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 llama_index.core import VectorStoreIndex, Document from llama_index.core.tools import QueryEngineTool, ToolMetadata from llama_index.agent.openai import OpenAIAgent class Agent: def __init__(self, name, role, doc_retrieval_gen, tokenizer): self.name = name self.role = role self.doc_retrieval_gen = doc_retrieval_gen self.tokenizer = tokenizer def generate_response(self, query, context): if self.role == "Information Retrieval": return self.retriever_logic(query, context) elif self.role == "Content Analysis": return self.analyzer_logic(query, context) elif self.role == "Response Generation": return self.generator_logic(query, context) elif self.role == "Task Coordination": return self.coordinator_logic(query, context) def retriever_logic(self, query, all_splits): query_embedding = self.doc_retrieval_gen.embeddings.encode(query, convert_to_tensor=True).cpu().numpy() distances, indices = self.doc_retrieval_gen.gpu_index.search(np.array([query_embedding]), k=3) relevant_docs = [all_splits[i] for i in indices[0] if distances[0][i] <= 1] return relevant_docs def analyzer_logic(self, query, relevant_docs): analysis_prompt = f"Analyze the following documents in relation to the query: '{query}'\n\nDocuments:\n" for doc in relevant_docs: analysis_prompt += f"- {doc.page_content}\n" analysis_prompt += "\nProvide a concise analysis of the key points relevant to the query." input_ids = self.tokenizer.encode(analysis_prompt, return_tensors="pt").to(self.doc_retrieval_gen.model.device) analysis = self.doc_retrieval_gen.model.generate(input_ids, max_length=200, num_return_sequences=1) return self.tokenizer.decode(analysis[0], skip_special_tokens=True) def generator_logic(self, query, analyzed_content): generation_prompt = f"Based on the following analysis, generate a comprehensive answer to the query: '{query}'\n\nAnalysis:\n{analyzed_content}\n\nGenerate a detailed response:" input_ids = self.tokenizer.encode(generation_prompt, return_tensors="pt").to(self.doc_retrieval_gen.model.device) response = self.doc_retrieval_gen.model.generate(input_ids, max_length=300, num_return_sequences=1) return self.tokenizer.decode(response[0], skip_special_tokens=True) def coordinator_logic(self, query, final_response): coordination_prompt = f"As a coordinator, review and refine the following response to the query: '{query}'\n\nResponse:\n{final_response}\n\nProvide a final, polished answer:" input_ids = self.tokenizer.encode(coordination_prompt, return_tensors="pt").to(self.doc_retrieval_gen.model.device) coordinated_response = self.doc_retrieval_gen.model.generate(input_ids, max_length=350, num_return_sequences=1) return self.tokenizer.decode(coordinated_response[0], skip_special_tokens=True) class MultiDocumentAgentSystem: def __init__(self, documents_dict, llm, embed_model): self.llm = llm self.embed_model = embed_model self.document_agents = {} self.create_document_agents(documents_dict) self.top_agent = self.create_top_agent() def create_document_agents(self, documents_dict): for doc_name, doc_content in documents_dict.items(): vector_index = VectorStoreIndex.from_documents([Document(doc_content)]) summary_index = VectorStoreIndex.from_documents([Document(doc_content)]) vector_query_engine = vector_index.as_query_engine(similarity_top_k=2) summary_query_engine = summary_index.as_query_engine() query_engine_tools = [ QueryEngineTool( query_engine=vector_query_engine, metadata=ToolMetadata( name=f"vector_tool_{doc_name}", description=f"Useful for specific questions about {doc_name}", ), ), QueryEngineTool( query_engine=summary_query_engine, metadata=ToolMetadata( name=f"summary_tool_{doc_name}", description=f"Useful for summarizing content about {doc_name}", ), ), ] self.document_agents[doc_name] = OpenAIAgent.from_tools( query_engine_tools, llm=self.llm, verbose=True, system_prompt=f"You are an agent designed to answer queries about {doc_name}.", ) def create_top_agent(self): all_tools = [] for doc_name, agent in self.document_agents.items(): doc_tool = QueryEngineTool( query_engine=agent, metadata=ToolMetadata( name=f"tool_{doc_name}", description=f"Use