(vector_store=vector_store)
            index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)
            logging.info("Created index from documents.")
            return index, None
    except Exception as e:
        log_and_exit(f"Error creating index: {e}")

def select_embedding_model(embedding_model_type, embedding_model):
    if embedding_model_type == "HF":
        return HuggingFaceEmbedding(model_name=embedding_model)
    elif embedding_model_type == "OAI":
        return OpenAIEmbedding()  # Implement OAI Embedding if needed

def create_bm25_nodes(documents, chunk_size):
    splitter = SentenceSplitter(chunk_size=chunk_size)
    nodes = splitter.get_nodes_from_documents(documents)
    return nodes

def select_retriever(index, nodes, retriever_method, top_k):
    logging.info(f"Selecting retriever with method: {retriever_method}")
    if nodes is not None:
        logging.info(f"Available document IDs: {list(range(len(nodes)))}")
    else:
        logging.warning("Nodes are None")

    if retriever_method == 'BM25':
        return BM25Retriever.from_defaults(nodes=nodes, similarity_top_k=top_k)
    elif retriever_method == "BM25+Vector":
        if index is None:
            log_and_exit("Index must be initialized when using BM25+Vector retriever method.")
        
        bm25_retriever = RetrieverTool.from_defaults(
            retriever=BM25Retriever.from_defaults(nodes=nodes, similarity_top_k=top_k),
            description="BM25 Retriever"
        )
        
        vector_retriever = RetrieverTool.from_defaults(
            retriever=VectorIndexRetriever(index=index),
            description="Vector Retriever"
        )
        
        router_retriever = RouterRetriever.from_defaults(
            retriever_tools=[bm25_retriever, vector_retriever],
            llm=llm,
            select_multi=True
        )
        return router_retriever
    elif retriever_method == "Vector Search":
        if index is None:
            log_and_exit("Index must be initialized when using Vector Search retriever method.")
        return VectorIndexRetriever(index=index, similarity_top_k=top_k)
    else:
        log_and_exit(f"Unsupported retriever method: {retriever_method}")

def setup_query_engine(index, response_mode, nodes=None, query_engine_method=None, retriever_method=None, top_k=2):
    global llm
    try:
        logging.info(f"Setting up query engine with retriever_method: {retriever_method} and query_engine_method: {query_engine_method}")
        retriever = select_retriever(index, nodes, retriever_method, top_k)
        
        if retriever is None:
            log_and_exit("Failed to create retriever. Index or nodes might be None.")

        response_synthesizer = get_response_synthesizer(response_mode=response_mode)
        index_query_engine = index.as_query_engine(similarity_top_k=top_k) if index else None

        if query_engine_method == "FLARE":
            query_engine = FLAREInstructQueryEngine(
                query_engine=index_query_engine,
                max_iterations=4,
                verbose=False
            )
        elif query_engine_method == "MS":
            query_engine = MultiStepQueryEngine(
                query_engine=index_query_engine,
                query_transform=StepDecomposeQueryTransform(llm=llm, verbose=False),
                index_summary="Used to answer questions about the regulation"
            )
        else:
            query_engine = RetrieverQueryEngine(retriever=retriever, response_synthesizer=response_synthesizer)
        
        if query_engine is None:
            log_and_exit("Failed to create query engine.")
        
        return query_engine
    except Exception as e:
        logging.error(f"Error setting up query engine: {e}")
        traceback.print_exc()
        log_and_exit(f"Error setting up query engine: {e}")

def log_to_google_sheets(data):
    try:
        sheet.append_row(data)
        logging.info("Logged data to Google Sheets.")
    except Exception as e:
        logging.error(f"Error logging data to Google Sheets: {e}")

def update_google_sheets(question_id, feedback=None, detailed_feedback=None, annotated_answer=None):
    try:
        existing_data = sheet.get_all_values()
        headers = existing_data[0]
        for i, row in enumerate(existing_data):
            if row[0] == question_id:
                if feedback is not None:
                    sheet.update_cell(i+1, headers.index("Feedback") + 1, feedback)
                if detailed_feedback is not None:
                    sheet.update_cell(i+1, headers.index("Detailed Feedback") + 1, detailed_feedback)
                if annotated_answer is not None:
                    sheet.update_cell(i+1, headers.index("annotated_answer") + 1, annotated_answer)
                logging.info("Updated data in Google Sheets.")
                return
    except Exception as e:
        logging.error(f"Error updating data in Google Sheets: {e}")

def run_streamlit_app():
    if 'query_engine' not in st.session_state:
        st.session_state.query_engine = None

    st.title("RAG Chat Application")

    col1, col2 = st.columns(2)

    with col1:
        pinecone_api_key = st.text_input("Pinecone API Key")
        azure_api_key = st.text_input("Azure API Key")
        groq_api_key = st.text_input("Groq API Key")

