SkillMesh

SkillMesh addresses the critical challenge of LLM agents becoming inefficient and costly when attempting to load an entire catalog of tools into their prompt. It solves this by implementing a sophisticated retrieval-based routing mechanism, which combines BM25 and dense retrieval with RRF fusion to intelligently select only the top-K most relevant "expert cards" for any given user query. These cards, encompassing detailed execution behavior, decision trees, anti-patterns, and composability hints, are then injected into the LLM's prompt, maintaining a small and accurate context while significantly reducing token costs. This approach empowers LLM agents to efficiently tackle complex, multi-domain tasks without performance degradation as the tool catalog expands.

主要功能

使用案例