English | 简体中文

HuixiangDou is a professional knowledge assistant based on LLM.

Advantages:

1. Design three-stage pipelines of preprocess, rejection and response
   • chat_in_group copes with group chat scenario, answer user questions without message flooding, see 2401.08772, 2405.02817, Hybrid Retrieval and Precision Report
   • chat_with_repo for real-time streaming chat
2. No training required, with CPU-only, 2G, 10G, 20G and 80G configuration
3. Offers a complete suite of Web, Android, and pipeline source code, industrial-grade and commercially viable

Check out the scenes in which HuixiangDou are running and join WeChat Group to try AI assistant inside.

If this helps you, please give it a star ⭐

🔆 New Features

Our Web version has been released to OpenXLab, where you can create knowledge base, update positive and negative examples, turn on web search, test chat, and integrate into Feishu/WeChat groups. See BiliBili and YouTube !

The Web version's API for Android also supports other devices. See Python sample code.

• [2024/09] Inverted indexer makes LLM prefer knowledge base🎯
• [2024/09] Code retrieval
• [2024/08] chat_with_readthedocs, see how to integrate 👍
• [2024/07] Image and text retrieval & Removal of langchain 👍
• [2024/07] Hybrid Knowledge Graph and Dense Retrieval improve 1.7% F1 score 🎯
• [2024/06] Evaluation of chunksize, splitter, and text2vec model 🎯
• [2024/05] wkteam WeChat access, parsing image & URL, support coreference resolution
• [2024/05] SFT LLM on NLP task, F1 increased by 29% 🎯

  🤗

  LoRA-Qwen1.5-14B

  LoRA-Qwen1.5-32B

  alpaca data

  arXiv

• [2024/04] RAG Annotation SFT Q&A Data and Examples
• [2024/04] Release Web Front and Back End Service Source Code 👍
• [2024/03] New Personal WeChat Integration and Prebuilt APK !
• [2024/02] [Experimental Feature] WeChat Group Integration of multimodal to achieve OCR

📖 Support Status

LLM	File Format	Retrieval Method	Integration	Preprocessing
InternLM2/InternLM2.5 Qwen1.5~2.5 puyu StepFun KIMI DeepSeek GLM (ZHIPU) SiliconCloud Xi-Api	pdf word excel ppt html markdown txt	Dense for Document Sparse for Code Knowledge Graph Internet Search SourceGraph Image and Text	WeChat(android/wkteam) Lark OpenXLab Web Gradio Demo HTTP Server Read the Docs	Coreference Resolution

📦 Hardware Requirements

The following are the GPU memory requirements for different features, the difference lies only in whether the options are turned on.

Configuration Example	GPU mem Requirements	Description	Verified on Linux
config-cpu.ini	-	Use siliconcloud API for text only
config-2G.ini	2GB	Use openai API (such as kimi, deepseek and stepfun to search for text only
config-multimodal.ini	10GB	Use openai API for LLM, image and text retrieval
[Standard Edition] config.ini	19GB	Local deployment of LLM, single modality
config-advanced.ini	80GB	local LLM, anaphora resolution, single modality, practical for WeChat group

🔥 Running the Standard Edition

We take the standard edition (local running LLM, text retrieval) as an introduction example. Other versions are just different in configuration options.

I. Download and install dependencies

Click to agree to the BCE model agreement, log in huggingface

huggingface-cli login

Install dependencies

# parsing `word` format requirements
apt update
apt install python-dev libxml2-dev libxslt1-dev antiword unrtf poppler-utils pstotext tesseract-ocr flac ffmpeg lame libmad0 libsox-fmt-mp3 sox libjpeg-dev swig libpulse-dev
# python requirements
pip install -r requirements.txt
# For python3.8, install faiss-gpu instead of faiss

II. Create knowledge base and ask questions

Use mmpose documents to build the mmpose knowledge base and filtering questions. If you have your own documents, just put them under repodir.

Copy and execute all the following commands (including the '#' symbol).

