I gave an LLM 248 tools and accuracy dropped to 12%. Here's what fixed it.

When exposing 248 Kubernetes API endpoints as tools to an LLM agent, accuracy dropped to 12% due to context overflow. Vector search failed because it finds individual tools but misses multi-step workflows. The author built graph-tool-call, a zero-dependency Python library that models tool relationships as a directed graph and uses weighted Reciprocal Rank Fusion to retrieve complete tool chains, restoring high accuracy and cutting token usage by 64–91%.