netease-youdao/QAnything

What it solves

QAnything is a local knowledge base question-answering system that allows users to upload various file formats and receive accurate, fast answers based on that data. It specifically addresses the problem of retrieval degradation in large-scale datasets, ensuring that accuracy remains stable or improves as more data is added.

How it works

The system uses a two-stage retrieval process to maintain high accuracy across massive datasets:

1. First-stage retrieval (embedding): Uses BCEmbedding for bilingual and cross-lingual semantic representation to find relevant documents.
2. Second-stage retrieval (reranking): Refines the results to ensure the most relevant information is surfaced, preventing the performance drop typically seen in single-stage RAG systems.

It is based on QwenLM, fine-tuned on professional QA datasets, and supports integration with other LLMs via OpenAI-compatible APIs (including Ollama).

Who it’s for

It is designed for users and organizations that require high data security (supporting fully offline use), need to process diverse file types (PDF, Word, PPT, XLSX, Markdown, etc.), and want a hardware-friendly deployment that can run on CPU-only environments across Windows, Mac, and Linux.

Highlights

• High Data Security: Supports complete offline installation and use.
• Cross-Language Support: Enables switching between Chinese and English for both queries and source files.
• Two-Stage Retrieval: Utilizes embedding and reranking to ensure SOTA performance on large-scale knowledge bases.
• Broad Format Support: Parses PDF, Word, PPT, XLSX, Markdown, Email, TXT, Images, CSV, and Web links.
• Hardware Friendly: Defaults to CPU-only operation with one-click Docker Compose deployment.