Hugging Face's logo

netease-youdao
/
QAnything

Model card Files Community
QAnything / README.md
renzh's picture
renzh
Update README.md
330cbbd
preview code
|
raw
history blame
11.9 kB
metadata
license: apache-2.0
metrics:
  - accuracy

Question and Answer based on Anything

Table of Contents

GitHub

👉To QAnything GitHub

What is QAnything?

Question and Answer based on Anything (QAnything) is a local knowledge base question-answering system designed to support a wide range of file formats and databases, allowing for offline installation and use.

With QAnything, you can simply drop any locally stored file of any format and receive accurate, fast, and reliable answers.

Currently supported formats include: PDF, Word (doc/docx), PPT, Markdown, Eml, TXT, Images (jpg, png, etc.), Web links and more formats coming soon…

Key features

  • Data Security, supports installation and usage with network cable unplugged throughout the process.
  • Cross-language QA support, freely switch between Chinese and English QA, regardless of the language of the document.
  • Supports massive data QA, two-stage retrieval ranking, solving the degradation problem of large-scale data retrieval; the more data, the better the performance.
  • High-performance production-grade system, directly deployable for enterprise applications.
  • User-friendly, no need for cumbersome configurations, one-click installation and deployment, ready to use.
  • Multi knowledge base QA Support selecting multiple knowledge bases for Q&A

Architecture

qanything_system

Why 2 stage retrieval?

In scenarios with a large volume of knowledge base data, the advantages of a two-stage approach are very clear. If only a first-stage embedding retrieval is used, there will be a problem of retrieval degradation as the data volume increases, as indicated by the green line in the following graph. However, after the second-stage reranking, there can be a stable increase in accuracy, the more data, the better the performance.

two stage retrievaal

QAnything uses the retrieval component BCEmbedding, which is distinguished for its bilingual and crosslingual proficiency. BCEmbedding excels in bridging Chinese and English linguistic gaps, which achieves

1st Retrieval（embedding）

Model Retrieval STS PairClassification Classification Reranking Clustering Avg
bge-base-en-v1.5 37.14 55.06 75.45 59.73 43.05 37.74 47.20
bge-base-zh-v1.5 47.60 63.72 77.40 63.38 54.85 32.56 53.60
bge-large-en-v1.5 37.15 54.09 75.00 59.24 42.68 37.32 46.82
bge-large-zh-v1.5 47.54 64.73 79.14 64.19 55.88 33.26 54.21
jina-embeddings-v2-base-en 31.58 54.28 74.84 58.42 41.16 34.67 44.29
m3e-base 46.29 63.93 71.84 64.08 52.38 37.84 53.54
m3e-large 34.85 59.74 67.69 60.07 48.99 31.62 46.78
bce-embedding-base_v1 57.60 65.73 74.96 69.00 57.29 38.95 59.43

2nd Retrieval（rerank）

Model Reranking Avg
bge-reranker-base 57.78 57.78
bge-reranker-large 59.69 59.69
bce-reranker-base_v1 60.06 60.06

RAG Evaluations in LlamaIndex（embedding and rerank）

rag evaluation in llamaindex

NOTE:

  • In WithoutReranker setting, our bce-embedding-base_v1 outperforms all the other embedding models.
  • With fixing the embedding model, our bce-reranker-base_v1 achieves the best performence.
  • The combination of bce-embedding-base_v1 and bce-reranker-base_v1 is SOTA.
  • If you want to use embedding and rerank separately, please refer to BCEmbedding

LLM

The open source version of QAnything is based on QwenLM and has been fine-tuned on a large number of professional question-answering datasets. It greatly enhances the ability of question-answering. If you need to use it for commercial purposes, please follow the license of QwenLM. For more details, please refer to: QwenLM

Getting Started

👉try QAnything online

Prerequisites

Required item Minimum Requirement Note
NVIDIA GPU Memory >= 16GB NVIDIA 3090 recommended
NVIDIA Driver Version >= 525.105.17
CUDA Version >= 12.0
docker compose version >=1.27.4 docker compose install

Installation

step1: pull qanything repository 

git clone https://github.com/netease-youdao/QAnything.git

step2: download the model and unzip it to the root directory of the current project. 

cd QAnything
git lfs install
git clone https://www.modelscope.cn/netease-youdao/qanything_models.git

unzip qanything_models/models.zip   # in root directory of the current project

step3: change config

in the Windows system

vim docker-compose-windows.yaml # change CUDA_VISIBLE_DEVICES to your gpu device id

in the Linux system

vim docker-compose-linux.yaml # change CUDA_VISIBLE_DEVICES to your gpu device id

step4: start server

in the Windows system

docker-compose -f docker-compose-windows.yaml up -d

in the Linux system

docker-compose -f docker-compose-linux.yaml up -d

After successful installation, you can experience the application by entering the following addresses in your web browser.

  • Frontend address: http://{your_host}:5052/qanything

  • API address: http://{your_host}:5052/api/

For detailed API documentation, please refer to QAnything API 文档

Usage

Cross-lingual: Multiple English paper Q&A

Information extraction

Various files

Web Q&A

API Document

If you need to access the API, please refer to the QAnything API documentation.

WeChat Group

Welcome to scan the QR code below and join the WeChat group.

Support

Reach out to the maintainer at one of the following places:

License

QAnything is licensed under Apache 2.0 License

Acknowledgments

QAnything adopts dependencies from the following:

  • Thanks to our BCEmbedding for the excellent embedding and rerank model.
  • Thanks to Qwen for strong base language models.
  • Thanks to Triton Inference Server for providing great open source inference serving.
  • Thanks to FasterTransformer for highly optimized LLM inference backend.
  • Thanks to Langchain for the wonderful llm application framework.
  • Thanks to Langchain-Chatchat for the inspiration provided on local knowledge base Q&A.
  • Thanks to Milvus for the excellent semantic search library.
  • Thanks to PaddleOCR for its ease-to-use OCR library.
  • Thanks to Sanic for the powerful web service framework.