THE DIRECT ANSWER

Google introduced Tool Retrieval on November 19, 2026. The preview lets Gemini load tool parameter schemas on demand instead of keeping every schema in context. It is aimed at larger agent tool libraries, with standard token billing and a choice of server-side or application-managed retrieval.

Read the dated primary source ↓

A smaller active toolbox for an agent

Google introduced Tool Retrieval in the Gemini Interactions API on November 19, 2026. The preview targets agents with large collections of functions: rather than placing every parameter schema in the model’s context, it can load those details when needed. Google lists support for Gemini 3.6 and later models, including the Gemini 4 family.

A schema is the specification that tells a model which arguments a tool accepts. In a business agent with separate functions for orders, invoices, inventory and support tickets, those specifications can become substantial before the user has asked a question.

The practical appeal is straightforward: keep less unused material in the active prompt. That could matter most when an application has many capabilities but each task uses only a few.

Public record: Google — Tool Retrieval launch, November 19, 2026

The preview has a specific API boundary

The documentation requires the v1beta API version and warns that the preview interface may change. Tool Retrieval is available through the Interactions API; generateContent does not support it.

Google previously made Interactions its primary interface for models and agents, while saying the legacy generateContent API remained supported. A team using that older endpoint should therefore treat adoption as an API integration decision, with its own implementation work.

Public record: Google AI Studio — Tool retrieval documentation · Google — Interactions API general availability and legacy API support

Choose who performs the lookup

Server-side setup adds tool_search and marks selected functions defer_loading: true. Names and descriptions remain visible; schemas load within the API call. Your application still executes its own function calls.

Client-side retrieval lets your application search its registry or filter tools per user, adding a round trip.

For product teams, the choice is also about ownership. If different customers can access different integrations, decide which layer supplies that customer’s available tools. Retrieval should fit the application’s permission model; discovering a function is not, by itself, permission to perform its action.

Public record: Google AI Studio — Tool retrieval documentation

No separate retrieval fee does not mean no token cost

Google says Tool Retrieval has no separate charge. Normal model input and output rates apply. The search tool adds a small fixed input-token overhead, and loaded schemas are billed as input for the rest of the conversation.

That means savings depend on how much stays unloaded. The launch recommends measuring actual traffic and keeping frequently used tools available upfront.

Consider a long support conversation that eventually uses most of the available functions. Its economics could differ from a short request that needs one rarely used function. Compare the total interaction, including follow-up turns and retries, rather than only the first prompt.

Public record: Google — Tool Retrieval launch, November 19, 2026

Two details to check before a trial

The documentation says to resend the complete tools array on each request, including when chaining turns with previous_interaction_id. Deferring context loading does not remove that request requirement.

In stateless workflows, preserve the returned tool_search_call and tool_search_result steps unchanged, including their signatures. Dropping that retrieval context can break the conversation flow.

A useful first trial would replay a representative set of tasks against the same model, once with the existing tool setup and once with retrieval. Include similar-sounding functions, a request with no matching tool and a follow-up that needs a different function. Record whether the right function was chosen, whether its arguments were valid, total tokens and time to an acceptable result.

Clear names and descriptions deserve review before that comparison. A tool library with overlapping labels gives retrieval little help distinguishing actions. The release offers a new way to manage context; whether it improves the product still has to be established on the work that product handles. Promptwire has not independently benchmarked the preview.

Public record: Google AI Studio — Tool retrieval documentation

COMMON QUESTIONS

Questions readers ask

Can an existing generateContent integration turn this on directly?

No. The preview requires the Interactions API with v1beta.

SOURCES & NOTES

Follow the evidence.

Primary sources checked November 20, 2026. Availability, mechanics and billing are attributed to Google. Promptwire has not independently tested the preview or measured its savings. Trial design and product implications are editorial analysis. Cover: AI-assisted conceptual illustration, not a product screenshot.

  1. Google — Tool Retrieval launch, November 19, 2026
  2. Google AI Studio — Tool retrieval documentation
  3. Google — Interactions API general availability and legacy API support

Public sources support the factual background. Analysis and illustrative exercises are identified in the text. How we assess AI · Editorial standards

Spot a factual error or an outdated detail?