import os import torch import logging from dotenv import load_dotenv from langchain.chains import create_retrieval_chain from langchain.chains.combine_documents import create_stuff_documents_chain from langchain_core.prompts import ChatPromptTemplate from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain_community.document_loaders import PyPDFLoader from langchain_community.vectorstores import Chroma from langchain_huggingface import ( HuggingFaceEndpoint, HuggingFaceEmbeddings, ChatHuggingFace, ) # Load environment variables load_dotenv() # Logging logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # Device DEVICE = "cuda" if torch.cuda.is_available() else "cpu" # Globals retrieval_chain = None chat_history = [] embeddings = None def init_llm(): global embeddings logger.info("Initializing embeddings...") embeddings = HuggingFaceEmbeddings( model_name="sentence-transformers/all-MiniLM-L6-v2", model_kwargs={"device": DEVICE}, ) logger.info("Embeddings initialized successfully.") def process_document(document_path): global retrieval_chain logger.info(f"Processing document: {document_path}") loader = PyPDFLoader(document_path) documents = loader.load() if not documents: raise ValueError("No text found in PDF (possibly scanned PDF).") text_splitter = RecursiveCharacterTextSplitter( chunk_size=1024, chunk_overlap=64, ) texts = text_splitter.split_documents(documents) db = Chroma.from_documents(texts, embedding=embeddings) retriever = db.as_retriever(search_kwargs={"k": 4}) # Hugging Face API token hf_token = os.getenv("HUGGINGFACEHUB_API_TOKEN") if not hf_token: raise ValueError("HUGGINGFACEHUB_API_TOKEN not found in environment variables") # Base LLM forcing the standard 'hf-inference' provider base_llm = HuggingFaceEndpoint( repo_id="meta-llama/Llama-3.1-8B-Instruct", task="text-generation", huggingfacehub_api_token=hf_token, temperature=0.1, max_new_tokens=600, provider="hf-inference", ) chat_model = ChatHuggingFace(llm=base_llm) system_prompt = ( "You are an AI assistant for document analysis. " "Use the following pieces of retrieved context to answer " "the user's question accurately. If you don't know the answer, " "say that you don't know.\n\n" "{context}" ) prompt = ChatPromptTemplate.from_messages([ ("system", system_prompt), ("human", "{input}"), ]) question_answer_chain = create_stuff_documents_chain(chat_model, prompt) retrieval_chain = create_retrieval_chain(retriever, question_answer_chain) logger.info("Document processed and LCEL retrieval chain ready.") def process_prompt(prompt): global chat_history, retrieval_chain if retrieval_chain is None: return "Please upload a PDF first." # Invoke modern LCEL chain response = retrieval_chain.invoke({"input": prompt}) answer = response.get("answer", "I couldn't generate a response.") chat_history.append((prompt, answer)) return answer.strip() # Initialize on startup init_llm()