app.py

import os
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 gradio as gr
import re
from threading import Thread

class MultiDocumentAgentSystem:
    def __init__(self, documents_dict, model, tokenizer, embeddings):
        self.model = model
        self.tokenizer = tokenizer
        self.embeddings = embeddings
        self.document_vectors = self.create_document_vectors(documents_dict)

    def create_document_vectors(self, documents_dict):
        document_vectors = {}
        for doc_name, content in documents_dict.items():
            vectors = self.embeddings.encode(content, convert_to_tensor=True)
            document_vectors[doc_name] = vectors
        return document_vectors

    def query(self, user_input):
        query_vector = self.embeddings.encode(user_input, convert_to_tensor=True)
        
        # Find the most similar document
        most_similar_doc = max(self.document_vectors.items(), 
                               key=lambda x: torch.cosine_similarity(query_vector, x[1], dim=0))
        
        # Generate response using the most similar document as context
        response = self.generate_response(user_input, most_similar_doc[0], most_similar_doc[1])
        return response

    def generate_response(self, query, doc_name, doc_vector):
        prompt = f"Based on the document '{doc_name}', answer the following question: {query}"
        input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(self.model.device)
        
        with torch.no_grad():
            output = self.model.generate(input_ids, max_length=150, num_return_sequences=1)
        
        response = self.tokenizer.decode(output[0], skip_special_tokens=True)
        return response

class DocumentRetrievalAndGeneration:
    def __init__(self, embedding_model_name, lm_model_id, data_folder):
        self.documents_dict = self.load_documents(data_folder)
        self.embeddings = SentenceTransformer(embedding_model_name)
        self.tokenizer, self.model = self.initialize_llm(lm_model_id)
        self.multi_doc_system = MultiDocumentAgentSystem(self.documents_dict, self.model, self.tokenizer, self.embeddings)

    def load_documents(self, folder_path):
        documents_dict = {}
        for file_name in os.listdir(folder_path):
            if file_name.endswith('.txt'):
                file_path = os.path.join(folder_path, file_name)
                with open(file_path, 'r', encoding='utf-8') as file:
                    content = file.read()
                    documents_dict[file_name[:-4]] = content
        return documents_dict

    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 generate_response_with_timeout(self, input_ids, max_new_tokens=1000):
        try:
            streamer = TextIteratorStreamer(self.tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
            generate_kwargs = dict(
                input_ids=input_ids,
                max_new_tokens=max_new_tokens,
                do_sample=True,
                top_p=1.0,
                top_k=20,
                temperature=0.8,
                repetition_penalty=1.2,
                eos_token_id=self.tokenizer.eos_token_id,
                streamer=streamer,
            )
            
            thread = Thread(target=self.model.generate, kwargs=generate_kwargs)
            thread.start()
            
            generated_text = ""
            for new_text in streamer:
                generated_text += new_text
            
            return generated_text
        except Exception as e:
            print(f"Error in generate_response_with_timeout: {str(e)}")
            return "Text generation process encountered an error"

    def query_and_generate_response(self, query):
        response = self.multi_doc_system.query(query)
        return str(response), ""

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

if __name__ == "__main__":
    embedding_model_name = 'sentence-transformers/all-MiniLM-L6-v2'
    lm_model_id = "facebook/opt-350m"  # You can change this to a different open-source model
    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"
        )

        interface.launch(debug=True)

    launch_interface()