Google Adds Tool Retrieval to Gemini API
The preview lets developers surface relevant tool descriptions only when needed, rather than sending every full schema with each request.
Updated on Nov. 19, 2026 in Artificial Intelligence

Google previewed Tool Retrieval in its Interactions API for Gemini, which surfaces relevant tool descriptions when needed instead of sending every full schema with each request. The feature is in preview.
Why it matters
Sending long lists of tool descriptions increases input tokens and processing time, and makes it harder for a model to select the right tool. Retrieval can miss a relevant tool, meaning it may not be available for the model to choose.
The announcement describes a qualitative tradeoff rather than measured results: fewer input tokens and faster processing versus the risk that retrieval omits a needed tool. No benchmark figures were included in the preview notice.
The players
Google is the American technology company behind the Gemini family of AI models and its developer API.
Gemini API
The Gemini API is Google's developer interface for building applications on its Gemini models, offering built-in tools and support for custom functions.
The details
Tool Retrieval selects which tool descriptions the model can see through a retrieval step, rather than including every function's schema in each request. Developers can already connect custom functions to the Gemini API, alongside built-in tools such as Google Search, File Search and Code Execution.
Timeline
Mid-2026: The Interactions API, which supports multi-turn workflows and server-side state management, reached general availability.
November 19, 2026: Google previewed Tool Retrieval in the Interactions API.
The Tech Race
Tool Retrieval extends the Interactions API, which reached general availability in mid-2026, into a more selective approach to tool use. It is part of Google's broader push to make its Gemini developer platform more efficient as AI agents rely on larger sets of connected tools.
Developers building AI applications may see lower input token use and faster responses if they adopt retrieval-based tool selection. The preview also means developers should verify that the tools they need are reliably retrieved before relying on the feature in production.
The takeaway
Developers building agents with many tools should test whether retrieval reliably surfaces the right tool before dropping full schemas from their requests. Trading fewer tokens for a possible missed tool is a tradeoff worth measuring in your own workloads.
Further reading
Explore more on Artificial Intelligence developments shaping how models use external tools.