    def update_api_based_on_model():
        selected_model = st.session_state['selected_model']
        if selected_model == 'gpt35':
            st.session_state['selected_api'] = 'azure'
        else:
            st.session_state['selected_api'] = 'groq'

    with col2:
        selected_model = st.selectbox("Select Model", ["llama3-8b", "llama3-70b", "mixtral-8x7b", "gemma-7b", "gpt35"], index=4, key='selected_model', on_change=update_api_based_on_model)
        selected_api = st.selectbox("Select API", ["azure", "groq"], index=0, key='selected_api', disabled=True)
        embedding_model_type = "HF"
        embedding_model = st.selectbox("Select Embedding Model", ["BAAI/bge-large-en-v1.5", "other_model"])
        retriever_method = st.selectbox("Select Retriever Method", ["Vector Search", "BM25", "BM25+Vector"])

    col3, col4 = st.columns(2)
    with col3:
        chunk_size = st.selectbox("Select Chunk Size", [128, 256, 512, 1024], index=2)
    with col4:
        top_k = st.selectbox("Select Top K", [1, 2, 3, 5, 6], index=1)

    if st.button("Initialize"):
        initialize_apis(st.session_state['selected_api'], selected_model, pinecone_api_key, groq_api_key, azure_api_key)
        documents = load_pdf_data(chunk_size)
        index, nodes = create_index(documents, embedding_model_type=embedding_model_type, embedding_model=embedding_model, retriever_method=retriever_method, chunk_size=chunk_size)
        st.session_state.query_engine = setup_query_engine(index, response_mode="compact", nodes=nodes, query_engine_method=None, retriever_method=retriever_method, top_k=top_k)
        st.success("Initialization complete.")

    if 'chat_history' not in st.session_state:
        st.session_state.chat_history = []

    for chat_index, chat in enumerate(st.session_state.chat_history):
        with st.chat_message("user"):
            st.markdown(chat['user'])
        with st.chat_message("bot"):
            st.markdown("### Retrieved Contexts")
            for node in chat.get('contexts', []):
                st.markdown(
                    f"<div style='border:1px solid #ccc; padding:10px; margin:10px 0; font-size:small;'>{node.text}</div>",
                    unsafe_allow_html=True
                )
            st.markdown("### Answer")
            st.markdown(chat['response'])

            col1, col2 = st.columns([1, 1])
            with col1:
                if st.button("Annotate 👎", key=f"annotate_{chat_index}"):
                    chat['annotate'] = True
                    chat['feedback'] = -1
                    st.session_state.chat_history[chat_index] = chat
                    update_google_sheets(chat['id'], feedback=-1)
                    st.rerun()
            with col2:
                if st.button("Approve 👍", key=f"approve_{chat_index}"):
                    chat['approved'] = True
                    chat['feedback'] = 1
                    st.session_state.chat_history[chat_index] = chat
                    update_google_sheets(chat['id'], feedback=1, annotated_answer=chat['response'])

            if chat.get('annotate', False):
                annotated_answer = st.text_area("Annotate Answer", value=chat['response'], key=f"annotate_text_{chat_index}")
                if st.button("Submit Annotated Answer", key=f"submit_annotate_{chat_index}"):
                    chat['annotated_answer'] = annotated_answer
                    chat['annotate'] = False
                    st.session_state.chat_history[chat_index] = chat
                    update_google_sheets(chat['id'], annotated_answer=annotated_answer)
            
            feedback = st.text_area("How was the response? Does it match the context? Does it answer the question fully?", key=f"textarea_{chat_index}")
            if st.button("Submit Feedback", key=f"submit_{chat_index}"):
                chat['detailed_feedback'] = feedback
                st.session_state.chat_history[chat_index] = chat
                update_google_sheets(chat['id'], detailed_feedback=feedback)

    if question := st.chat_input("Enter your question"):
        if st.session_state.query_engine:
            with st.spinner('Generating response...'):
                # Compile chat history for context
                history = "\n".join([f"Q: {chat['user']}\nA: {chat['response']}" for chat in st.session_state.chat_history])
                full_query = f"{history}\nQ: {question}"
                response = st.session_state.query_engine.query(full_query)
                logging.info(f"Generated response: {response.response}")
                logging.info(f"Retrieved contexts: {[node.text for node in response.source_nodes]}")
            question_id = str(uuid.uuid4())
            timestamp = datetime.now().isoformat()
            st.session_state.chat_history.append({'id': question_id, 'user': question, 'response': response.response, 'contexts': response.source_nodes, 'feedback': 0, 'detailed_feedback': '', 'annotated_answer': '', 'timestamp': timestamp})
            
            # Log initial query and response to Google Sheets without feedback
            log_to_google_sheets([question_id, question, response.response, st.session_state['selected_api'], selected_model, embedding_model, retriever_method, chunk_size, top_k, 0, "", "", timestamp])
            
            st.rerun()
        else:
            st.error("Query engine is not initialized. Please initialize it first.")

if __name__ == "__main__":
    run_streamlit_app()