import torch import os try: from llama_cpp import Llama except: if torch.cuda.is_available(): print("CUDA is available on this system.") os.system('CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python==0.1.78 --force-reinstall --upgrade --no-cache-dir --verbose') else: print("CUDA is not available on this system.") os.system('pip install llama-cpp-python==0.1.78') import gradio as gr from langchain.embeddings.openai import OpenAIEmbeddings from langchain.text_splitter import CharacterTextSplitter, RecursiveCharacterTextSplitter from langchain.vectorstores import DocArrayInMemorySearch from langchain.chains import RetrievalQA, ConversationalRetrievalChain from langchain.memory import ConversationBufferMemory from langchain.chat_models import ChatOpenAI from langchain.embeddings import HuggingFaceEmbeddings from langchain import HuggingFaceHub from langchain.llms import LlamaCpp from huggingface_hub import hf_hub_download from langchain.document_loaders import ( EverNoteLoader, TextLoader, UnstructuredEPubLoader, UnstructuredHTMLLoader, UnstructuredMarkdownLoader, UnstructuredODTLoader, UnstructuredPowerPointLoader, UnstructuredWordDocumentLoader, PyPDFLoader, ) import param from conversadocs.bones import DocChat from conversadocs.llm_chess import ChessGame My_hf_token = os.getenv("My_hf_token") dc = DocChat() cg = ChessGame(dc) ##### GRADIO CONFIG #### css=""" #col-container {max-width: 1500px; margin-left: auto; margin-right: auto;} """ title = """

Chat with Documents 📚 - Falcon, Llama-2 and OpenAI

Upload txt, pdf, doc, docx, enex, epub, html, md, odt, ptt and pttx. Wait for the Status to show Loaded documents, start typing your questions. This DEMO uses Falcon 7B, so the answers may not be optimal. You can use the Colab with GPU and Llama2 to have high-quality responses. Oficial Repository ConversaDocs.

""" description = """ # Application Information - Notebook for run ConversaDocs in Colab [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/R3gm/ConversaDocs/blob/main/ConversaDocs_Colab.ipynb) - Oficial Repository [![a](https://img.shields.io/badge/GitHub-Repository-black?style=flat-square&logo=github)](https://github.com/R3gm/ConversaDocs/) - You can upload multiple documents at once to a single database. - Every time a new database is created, the previous one is deleted. - For maximum privacy, you can click "Load LLAMA GGML Model" to use a Llama 2 model. By default, the model llama-2_7B-Chat is loaded. - This application works on both CPU and GPU. For fast inference with GGML models, use the GPU. - For more information about what GGML models are, you can visit this notebook [![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/R3gm/InsightSolver-Colab/blob/main/LLM_Inference_with_llama_cpp_python__Llama_2_13b_chat.ipynb) ## 📖 News 🔥 2023/07/24: Document summarization was added. 🔥 2023/07/29: Error with llama 70B was fixed. 🔥 2023/08/07: ♟️ Chessboard was added for playing with a LLM. """ theme='aliabid94/new-theme' def flag(): return "PROCESSING..." def upload_file(files, max_docs): file_paths = [file.name for file in files] return dc.call_load_db(file_paths, max_docs) def predict(message, chat_history, max_k, check_memory): print(message) bot_message = dc.convchain(message, max_k, check_memory) print(bot_message) return "", dc.get_chats() def convert(): docs = dc.get_sources() data_docs = "" for i in range(0,len(docs),2): txt = docs[i][1].replace("\n","
") sc = "Archive: " + docs[i+1][1]["source"] try: pg = "Page: " + str(docs[i+1][1]["page"]) except: pg = "Document Data" data_docs += f"

