Tool Discovery — References

The enumeration mechanics protocols provide, and the retrieval research that applies once the enumerated surface is too large to show in full.

Enumerating what is available

• Model Context Protocol — listing operations and the change notifications that tell a consumer the advertised set has moved. Linked at the root; deep paths into the specification have been reorganised.
• OpenAPI Specification — a published interface description acting as a discovery source, including the operation summaries and descriptions that become retrieval text.
• RFC 8615: Well-Known Uniform Resource Identifiers — the convention for publishing a discovery document at a predictable path.
• RFC 9111: HTTP Caching — freshness and revalidation semantics, the established vocabulary for how stale an enumerated copy is allowed to be.

Retrieval when the surface is large

• ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs (Qin et al., 2023), arXiv:2307.16789 — explicit API retrieval as a separate stage before invocation, which is the structure this capability describes.
• Gorilla: Large Language Model Connected with Massive APIs (Patil et al., 2023), arXiv:2305.15334 — retriever-aware invocation against a large API corpus, and the effect of retrieval errors on downstream call accuracy.
• Dense Passage Retrieval for Open-Domain Question Answering (Karpukhin et al., 2020), arXiv:2004.04906 — the dual-encoder formulation that description-based tool retrieval is a small-document instance of.
• HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels (Gao et al., 2022), arXiv:2212.10496 — generating a hypothetical document to bridge a query-document vocabulary gap, the technique behind indexing generated intent text.

Trusting what advertises itself

• Not What You've Signed Up For (Greshake et al., 2023), on compromising real-world LLM-integrated applications through indirect prompt injection, arXiv:2302.12173 — why text supplied by a third party and placed in a model's context is an instruction channel, which applies to advertised descriptions as much as to results.

No verified reference is recorded for measuring candidate-set recall in agent systems, or for conventions that prevent description text from being written to win retrieval. Both are open practice problems.