Wrote a tutorial on loading agent tools on demand using local ONNX embeddings

[anktsrkr]

Had ~80 tools in an agent and the model kept picking the wrong ones or just ignoring them. turned out I was burning most of the context window just on tool definitions before any actual conversation.

Built a "RAG for tools" pattern to fix it — agent starts with one SearchTools function, calls it when it needs something, local embedding model finds the relevant tools and loads them into the session on the fly. Runs offline using all-MiniLM-L6-v2 via ONNX Runtime with FusionCache so there's no inference overhead on repeat queries.

Full writeup: https://microsoft-agent-framework.github.io/learn/agent-capabilities/semantic-tool-search/