One tool, and the interesting part sits behind it: because the query goes through Higress, the same call can reach the open web, Arxiv and an internal knowledge base without the client needing to know which is which.
A single-tool server that sends a query to Higress and returns an answer grounded in what the search brought back. Which engines are in play is set by your Higress ai-search configuration: Google, Bing and Quark for the web, Arxiv for papers, and your own internal knowledge bases.
- A question answered against live web results rather than against what the model remembers
- Academic search over Arxiv, for questions where the paper is the answer
- Internal knowledge bases searched alongside the public web, once Higress is configured for them
- The search engines and the answering model both chosen in your gateway configuration, not baked into the server
A Higress instance with the ai-search and ai-proxy plugins configured and reachable from the machine running the server; the default address is a local one. You name the model the gateway should answer with, and describe your internal knowledge bases so the tool knows when they are worth reaching for. Installed with uv, or run from a local checkout.
One command — uvx higress-ai-search-mcp-server