🔍 Read the full analysis: Exploring Tool Retrieval In Google's Gemini API on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google announced Tool Retrieval in preview for the Gemini Interactions API on Nov. 19, 2026. It keeps tool names and descriptions available while loading full parameter schemas only when requested; Google says costs and potential savings depend on actual use.
Google announced Tool Retrieval in preview for its Gemini Interactions API on Nov. 19, 2026, giving developers a way to keep full function parameter schemas out of a request until a model asks to use a tool, as described in the original analysis. The change targets agents with large tool libraries, where sending every schema may add input tokens even when most tools are irrelevant to a particular task.
With the server-side setup described by Google, developers add a tool_search tool and mark functions for deferred loading with defer_loading: true, a pattern that complements Gemini’s decision-making tools. The model can still see each deferred tool’s name and description, but not its full parameter schema. When it requests a tool by name, the API supplies that schema so the model can make the function call.
Google says developers can inspect this sequence in the interaction steps: a tool search call, a result containing the requested schema, and then a function call, offering another view of Google’s Gemini 4 tooling. The application handles that function and returns its result as usual. Teams can leave commonly used tools loaded from the start and defer less frequently used ones. Google says server-side retrieval does not require developers to build or host an index.
The feature also supports tools exposed by remote MCP servers. Developers with their own registry, embedding index or user-permission filters can instead use client-side retrieval: their application searches its own system and returns relevant schemas. That option adds a round trip for each search, according to the source material. Google says the feature works with Gemini 3.6 and later, including all Gemini 4 models, through the Interactions API.
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 Retrieval changes how an agent can access a large collection of functions without placing every parameter definition in the initial request. For teams whose agents connect to dozens or hundreds of tools, deferring schemas may reduce the input context devoted to functions that a task never uses. It may also make it easier to keep a broad library available to an agent while loading details as needed.
Those benefits are a design goal, not a published performance result. Google says the retrieval feature has no separate charge, but the search adds a small fixed block of input tokens. Each schema loaded is then billed as input at the model’s standard rate for the rest of the conversation. The net cost depends on whether the schemas that remain unused outweigh the retrieval overhead and the schemas that are loaded.
That trade-off means teams should not assume that deferred loading will lower costs or latency in every case. A workload that uses many tools on most requests could gain little, while client-side retrieval also introduces an additional network step. Google advises developers to measure the feature against their own traffic; the announcement provides no general savings estimate.
Products Worth Considering

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
From Full Schemas to On-Demand Loading
In conventional function calling, a developer supplies tool definitions, including parameter schemas, alongside a request. Tool Retrieval separates that information: tool names and descriptions remain available for selection, while full schemas are supplied only after the model identifies a function it wants to call. The intended effect is to keep unused parameter details out of the initial context while preserving ordinary function execution.
Google says deferred schemas are compatible with prefix caching. Because those schemas enter later in the interaction rather than at the beginning of the prompt, Google says the cached prefix can remain unchanged across turns. The announcement also places the feature within the Interactions API’s support for tool use and managed agent workflows. The new mechanism is in preview, not described as a general-availability release.
“The cost outcome depends on usage.”
Savings and Preview Limits
Google has not provided benchmark results for token savings, latency or tool-selection accuracy, and no independent evaluation is included in the announcement. The degree of benefit is therefore unknown and may vary with tool-library size, how often functions are called, and how requests are structured. The source material says Google recommends measuring costs against actual traffic.
The feature remains in preview, and the announcement does not specify a general-availability date or detail all preview limitations. It also does not establish whether searching a larger tool library affects the accuracy or consistency of tool selection. Those points require testing and further information.
Testing Tool Retrieval in Preview
Developers can consult Google’s setup documentation and test Tool Retrieval with supported models through the Interactions API. Google has invited users to discuss the feature and share feedback. For teams evaluating it, the immediate step is to compare requests with and without deferred schemas under representative workloads, tracking total input tokens, latency, tool selection and any added client-side round trips.
Further evidence will depend on results from those tests and any updates Google provides as the preview progresses. Until Google publishes broader measurements or announces availability beyond preview, expected savings and effects on performance remain workload-specific.
Key Questions
What does Gemini API Tool Retrieval do?
It lets developers defer full function parameter schemas, keeping tool names and descriptions available while supplying a schema when the model requests that tool.
Does Tool Retrieval guarantee lower API costs?
No. Google says the search adds 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 a workload does not load.
Which Gemini models support the feature?
According to the source material, Tool Retrieval works through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models.
Is Tool Retrieval generally available?
No. Google announced it in preview on Nov. 19, 2026. The source material does not give a general-availability date.
Primary source: Google AI · via ThorstenMeyerAI.com
