import gradio as gr import openai, os 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 dotenv import load_dotenv, find_dotenv _ = load_dotenv(find_dotenv()) #openai.api_key = os.environ["OPENAI_API_KEY"] 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 Straehle." 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" MODEL_NAME = "gpt-4" def invoke(openai_api_key, use_rag, prompt): try: llm = ChatOpenAI(model_name = MODEL_NAME, openai_api_key = openai_api_key, temperature = 0) if (use_rag): # Document loading #docs = [] # Load PDF #loader = PyPDFLoader(PDF_URL) #docs.extend(loader.load()) # Load Web #loader = WebBaseLoader(WEB_URL_1) #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) # Document storage #vector_db = Chroma.from_documents(documents = splits, # embedding = OpenAIEmbeddings(disallowed_special = ()), # persist_directory = CHROMA_DIR) # Document retrieval vector_db = Chroma(embedding_function = OpenAIEmbeddings(), persist_directory = CHROMA_DIR) rag_chain = RetrievalQA.from_chain_type(llm, chain_type_kwargs = {"prompt": RAG_CHAIN_PROMPT}, retriever = vector_db.as_retriever(search_kwargs = {"k": 3}), return_source_documents = True) result = rag_chain({"query": prompt}) result = result["result"] else: chain = LLMChain(llm = llm, prompt = LLM_CHAIN_PROMPT) result = chain.run({"question": prompt}) except Exception as e: raise gr.Error(e.Message) return result description = """Overview: Reasoning application that demonstrates a Large Language Model (LLM) with Retrieval Augmented Generation (RAG) on external data.\n\n Instructions: Enter an OpenAI API key and perform LLM use cases (semantic search, summarization, translation, etc.) on YouTube, PDF, and Web GPT-4 data (created after LLM training cutoff). \n\n Technology: Gradio UI using OpenAI API via AI-first LangChain toolkit with GPT-4 foundation model and AI-native Chroma embedding database. Speech-to-text via Whisper foundation model.""" gr.close_all() demo = gr.Interface(fn=invoke, inputs = [gr.Textbox(label = "OpenAI API Key", value = "sk-", lines = 1), gr.Radio([True, False], label="Retrieval Augmented Generation", value = False), 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()