When ReAct Tool Selection Becomes a RAG

Draft — not listed anywhere. Reachable only at this URL.

Outline:

• The comfortable phase: a handful of tools, all schemas in context, the model picks correctly and nobody thinks about it.
• Where it breaks — context cost, latency, and accuracy all degrading as the tool count climbs. Name collisions and near-duplicate descriptions.
• The reframe: retrieving tools is the same problem as retrieving documents, so the same techniques apply. See Hybrid Search.
• Embedding tool descriptions vs. embedding example invocations — the question extraction trick, applied to capabilities.
• Hierarchies: namespaces, toolsets, and a router that picks the group before the tool.
• MCP in the picture — many servers, each contributing tools you didn't write.
• Keeping the exact-match path: when the user names the tool, don't let a vector search talk you out of it.
• Evaluating tool selection separately from evaluating the agent.