Google launched Tool Retrieval in preview on November 19, letting the Gemini Interactions API load tool parameter schemas on demand. There is a detail to keep visible during an enterprise rollout: in server-side retrieval, deferred tools still have names and descriptions presented to the model. A tool that is never called can still contribute information to the conversation.
AI disclosure: This article was generated with AI assistance using Keith Azodeh’s public project documentation and the linked primary sources. It has not been reviewed by Keith before publication.
The practical question starts before the first search result: what does the tool catalog itself disclose?
The catalog is part of the data flow
Tool names often begin as implementation details. A developer writes something descriptive, adds a sentence explaining the function, and moves on to the harder work of validating arguments and returning records.
Once an assistant reads that catalog, those details become model input.
Consider a fictional archive tool named search_project_orchid_acquisition_notes. It might return nothing, reject every unauthorized request, and never be called in a particular conversation. Its name still describes a specific internal subject. Deferring its parameters would not remove that subject from the server-side tool list.
This is an invented example, not a tool or project from Keith’s work. The point is that exposure can happen through labels as well as through query results.
A capability-level name such as archive_search can be more appropriate, paired with a useful description of what it retrieves. Actual project names, customer names, and record examples do not need to become permanent catalog text.
That does not make the eventual function arguments or results public-safe. Those remain separate parts of the data flow to review.
Illumia provides a concrete boundary
The public Illumia Archive Intelligence case study describes six bounded, read-only Claude Desktop tools: archive information, filter lookup, search, case retrieval, work-record retrieval, and exact aggregates. It excludes arbitrary SQL, shell access, and filesystem writes.
The application retrieves locally, while selected evidence can be sent to online Claude when approved for the organization. The case study distinguishes synthetic-record validation from staged workplace acceptance.
That documented design makes an important decision explicit: which capabilities and evidence an assistant receives.
A proposed Gemini adaptation would need to examine another layer of that same decision: the metadata used to advertise those capabilities. The public documentation does not describe a Gemini integration, and this article does not claim one exists.
Keeping records local is only one fact about a connected workflow. Its tool definitions, arguments, and returned evidence need their own descriptions of where they go.
Compact descriptions still need useful meaning
Google’s technical guide gives more precise behavior than the launch headline alone. The deferred list uses a function’s short_description, or the first sentence of its description when that field is absent. Full parameter schemas load later.
That changes where a developer should look during a catalog review. An internal detail in the first sentence can be visible even when the detailed schema stays unloaded.
The answer is not to make every tool sound like “get data.” A catalog also needs enough meaning for the model to choose an appropriate operation.
For a proposed archive interface, a compact description might say, “Retrieve an existing service case by its stable identifier.” It communicates the object and the required kind of reference without embedding a real identifier or copying a record into the description.
Another might describe searching work notes and returning their parent-case references. These are illustrative descriptions, not quotations from Illumia’s implementation.
Reviewing metadata should preserve the distinctions that help the assistant work. Remove unnecessary instance-specific details; keep the capability understandable.
Choose the catalog before deferring it
For an application with different user scopes, a proposed sequence is to establish the signed-in user’s permitted capabilities, construct that user’s catalog, and then decide which schemas to defer. Execution-time authorization still checks the actual request.
Deferral answers when a schema enters model context. The application decides which capabilities belong in the conversation.
Google also documents client-side retrieval, where an application searches its own registry and returns matching declarations. That can support application-owned filtering, with an additional round trip for each search. It creates another interface to review rather than eliminating the need for one: the registry-search description and returned declarations are model input too.
Before adopting either pattern, inspect a synthetic conversation in which no archive tool runs. Record which tool names and compact descriptions are available at that point. Then compare two fictional user scopes and confirm that the catalog builder returns the intended capabilities for each.
Those are proposed acceptance checks, not tests already performed against Illumia or Gemini.
Make the rollout decision at the right level
Tool Retrieval addresses the cost and complexity of large tool libraries. A six-tool interface does not automatically need the same loading strategy as an assistant connected to hundreds of operations.
For enterprise teams, the useful release artifact is a reviewed catalog alongside the usual argument and result contracts. Someone should be able to read it and understand what the model learns before retrieving a single record.
Google’s preview makes schema loading more selective. Use that flexibility after deciding what the assistant should be told exists.
Top comments (0)