# Download the knowledge base, we only take the documents of mmpose as an example. You can put any of your own documents under `repodir`
cd HuixiangDou
mkdir repodir
git clone https://github.com/open-mmlab/mmpose    --depth=1 repodir/mmpose

# Save the features of repodir to workdir, and update the positive and negative example thresholds into `config.ini`
mkdir workdir
python3 -m huixiangdou.service.feature_store

After running, test with python3 -m huixiangdou.main --standalone. At this time, reply to mmpose related questions (related to the knowledge base), while not responding to weather questions.

python3 -m huixiangdou.main --standalone

+---------------------------+---------+----------------------------+-----------------+
|         Query             |  State  |         Reply              |   References    |
+===========================+=========+============================+=================+
| How to install mmpose?    | success | To install mmpose, plea..  | installation.md |
--------------------------------------------------------------------------------------
| How is the weather today? | unrelated.. | ..                     |                 |
+-----------------------+---------+--------------------------------+-----------------+
🔆 Input your question here, type `bye` for exit:
..

Note

If restarting LLM every time is too slow, first python3 -m huixiangdou.service.llm_server_hybrid; then open a new window, and each time only execute python3 -m huixiangdou.main without restarting LLM.

💡 Also run a simple Web UI with gradio:

python3 -m huixiangdou.gradio_ui

output.mp4

Or run a server to listen 23333, default pipeline is chat_with_repo:

python3 -m huixiangdou.server

# test async API 
curl -X POST http://127.0.0.1:23333/huixiangdou_stream  -H "Content-Type: application/json" -d '{"text": "how to install mmpose","image": ""}'
# cURL sync API
curl -X POST http://127.0.0.1:23333/huixiangdou_inference  -H "Content-Type: application/json" -d '{"text": "how to install mmpose","image": ""}'

Please update the repodir documents, good_questions and bad_questions, and try your own domain knowledge (medical, financial, power, etc.).

III. Integration into Feishu, WeChat group

• One-way sending to Feishu group
• Two-way Feishu group receiving and sending, recalling
• Personal WeChat Android access
• Personal WeChat wkteam access

IV. Deploy web front and back end

We provide typescript front-end and python back-end source code:

• Multi-tenant management supported
• Zero programming access to Feishu and WeChat
• k8s friendly

Same as OpenXlab APP, please read the web deployment document.

🍴 Other Configurations

CPU-only Edition

If there is no GPU available, model inference can be completed using the siliconcloud API.

Taking docker miniconda+Python3.11 as an example, install CPU dependencies and run:

# Start container
docker run -v /path/to/huixiangdou:/huixiangdou -p 7860:7860 -p 23333:23333 -it continuumio/miniconda3 /bin/bash
# Install dependencies
apt update
apt install python-dev libxml2-dev libxslt1-dev antiword unrtf poppler-utils pstotext tesseract-ocr flac ffmpeg lame libmad0 libsox-fmt-mp3 sox libjpeg-dev swig libpulse-dev
python3 -m pip install -r requirements-cpu.txt
# Establish knowledge base
python3 -m huixiangdou.service.feature_store --config_path config-cpu.ini
# Q&A test
python3 -m huixiangdou.main --standalone --config_path config-cpu.ini
# gradio UI
python3 -m huixiangdou.gradio_ui --config_path config-cpu.ini

If you find the installation too slow, a pre-installed image is provided in Docker Hub. Simply replace it when starting the docker.

2G Cost-effective Edition

If your GPU mem exceeds 1.8G, or you pursue cost-effectiveness. This configuration discards the local LLM and uses remote LLM instead, which is the same as the standard edition.

Take siliconcloud as an example, fill in the API TOKEN applied from the official website into config-2G.ini

# config-2G.ini
[llm]
enable_local = 0   # Turn off local LLM
enable_remote = 1  # Only use remote
..
remote_type = "siliconcloud"   # Choose siliconcloud
remote_api_key = "YOUR-API-KEY-HERE" # Your API key
remote_llm_model = "alibaba/Qwen1.5-110B-Chat"

Note

Each Q&A scenario requires calling the LLM 7 times at worst, subject to the free user RPM limit, you can modify the rpm parameter in config.ini

Execute the following to get the Q&A results

python3 -m huixiangdou.main --standalone --config-path config-2G.ini # Start all services at once

10G Multimodal Edition

If you have 10G GPU mem, you can further support image and text retrieval. Just modify the model used in config.ini.

