Easily configure and deploy a fully self-hosted chatbot web service based on open source Large Language Models (LLMs), such as Mixtral or Llama 2, without the need for knowledge in machine learning.
- 🌐 Free and Open Source chatbot web service with UI and API.
- 🏡 Fully self-hosted, not tied to any service, and offline capable. Forget about API keys! Models and embeddings can be pre-downloaded, and the training and inference processes can run off-line if necessary.
- 🔌 Web API described using OpenAPI specs: GET/POST operations, websocket for streaming response
- 🪶 Chat web UI working well on desktop and mobile, with streaming response, and markdown rendering. Alternative gradio-based UI also available.
- 🚀 Easy to setup, no need to program, just configure the service with a YAML file, and start it with 1 command
- 📦 Available as a
pippackage 🐍, ordockerimage 🐳 - 🐌 No need for GPU, this will work even on your laptop CPU! That said, just running on CPUs can be quite slow (up to 1min to answer a documents-base question on recent laptops).
- 🦜 Powered by
LangChainandllama.cppto perform inference locally. - 🤖 Various types of agents can be deployed:
- 💬 Generic conversation: do not need any additional training, just configure settings such as the template prompt
- 📚 Documents-based question answering (experimental): automatically build similarity vectors from documents uploaded through the API UI, the chatbot will use them to answer your question, and return which documents were used to generate the answer (PDF, CSV, HTML, JSON, markdown, and more supported).
- 🔍 Readable logs to understand what is going on.
For more details on how to use Libre Chat check the documentation at vemonet.github.io/libre-chat
Warning
This project is a work in progress, use it with caution.
Those checkpoints are features we plan to work on in the future, feel free to let us know in the issues if you have any comment or request.
- Stream response to the websocket to show words as they are generated
- Add button to let the user stop the chatbot generation
- Add authentication mechanisms? (OAuth/OpenID Connect) #5
- Add conversational history? https://milvus.io/blog/conversational-memory-in-langchain.md
- Add an admin dashboard web UI to enable users to upload/inspect/delete documents for QA, see/edit the config of the chatbot.
- Kubernetes deployment (Helm chart?)
If you just want to quickly deploy it using the pre-trained model Mixtral-8x7B-Instruct, you can use docker:
docker run -it -p 8000:8000 ghcr.io/vemonet/libre-chat:mainYou can configure the deployment using environment variables. For this using a docker compose and a .env file is easier, first create the docker-compose.yml file:
version: "3"
services:
libre-chat:
image: ghcr.io/vemonet/libre-chat:main
volumes:
# ⚠️ Share folders from the current directory to the /data dir in the container
- ./chat.yml:/data/chat.yml
- ./models:/data/models
- ./documents:/data/documents
- ./embeddings:/data/embeddings
- ./vectorstore:/data/vectorstore
ports:
- 8000:8000And create a chat.yml file with your configuration in the same folder as the docker-compose.yml:
llm:
model_path: ./models/mixtral-8x7b-instruct-v0.1.Q2_K.gguf
model_download: https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF/resolve/main/mixtral-8x7b-instruct-v0.1.Q2_K.gguf
temperature: 0.01 # Config how creative, but also potentially wrong, the model can be. 0 is safe, 1 is adventurous
max_new_tokens: 1024 # Max number of words the LLM can generate
# Always use input for the human input variable with a generic agent
prompt_variables: [input, history]
prompt_template: |
Your are an assistant, please help me
{history}
User: {input}
AI Assistant:
vector:
vector_path: null # Path to the vectorstore to do QA retrieval, e.g. ./vectorstore/db_faiss
# Set to null to deploy a generic conversational agent
vector_download: null
embeddings_path: ./embeddings/all-MiniLM-L6-v2 # Path to embeddings used to generate the vectors, or use directly from HuggingFace: sentence-transformers/all-MiniLM-L6-v2
embeddings_download: https://public.ukp.informatik.tu-darmstadt.de/reimers/sentence-transformers/v0.2/all-MiniLM-L6-v2.zip
documents_path: ./documents # Path to documents to vectorize
chunk_size: 500 # Maximum size of chunks, in terms of number of characters
chunk_overlap: 50 # Overlap in characters between chunks
chain_type: stuff # Or: map_reduce, reduce, map_rerank. More details: https://docs.langchain.com/docs/components/chains/index_related_chains
search_type: similarity # Or: similarity_score_threshold, mmr. More details: https://python.langchain.com/docs/modules/data_connection/retrievers/vectorstore
return_sources_count: 2 # Number of sources to return when generating an answer
score_threshold: null # If using the similarity_score_threshold search type. Between 0 and 1
info:
title: "Libre Chat"
version: "0.1.0"
description: |
Open source and free chatbot powered by [LangChain](https://python.langchain.com) and [llama.cpp](https://github.com/ggerganov/llama.cpp)
examples:
- What is the capital of the Netherlands?
- Which drugs are approved by the FDA to mitigate Alzheimer symptoms?
- How can I create a logger with timestamp using python logging?
favicon: https://raw.github.com/vemonet/libre-chat/main/docs/docs/assets/logo.png
repository_url: https://github.com/vemonet/libre-chat
public_url: https://chat.semanticscience.org
contact:
name: Vincent Emonet
email: [email protected]
license_info:
name: MIT license
url: https://raw.github.com/vemonet/libre-chat/main/LICENSE.txtFinally start your chat service with:
docker compose upThis package requires Python >=3.8, simply install it with pipx or pip:
pip install libre-chatYou can easily start a new chat web service including UI and API using your terminal:
libre-chat startProvide a specific config file:
libre-chat start config/chat-vectorstore-qa.ymlFor re-build of the vectorstore:
libre-chat build --vector vectorstore/db_faiss --documents documentsGet a full rundown of the available options with:
libre-chat --helpOr you can use this package in python scripts:
import logging
import uvicorn
from libre_chat import ChatConf, ChatEndpoint, Llm
logging.basicConfig(level=logging.getLevelName("INFO"))
conf = ChatConf(
model_path="./models/mixtral-8x7b-instruct-v0.1.Q2_K.gguf",
vector_path=None
)
llm = Llm(conf=conf)
print(llm.query("What is the capital of the Netherlands?"))
# Create and deploy a FastAPI app based on your LLM
app = ChatEndpoint(llm=llm, conf=conf)
uvicorn.run(app)Inspired by:
