Tool Routing
A mechanism in AI agent systems that dynamically selects and injects a relevant subset of tools based on user intent. Using intent classification or keyword matching, it reduces the number of tools present in the context, lowering token costs and improving model decision accuracy.
Timeline (last 90 days)
Tool routing is first using a lightweight model or rule engine to classify a user request into a certain domain, and then injecting only the tools relevant to that domain into the main call context, which can simultaneously improve accuracy and reduce latency and cost.
Tool routing itself can also go wrong, so when classification fails it should fall back to a conservative mode containing all tools rather than directly rejecting the request.
The tool routing mechanism first uses embedding vector retrieval or a classifier to pre-judge user intent, narrowing the candidate tool pool to 3–5 options, which are then passed to the LLM for precise selection, similar to the recall-ranking two-stage architecture of a search engine.
All Facts (3)
Tool routing is first using a lightweight model or rule engine to classify a user request into a certain domain, and then injecting only the tools relevant to that domain into the main call context, which can simultaneously improve accuracy and reduce latency and cost.
50%UnverifiedTool routing itself can also go wrong, so when classification fails it should fall back to a conservative mode containing all tools rather than directly rejecting the request.
50%UnverifiedThe tool routing mechanism first uses embedding vector retrieval or a classifier to pre-judge user intent, narrowing the candidate tool pool to 3–5 options, which are then passed to the LLM for precise selection, similar to the recall-ranking two-stage architecture of a search engine.
50%