app.py

import os
import multiprocessing
import concurrent.futures
from langchain.document_loaders import TextLoader, DirectoryLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import FAISS
from sentence_transformers import SentenceTransformer
import faiss
import torch
import numpy as np
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, BitsAndBytesConfig
from datetime import datetime
import json
import gradio as gr
import re
from threading import Thread
from transformers.agents import Tool, HfEngine, ReactJsonAgent
from huggingface_hub import InferenceClient
import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

class DocumentRetrievalAndGeneration:
    def __init__(self, embedding_model_name, lm_model_id, data_folder):
        self.all_splits = self.load_documents(data_folder)
        self.embeddings = SentenceTransformer(embedding_model_name)
        self.vectordb = self.create_faiss_index()
        self.tokenizer, self.model = self.initialize_llm(lm_model_id)
        self.retriever_tool = self.create_retriever_tool()
        self.agent = self.create_agent()

    def load_documents(self, folder_path):
        loader = DirectoryLoader(folder_path, loader_cls=TextLoader)
        documents = loader.load()
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=200, chunk_overlap=20)
        all_splits = text_splitter.split_documents(documents)
        logger.info(f'Loaded {len(documents)} documents')
        logger.info(f"Split into {len(all_splits)} chunks")
        return all_splits

    def create_faiss_index(self):
        all_texts = [split.page_content for split in self.all_splits]
        embeddings = self.embeddings.encode(all_texts, convert_to_tensor=True).cpu().numpy()
        vectordb = FAISS.from_embeddings(
            embeddings,
            self.embeddings,
            metadatas=[{"source": f"doc_{i}"} for i in range(len(all_texts))]
        )
        return vectordb

    def initialize_llm(self, model_id):
        quantization_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16
        )
        tokenizer = AutoTokenizer.from_pretrained(model_id)
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            torch_dtype=torch.bfloat16,
            device_map="auto",
            quantization_config=quantization_config
        )
        return tokenizer, model

    def create_retriever_tool(self):
        class RetrieverTool(Tool):
            name = "retriever"
            description = "Retrieves documents from the knowledge base that are semantically similar to the input query."
            inputs = {
                "query": {
                    "type": "text",
                    "description": "The query to perform. Use affirmative form rather than a question.",
                }
            }
            output_type = "text"

            def __init__(self, vectordb, **kwargs):
                super().__init__(**kwargs)
                self.vectordb = vectordb

            def forward(self, query: str) -> str:
                docs = self.vectordb.similarity_search(query, k=3)
                return "\nRetrieved documents:\n" + "".join(
                    [f"===== Document {str(i)} =====\n" + doc.page_content for i, doc in enumerate(docs)]
                )

        return RetrieverTool(self.vectordb)

    def create_agent(self):
        llm_engine = HfEngine("meta-llama/Meta-Llama-3.1-8B-Instruct")
        return ReactJsonAgent(tools=[self.retriever_tool], llm_engine=llm_engine, max_iterations=4, verbose=2)

    def run_agentic_rag(self, question: str) -> str:
        enhanced_question = f"""Using the information in your knowledge base, accessible with the 'retriever' tool,
give a comprehensive answer to the question below.
Respond only to the question asked, be concise and relevant.
If you can't find information, try calling your retriever again with different arguments.
Make sure to cover the question completely by calling the retriever tool several times with semantically different queries.
Your queries should be in affirmative form, not questions.

Question:
{question}"""

        return self.agent.run(enhanced_question)

    def run_standard_rag(self, question: str) -> str:
        context = self.retriever_tool(query=question)

        prompt = f"""Given the question and supporting documents below, give a comprehensive answer to the question.
Respond only to the question asked, be concise and relevant.
Provide the number of the source document when relevant.

Question:
{question}

{context}
"""
        messages = [{"role": "user", "content": prompt}]

        reader_llm = InferenceClient("meta-llama/Meta-Llama-3.1-8B-Instruct")

        return reader_llm.chat_completion(messages).choices[0].message.content

    def query_and_generate_response(self, query):
        agentic_answer = self.run_agentic_rag(query)
        standard_answer = self.run_standard_rag(query)
        
        combined_answer = f"Agentic RAG Answer:\n{agentic_answer}\n\nStandard RAG Answer:\n{standard_answer}"
        return combined_answer, ""  # Return empty string for 'content' as it's not used in this implementation

    def qa_infer_gradio(self, query):
        response = self.query_and_generate_response(query)
        return response

if __name__ == "__main__":
    embedding_model_name = 'thenlper/gte-small'
    lm_model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
    data_folder = 'sample_embedding_folder2'

    doc_retrieval_gen = DocumentRetrievalAndGeneration(embedding_model_name, lm_model_id, data_folder)

    def launch_interface():
        css_code = """
            .gradio-container {
                background-color: #daccdb;
            }
            button {
                background-color: #927fc7;
                color: black;
                border: 1px solid black;
                padding: 10px;
                margin-right: 10px;
                font-size: 16px;
                font-weight: bold;
            }
        """
        EXAMPLES = [
            "On which devices can the VIP and CSI2 modules operate simultaneously?", 
            "I'm using Code Composer Studio 5.4.0.00091 and enabled FPv4SPD16 floating point support for CortexM4 in TDA2. However, after building the project, the .asm file shows --float_support=vfplib instead of FPv4SPD16. Why is this happening?", 
            "Could you clarify the maximum number of cameras that can be connected simultaneously to the video input ports on the TDA2x SoC, considering it supports up to 10 multiplexed input ports and includes 3 dedicated video input modules?"
        ]

        interface = gr.Interface(
            fn=doc_retrieval_gen.qa_infer_gradio,
            inputs=[gr.Textbox(label="QUERY", placeholder="Enter your query here")],
            allow_flagging='never',
            examples=EXAMPLES,
            cache_examples=False,
            outputs=[gr.Textbox(label="RESPONSE"), gr.Textbox(label="RELATED QUERIES")],
            css=css_code,
            title="TI E2E FORUM Multi-Agent RAG"
        )

        interface.launch(debug=True)

    launch_interface()