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 class MultiAgentRAG: 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.gpu_index = self.create_faiss_index() self.tokenizer, self.model = self.initialize_llm(lm_model_id) def load_documents(self, folder_path): loader = DirectoryLoader(folder_path, loader_cls=TextLoader) documents = loader.load() text_splitter = RecursiveCharacterTextSplitter(chunk_size=5000, chunk_overlap=250) all_splits = text_splitter.split_documents(documents) print('Length of documents:', len(documents)) print("LEN of all_splits", len(all_splits)) for i in range(3): print(all_splits[i].page_content) 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() index = faiss.IndexFlatL2(embeddings.shape[1]) index.add(embeddings) gpu_resource = faiss.StandardGpuResources() gpu_index = faiss.index_cpu_to_gpu(gpu_resource, 0, index) return gpu_index 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=[128001, 128008, 128009], 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 retrieval_agent(self, query): query_embedding = self.embeddings.encode(query, convert_to_tensor=True).cpu().numpy() distances, indices = self.gpu_index.search(np.array([query_embedding]), k=3) content = "" for idx, distance in zip(indices[0], distances[0]): content += "-" * 50 + "\n" content += self.all_splits[idx].page_content + "\n" return content def grading_agent(self, query, retrieved_content): grading_prompt = f""" Evaluate the relevance of the following retrieved content to the given query: Query: {query} Retrieved Content: {retrieved_content} Rate the relevance on a scale of 1-10 and explain your rating: """ input_ids = self.tokenizer.encode(grading_prompt, return_tensors="pt").to(self.model.device) grading_response = self.generate_response_with_timeout(input_ids) # Extract the numerical rating from the response rating = int(re.search(r'\d+', grading_response).group()) return rating, grading_response def query_rewrite_agent(self, original_query): rewrite_prompt = f""" The following query did not yield relevant results. Please rewrite it to potentially improve retrieval: Original Query: {original_query} Rewritten Query: """ input_ids = self.tokenizer.encode(rewrite_prompt, return_tensors="pt").to(self.model.device) rewritten_query = self.generate_response_with_timeout(input_ids) return rewritten_query.strip() def generation_agent(self, query, retrieved_content): conversation = [ {"role": "system", "content": "You are a knowledgeable assistant with access to a comprehensive database."}, {"role": "user", "content": f""" I need you to answer my question and provide related information in a specific format. I have provided five relatable json files {retrieved_content}, choose the most suitable chunks for answering the query. RETURN ONLY SOLUTION without additional comments, sign-offs, retrived chunks, refrence to any Ticket or extra phrases. Be direct and to the point. IF THERE IS NO ANSWER RELATABLE IN RETRIEVED CHUNKS, RETURN "NO SOLUTION AVAILABLE". DO NOT GIVE REFRENCE TO ANY CHUNKS OR TICKETS,BE ON POINT. Here's my question: Query: {query} Solution==> """} ] input_ids = self.tokenizer.encode(self.tokenizer.apply_chat_template(conversation, tokenize=False), return_tensors="pt").to(self.model.device) return self.generate_response_with_timeout(input_ids) def run_multi_agent_rag(self, query): max_iterations = 3 for i in range(max_iterations): # Retrieval step retrieved_content = self.retrieval_agent(query) # Grading step relevance_score, grading_explanation = self.grading_agent(query, retrieved_content) if relevance_score >= 7: # Assuming 7 out of 10 is the threshold for relevance # Generation step answer = self.generation_agent(query, retrieved_content) return answer, retrieved_content, grading_explanation else: # Query rewrite step query = self.query_rewrite_agent(query) return "Unable to find a relevant answer after multiple attempts.", "", "Low relevance across all attempts." def qa_infer_gradio(self, query): answer, retrieved_content, grading_explanation = self.run_multi_agent_rag(query) return answer, f"Retrieved Content:\n{retrieved_content}\n\nGrading Explanation:\n{grading_explanation}" def launch_interface(doc_retrieval_gen): 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) if __name__ == "__main__": embedding_model_name = 'flax-sentence-embeddings/all_datasets_v3_MiniLM-L12' lm_model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct" data_folder = 'sample_embedding_folder2' multi_agent_rag = MultiAgentRAG(embedding_model_name, lm_model_id, data_folder) launch_interface(multi_agent_rag)