🔍 Read the full analysis: Get Started With Tool Retrieval In The Gemini API on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Google announced Tool Retrieval in preview for the Gemini Interactions API on Nov. 19, 2026. Developers can keep tool names and descriptions visible while deferring full parameter schemas until a model requests a tool; Google has not published general savings figures.
Google announced Tool Retrieval in preview for the Gemini Interactions API on Nov. 19, 2026, giving developers a way to defer full function parameter schemas until a model needs them, as described in the original analysis. The feature is aimed at agents connected to large tool libraries, where including every schema in each request can add input tokens; Google has not published a general estimate of potential savings.
To use Google’s server-side approach, developers add a tool_search tool and set defer_loading: true for functions whose schemas should be held back. The model can still see a deferred function’s name and description, but not its full parameter schema. If it asks for that function, the API supplies the schema, after which the model can make the function call within the same request.
Google says the interaction’s steps show the tool search call, the result containing the requested schema, and the subsequent function call. The application handles that function and returns its result as it normally would. Developers may keep frequently used tools loaded while deferring less commonly used ones. Google says server-side retrieval does not require developers to build or host an index.
The feature also supports remote MCP servers, whose tools can be exposed on demand. Developers with their own tool registry, embedding index or user-permission filters can instead use client-side retrieval: their application searches its own system and returns matching schemas. That option adds a round trip for each search, according to the source material.
Introducing Tool Retrieval in the Gemini API
Google’s preview feature loads tool parameter schemas on demand, helping agents work with large tool libraries while keeping unused details out of the initial request.
What changes for agents
Tool names and descriptions remain visible; full parameter schemas load when requested.
Developers mark selected functions for deferred loading.
The model requests a tool by name and receives its schema.
The model calls it, and the application handles the result as usual.
Choose a retrieval approach
Server-side retrieval suits managed tool libraries; client-side retrieval supports custom search systems.
- Uses the tool_search tool.
- No index to build or host.
- Supports remote MCP server tools.
- Searches a custom registry or index.
- Can apply per-user permission filters.
- Adds a round trip for each search.
Understand the cost trade-off
Tool Retrieval has no separate charge; standard model token rates apply.
Skip unused schemas
Savings depend on whether schemas left unloaded outweigh schemas retrieved.
Search and loaded schemas
Search adds a small fixed block of input tokens; loaded schemas count as input for the rest of the conversation.
Managing Large Gemini Tool Libraries
Tool libraries can grow to dozens or hundreds of functions, while a given task may need only one. Sending every parameter schema with every request can use input tokens for tools that are not selected. Google’s approach separates the information needed to identify a potential tool—its name and description—from the fuller schema needed to call it.
Google says the feature is designed to reduce input token use, latency and difficulty selecting tools as libraries grow. Those are stated aims, not results established by published benchmarks. The cost calculation is workload-dependent: Google says there is no separate charge for the retrieval feature, but the search adds a small fixed block of input tokens, and schemas loaded are billed as input at the model’s standard rate for the rest of the conversation.
In practice, potential savings depend on how many schemas remain unused compared with how many are loaded. A system that frequently uses most of its tools may see less benefit than one with many rarely selected functions. Developers therefore need to compare total token use and request behavior against their own traffic rather than assume a particular reduction.
Products Worth Considering

Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, and IPX4 Water Resistance

Jackery Solar Generator 1000 V2 with 100W Solar Panel

DJI Mini 4K Drone Combo with 4K Camera, 3-Axis Gimbal, and Two Batteries
As an affiliate, we earn on qualifying purchases.
From Full Schemas to On-Demand Loading
In conventional function calling, developers provide tool definitions, including parameter schemas, with a request. Tool Retrieval changes when some of that information is supplied: the model initially sees names and descriptions, then requests a full schema if it identifies a function it may need. The function still runs through the application’s usual handling after the model calls it.
Google says deferred schemas are compatible with prefix caching. Because those schemas enter later rather than at the start of the prompt, the company says the cached prefix can remain the same across turns. The announcement says Tool Retrieval works through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models. These are Google’s stated compatibility details; the source does not include independent testing.
“Tool Retrieval lets your agent keep its full tool library on standby and load parameter schemas only when the current task needs them.”
Savings and Preview Limits Unreported
Google has not supplied a general figure for token savings, latency changes or tool-selection accuracy. The results may vary with the size of a tool library, how often each function is used and how requests are structured. The company advises developers to measure the effect on their own traffic.
The announcement also does not give a general-availability date, detail all preview limitations or provide independent evaluations of how accurately models find tools as libraries expand. The fixed token cost of searches and the additional round trip in client-side retrieval are relevant to testing, but the source does not quantify their effect across workloads.
Measure Results During Preview
Developers can consult Google’s setup documentation and test the feature with supported models in the Interactions API. The next practical step is to compare a representative workload with and without deferred schemas, tracking total input tokens, search behavior, latency and whether the model selects the appropriate function. That comparison should account for both schemas that are loaded and the search overhead.
Google has invited users to discuss the feature and share feedback. No later release date or benchmark publication was specified in the source material, so the timing of any move beyond preview remains unknown.
Key Questions
What does Tool Retrieval do?
It lets developers defer a function’s full parameter schema until a model requests that tool. The model can still see the tool’s name and description before then.
Does Tool Retrieval guarantee lower costs?
No. Google says the feature has no separate charge, but searches add a small fixed block of input tokens and loaded schemas are billed at the model’s standard input rate. Savings depend on how many schemas are loaded versus left unused.
Can developers use their own tool registry?
Yes. The client-side option lets an application search its own registry or index and return matching schemas. According to the source material, that approach adds a round trip for each search.
Which Gemini models support the preview?
Google’s announcement says Tool Retrieval works with Gemini 3.6 and later, including all Gemini 4 models, through the Interactions API.
When will Tool Retrieval leave preview?
The source material does not provide a general-availability date. It describes the feature as being in preview and says Google is seeking feedback.
Primary source: Google AI · via ThorstenMeyerAI.com
