import gradio as gr import openai, os, wandb from langchain.chains import LLMChain, RetrievalQA from langchain.chat_models import ChatOpenAI from langchain.document_loaders import PyPDFLoader, WebBaseLoader from langchain.document_loaders.blob_loaders.youtube_audio import YoutubeAudioLoader from langchain.document_loaders.generic import GenericLoader from langchain.document_loaders.parsers import OpenAIWhisperParser from langchain.embeddings.openai import OpenAIEmbeddings from langchain.prompts import PromptTemplate from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain.vectorstores import Chroma from langchain.vectorstores import MongoDBAtlasVectorSearch from pymongo import MongoClient from dotenv import load_dotenv, find_dotenv _ = load_dotenv(find_dotenv()) #openai.api_key = os.environ["OPENAI_API_KEY"] wandb_api_key = os.environ["WANDB_API_KEY"] MONGODB_URI = os.environ["MONGODB_ATLAS_CLUSTER_URI"] client = MongoClient(MONGODB_URI) MONGODB_DB_NAME = "langchain_db" MONGODB_COLLECTION_NAME = "gpt-4" MONGODB_COLLECTION = client[MONGODB_DB_NAME][MONGODB_COLLECTION_NAME] MONGODB_INDEX_NAME = "default" config = { "model": "gpt-4", "temperature": 0, } template = """If you don't know the answer, just say that you don't know, don't try to make up an answer. Keep the answer as concise as possible. Always say "🧠 Thanks for using the app - Bernd" at the end of the answer. """ llm_template = "Answer the question at the end. " + template + "Question: {question} Helpful Answer: " rag_template = "Use the following pieces of context to answer the question at the end. " + template + "{context} Question: {question} Helpful Answer: " LLM_CHAIN_PROMPT = PromptTemplate(input_variables = ["question"], template = llm_template) RAG_CHAIN_PROMPT = PromptTemplate(input_variables = ["context", "question"], template = rag_template) CHROMA_DIR = "/data/chroma" YOUTUBE_DIR = "/data/youtube" PDF_URL = "https://arxiv.org/pdf/2303.08774.pdf" WEB_URL = "https://openai.com/research/gpt-4" YOUTUBE_URL_1 = "https://www.youtube.com/watch?v=--khbXchTeE" YOUTUBE_URL_2 = "https://www.youtube.com/watch?v=hdhZwyf24mE" YOUTUBE_URL_3 = "https://www.youtube.com/watch?v=vw-KWfKwvTQ" def document_loading_splitting(): # Document loading docs = [] # Load PDF loader = PyPDFLoader(PDF_URL) docs.extend(loader.load()) # Load Web loader = WebBaseLoader(WEB_URL) docs.extend(loader.load()) # Load YouTube loader = GenericLoader(YoutubeAudioLoader([YOUTUBE_URL_1, YOUTUBE_URL_2, YOUTUBE_URL_3], YOUTUBE_DIR), OpenAIWhisperParser()) docs.extend(loader.load()) # Document splitting text_splitter = RecursiveCharacterTextSplitter(chunk_overlap = 150, chunk_size = 1500) splits = text_splitter.split_documents(docs) return splits def document_storage_chroma(splits): Chroma.from_documents(documents = splits, embedding = OpenAIEmbeddings(disallowed_special = ()), persist_directory = CHROMA_DIR) def document_storage_mongodb(splits): MongoDBAtlasVectorSearch.from_documents(documents = splits, embedding = OpenAIEmbeddings(disallowed_special = ()), collection = MONGODB_COLLECTION, index_name = MONGODB_INDEX_NAME) def document_retrieval_chroma(llm, prompt): db = Chroma(embedding_function = OpenAIEmbeddings(), persist_directory = CHROMA_DIR) return db def document_retrieval_mongodb(llm, prompt): db = MongoDBAtlasVectorSearch.from_connection_string(MONGODB_URI, MONGODB_DB_NAME + "." + MONGODB_COLLECTION_NAME, OpenAIEmbeddings(disallowed_special = ()), index_name = MONGODB_INDEX_NAME) return db def llm_chain(llm, prompt): llm_chain = LLMChain(llm = llm, prompt = LLM_CHAIN_PROMPT) completion = llm_chain.run({"question": prompt}) return completion def rag_chain(llm, prompt, db): rag_chain = RetrievalQA.from_chain_type(llm, chain_type_kwargs = {"prompt": RAG_CHAIN_PROMPT}, retriever = db.as_retriever(search_kwargs = {"k": 3}), return_source_documents = True) completion = rag_chain({"query": prompt}) return completion["result"] def evaluate(prompt, completion, rag_option): wandb.login(key = wandb_api_key) wandb.config["model"] = config.model wandb.config["temperature"] = config.temperature wandb.config["rag_option"] = rag_option wandb.config["prompt"] = prompt wandb.config["completion"] = completion wandb.init(project = "openai-llm-rag", config = config) #config = wandb.config wandb.log({"prompt": prompt, "completion": completion}) wandb.logout() def invoke(openai_api_key, rag_option, prompt): if (openai_api_key == ""): raise gr.Error("OpenAI API Key is required.") if (rag_option is None): raise gr.Error("Retrieval Augmented Generation is required.") if (prompt == ""): raise gr.Error("Prompt is required.") try: llm = ChatOpenAI(model_name = config.model, openai_api_key = openai_api_key, temperature = config.temperature) if (rag_option == "Chroma"): #splits = document_loading_splitting() #document_storage_chroma(splits) db = document_retrieval_chroma(llm, prompt) completion = rag_chain(llm, prompt, db) elif (rag_option == "MongoDB"): #splits = document_loading_splitting() #document_storage_mongodb(splits) db = document_retrieval_mongodb(llm, prompt) completion = rag_chain(llm, prompt, db) else: completion = llm_chain(llm, prompt) except Exception as e: raise gr.Error(e) evaluate(prompt, completion, rag_option) return completion description = """Overview: Context-aware multimodal reasoning application that demonstrates a large language model (LLM) with retrieval augmented generation (RAG). See architecture diagram on GitHub.\n\n Instructions: Enter an OpenAI API key and perform text generation use cases on YouTube, PDF, and web data published after LLM knowledge cutoff (example: GPT-4 data). \n\n Technology: Gradio UI using the OpenAI API and AI-native Chroma embedding database or MongoDB vector search. Speech-to-text via whisper-1 model, text embedding via text-embedding-ada-002 model, and text generation via gpt-4 model. Implementation via AI-first LangChain toolkit. Performance evaluation via Weights & Biases.""" gr.close_all() demo = gr.Interface(fn=invoke, inputs = [gr.Textbox(label = "OpenAI API Key", value = "sk-", lines = 1), gr.Radio(["Off", "Chroma", "MongoDB"], label="Retrieval Augmented Generation", value = "Off"), gr.Textbox(label = "Prompt", value = "What is GPT-4?", lines = 1)], outputs = [gr.Textbox(label = "Completion", lines = 1)], title = "Generative AI - LLM & RAG", description = description) demo.launch()