# config-multimodal.ini
# !!! Download `https://huggingface.co/BAAI/bge-visualized/blob/main/Visualized_m3.pth`    to `bge-m3` folder !!!
embedding_model_path = "BAAI/bge-m3"
reranker_model_path = "BAAI/bge-reranker-v2-minicpm-layerwise"

Note:

• You need to manually download Visualized_m3.pth to the bge-m3 directory
• Install FlagEmbedding on main branch, we have made bugfix. Here you can download bpe_simple_vocab_16e6.txt.gz
• Install requirements/multimodal.txt

Run gradio to test, see the image and text retrieval result here.

python3 tests/test_query_gradio.py

80G Complete Edition

The "HuiXiangDou" in the WeChat experience group has enabled all features:

• Serper search and SourceGraph search enhancement
• Group chat images, WeChat public account parsing
• Text coreference resolution
• Hybrid LLM
• Knowledge base is related to openmmlab's 12 repositories (1700 documents), refusing small talk

Please read the following topics:

• Hybrid knowledge graph and dense retrieval
• Refer to config-advanced.ini configuration to improve effects
• Group chat scenario anaphora resolution training
• Use wkteam WeChat access, integrate images, public account parsing, and anaphora resolution
• Use rag.py to annotate SFT training data

Android Tools

Contributors have provided Android tools to interact with WeChat. The solution is based on system-level APIs, and in principle, it can control any UI (not limited to communication software).

🛠️ FAQ

1. What if the robot is too cold/too chatty?

   • Fill in the questions that should be answered in the real scenario into resource/good_questions.json, and fill the ones that should be rejected into resource/bad_questions.json.
   • Adjust the theme content in repodir to ensure that the markdown documents in the main library do not contain irrelevant content.

   Re-run feature_store to update thresholds and feature libraries.

   ⚠️ You can directly modify reject_throttle in config.ini. Generally speaking, 0.5 is a high value; 0.2 is too low.

2. Launch is normal, but out of memory during runtime?

   LLM long text based on transformers structure requires more memory. At this time, kv cache quantization needs to be done on the model, such as lmdeploy quantization description. Then use docker to independently deploy Hybrid LLM Service.

3. How to access other local LLM / After access, the effect is not ideal?

   • Open hybrid llm service, add a new LLM inference implementation.
   • Refer to test_intention_prompt and test data, adjust prompt and threshold for the new model, and update them into prompt.py.

4. What if the response is too slow/request always fails?

   • Refer to hybrid llm service to add exponential backoff and retransmission.
   • Replace local LLM with an inference framework such as lmdeploy, instead of the native huggingface/transformers.

5. What if the GPU memory is too low?

   At this time, it is impossible to run local LLM, and only remote LLM can be used in conjunction with text2vec to execute the pipeline. Please make sure that config.ini only uses remote LLM and turn off local LLM.

🍀 Acknowledgements

• KIMI: Long text LLM, supports direct file upload
• FlagEmbedding: BAAI RAG group
• BCEmbedding: Chinese-English bilingual feature model
• Langchain-ChatChat: Application of Langchain and ChatGLM
• GrabRedEnvelope: WeChat red packet grab

📝 Citation

@misc{kong2024huixiangdou,
      title={HuiXiangDou: Overcoming Group Chat Scenarios with LLM-based Technical Assistance},
      author={Huanjun Kong and Songyang Zhang and Jiaying Li and Min Xiao and Jun Xu and Kai Chen},
      year={2024},
      eprint={2401.08772},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{kong2024labelingsupervisedfinetuningdata,
      title={Labeling supervised fine-tuning data with the scaling law}, 
      author={Huanjun Kong},
      year={2024},
      eprint={2405.02817},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2405.02817}, 
}