this tool for questions about {doc_name}", ), ) all_tools.append(doc_tool) obj_index = VectorStoreIndex.from_objects(all_tools, embed_model=self.embed_model) return OpenAIAgent.from_tools( all_tools, llm=self.llm, verbose=True, system_prompt="You are an agent designed to answer queries about multiple documents.", tool_retriever=obj_index.as_retriever(similarity_top_k=3), ) def query(self, user_input): return self.top_agent.chat(user_input) 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.gpu_index = self.create_faiss_index() self.tokenizer, self.model = self.initialize_llm(lm_model_id) self.agents = self.initialize_agents() documents_dict = self.load_documents(data_folder) self.multi_doc_system = MultiDocumentAgentSystem(documents_dict, self.model, self.embeddings) def initialize_agents(self): agents = [ Agent("Retriever", "Information Retrieval", self, self.tokenizer), Agent("Analyzer", "Content Analysis", self, self.tokenizer), Agent("Generator", "Response Generation", self, self.tokenizer), Agent("Coordinator", "Task Coordination", self, self.tokenizer) ] return agents 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 # Use filename without .txt as key return documents_dict 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 coordinate_agents(self, query): coordinator = next(agent for agent in self.agents if agent.name == "Coordinator") # Step 1: Information Retrieval retriever = next(agent for agent in self.agents if agent.name == "Retriever") relevant_docs = retriever.generate_response(query, self.all_splits) # Step 2: Content Analysis analyzer = next(agent for agent in self.agents if agent.name == "Analyzer") analyzed_content = analyzer.generate_response(query, relevant_docs) # Step 3: Response Generation generator = next(agent for agent in self.agents if agent.name == "Generator") final_response = generator.generate_response(query, analyzed_content) # Step 4: Coordination and Refinement coordinated_response = coordinator.generate_response(query, final_response) return coordinated_response, "\n".join([doc.page_content for doc in relevant_docs]) def query_and_generate_response(self, query): response = self.multi_doc_system.query(query) return str(response), "" 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 query_and_generate_response(self, query): similarityThreshold = 1 query_embedding = self.embeddings.encode(query, convert_to_tensor=True).cpu().numpy() distances, indices = self.gpu_index.search(np.array([query_embedding]), k=3) print("Distance", distances, "indices", indices) content = "" filtered_results = [] for idx, distance in zip(indices[0], distances[0]): if distance <= similarityThreshold: filtered_results.append(idx) for i in filtered_results: print(self.all_splits[i].page_content) content += "-" * 50 + "\n" content += self.all_splits[idx].page_content + "\n" print("CHUNK", idx) print("Distance:", distance) print("indices:", indices) print(self.all_splits[idx].page_content) print("############################") 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 {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==> """} ] #Include a final answer without additional comments, sign-offs, or extra phrases. Be direct and to the point. input_ids = self.tokenizer.apply_chat_template(conversation, return_tensors="pt").to(self.model.device) start_time = datetime.now() generated_response = self.generate_response_with_timeout(input_ids) elapsed_time = datetime.now() - start_time print("Generated response:", generated_response) print("Time elapsed:", elapsed_time) print("Device in use:", self.model.device) solution_text = generated_response.strip() if "Solution:" in solution_text: solution_text = solution_text.split("Solution:", 1)[1].strip() # Post-processing to remove "assistant" prefix solution_text = re.sub(r'^assistant\s*', '', solution_text, flags=re.IGNORECASE) solution_text = solution_text.strip() return solution_text, content 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 = 'flax-sentence-embeddings/all_datasets_v3_MiniLM-L12' 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" ) interface.launch(debug=True) launch_interface()