gnosis-mcp

概述

Gnosis MCP 是一个强大的文档搜索服务器，相比传统方法可将每次查找的 token 使用量减少 5-10 倍。它支持关键词(BM25)和本地 ONNX 嵌入的语义搜索，确保数据始终留在本地机器。服务器可索引多种文档格式，包括 markdown、git 历史记录和爬取的网站，并内置评估工具来衡量检索质量。

试试问 AI

装完之后，这里有 5 个你可以让 AI 做的事：

• search_docs

  Search through indexed documentation with keyword or hybrid search.

  试试:Search for information about authentication试试:Find documentation related to API setup试试:How do I configure the database connection?
• get_doc

  Retrieve the full content of a specific document by path.

  试试:Get the full content of the README file试试:Show me the complete documentation on embeddings试试:Retrieve the content of src/auth.py
• get_related

  Find documents related to a specific document based on relationships.

  试试:Find documents related to the authentication module试试:Show me documents that reference the database configuration试试:What's connected to the payment processing documentation?
• ingest

  Ingest local documentation files into the search index.

  试试:Ingest all files in the docs/ directory试试:Add new API documentation to the index试试:Ingest the project's source code documentation
• ingest_git

  Ingest git commit history as searchable documentation.

  试试:Ingest git history for the current repository试试:Add commit messages from the last 6 months试试:Ingest git history for the authentication module
• crawl

  Crawl and ingest documentation from a website.

  试试:Crawl the Stripe documentation website试试:Ingest tutorials from the FastAPI documentation试试:Fetch documentation from docs.python.org
• stats

  View statistics about the indexed documents.

  试试:Show me statistics about the indexed documentation试试:How many documents are in the index?试试:What's the total token count of all indexed documents?
• prune

  Remove stale documents from the index that no longer exist on disk.

  试试:Prune deleted documents from the index试试:Clean up the index after removing docs试试:Show me what would be pruned without actually deleting
• embed

  Generate embeddings for documents to enable semantic search.

  试试:Generate embeddings for all indexed documents试试:Create embeddings for the API documentation试试:Update embeddings with the latest model
• savings

  Calculate token savings from using search instead of full document retrieval.

  试试:Show me token savings from the last 7 days试试:How many tokens have I saved by using search?试试:Calculate token savings for the last month
• eval

  Evaluate the search performance with a test dataset.

  试试:Run evaluation on the current index试试:Test search performance with golden queries试试:Show me Hit@K and MRR metrics
• serve

  Start the MCP server to expose tools to AI clients.

  试试:Start the MCP server with default settings试试:Serve the documentation index on port 8000试试:Run the server with HTTP transport

可对比工具

安装

pip install gnosis-mcp           # 或：uv tool install gnosis-mcp
gnosis-mcp ingest ./docs/        # 将文档加载到 SQLite（自动创建）
gnosis-mcp serve                 # 启动 MCP 服务器

{
  "mcpServers": {
    "docs": {
      "command": "gnosis-mcp",
      "args": ["serve"]
    }
  }
}

FAQ

Gnosis MCP 支持哪些文档格式？

支持 .md、.txt、.ipynb、.toml、.csv、.json，以及可选的 .rst 和 .pdf 文件。

我可以不连接互联网使用 Gnosis MCP 吗？

是的，Gnosis MCP 默认完全离线工作，使用 SQLite，并使用本地 ONNX 嵌入进行语义搜索，无需 API 密钥。

gnosis-mcp 对比