[Feature Request] Add support for Adapt Search

[gxlover0625]

Thank you for your wonderful work, which provides an open-source solution for enabling collaboration between strong and weak models.

Beyond reasoning tasks, retrieval—is another crucial application scenario for large language models. However, not every user query actually requires retrieval. Passive approaches like RAG can introduce irrelevant documents, which may mislead the model, produce incorrect responses, and waste resources. This makes automatic retrieval a key capability for improving both efficiency and accuracy.

I hope that the semantic-router can support retrieval routing. In discussions with chinese developers from vLLM in little red book (china social media), we agreed that this represents a particularly important user case.

[rootfs]

Thanks for this idea, @gxlover0625!

In a quick note, I think the retrieval use case might have at least these supporting functionalities: function calling and model selection. Not all the LLM can do retrieval the same and once LLM comes to pick up the right tool with the retrieval results, the tool picking could also be obscured by model capability and prompt length.

Semantic Router can take small steps towards your use case by adding awareness of LLM's retrieval and function calling capabilities, then once the prompts are classified as retrieval, the router can find the best model with fine-tuned function call params to route the prompts.

[Xunzhuo]

/area research core
/priority P2
/milestone backlog