{pg}

{txt}

{sc}

" return data_docs def clear_api_key(api_key): return 'api_key...', dc.openai_model(api_key) # Max values in generation DOC_DB_LIMIT = 5 MAX_NEW_TOKENS = 2048 # Limit in HF, no need to set it if "SET_LIMIT" == os.getenv("DEMO"): DOC_DB_LIMIT = 4 MAX_NEW_TOKENS = 32 with gr.Blocks(theme=theme, css=css) as demo: with gr.Tab("Chat"): with gr.Column(): gr.HTML(title) upload_button = gr.UploadButton("Click to Upload Files", file_count="multiple") file_output = gr.HTML() chatbot = gr.Chatbot([], elem_id="chatbot") #.style(height=300) msg = gr.Textbox(label="Question", placeholder="Type your question and hit Enter ") with gr.Row(): check_memory = gr.inputs.Checkbox(label="Remember previous messages") clear_button = gr.Button("CLEAR CHAT HISTORY", ) max_docs = gr.inputs.Slider(1, DOC_DB_LIMIT, default=3, label="Maximum querys to the DB.", step=1) with gr.Column(): link_output = gr.HTML("") sou = gr.HTML("") clear_button.click(flag,[],[link_output]).then(dc.clr_history,[], [link_output]).then(lambda: None, None, chatbot, queue=False) upload_button.upload(flag,[],[file_output]).then(upload_file, [upload_button, max_docs], file_output).then(dc.clr_history,[], [link_output]) with gr.Tab("Experimental Summarization"): default_model = gr.HTML("
From DB
It may take approximately 5 minutes to complete 15 pages in GPU. Please use files with fewer pages if you want to use summarization.
") summarize_button = gr.Button("Start summarization") summarize_verify = gr.HTML(" ") summarize_button.click(dc.summarize, [], [summarize_verify]) with gr.Tab("♟️ Chess Game with a LLM"): with gr.Column(): gr.HTML('♟️ Click to start the Chessboard ♟️') start_chess = gr.Button("START GAME") board_chess = gr.HTML() info_chess = gr.HTML() input_chess = gr.Textbox(label="Type a valid move", placeholder="") start_chess.click(cg.start_game,[],[board_chess, info_chess]) input_chess.submit(cg.user_move,[input_chess],[board_chess, info_chess, input_chess]) with gr.Tab("Config llama-2 model"): gr.HTML("

Only models from the GGML library are accepted. To apply the new configurations, please reload the model.

") repo_ = gr.Textbox(label="Repository" ,value="TheBloke/Llama-2-7B-Chat-GGML") file_ = gr.Textbox(label="File name" ,value="llama-2-7b-chat.ggmlv3.q2_K.bin") max_tokens = gr.inputs.Slider(1, MAX_NEW_TOKENS, default=16, label="Max new tokens", step=1) temperature = gr.inputs.Slider(0.1, 1., default=0.2, label="Temperature", step=0.1) top_k = gr.inputs.Slider(0.01, 1., default=0.95, label="Top K", step=0.01) top_p = gr.inputs.Slider(0, 100, default=50, label="Top P", step=1) repeat_penalty = gr.inputs.Slider(0.1, 100., default=1.2, label="Repeat penalty", step=0.1) change_model_button = gr.Button("Load Llama GGML Model") model_verify_ggml = gr.HTML("Loaded model Llama-2") with gr.Tab("API Models"): default_model = gr.HTML("
Falcon Model") hf_key = gr.Textbox(label="HF TOKEN", value=My_hf_token, visible=False) falcon_button = gr.Button("Load FALCON 7B-Instruct") openai_gpt_model = gr.HTML("
OpenAI Model gpt-3.5-turbo") api_key = gr.Textbox(label="API KEY", value="api_key...") openai_button = gr.Button("Load gpt-3.5-turbo") line_ = gr.HTML("
") model_verify = gr.HTML(" ") with gr.Tab("Help"): description_md = gr.Markdown(description) msg.submit(predict,[msg, chatbot, max_docs, check_memory],[msg, chatbot]).then(convert,[],[sou]) change_model_button.click(dc.change_llm,[repo_, file_, max_tokens, temperature, top_p, top_k, repeat_penalty, max_docs],[model_verify_ggml]) falcon_button.click(dc.default_falcon_model, [hf_key], [model_verify]) openai_button.click(clear_api_key, [api_key], [api_key, model_verify]) demo.launch(enable_queue=True)