🔍 Read the full analysis: Why Use Tool Retrieval In The 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 lets developers defer full function parameter schemas until the model requests a tool, while keeping tool names and descriptions available. Google has not provided general savings figures; costs and performance effects depend on which schemas are loaded and how applications use the feature.
Google announced Tool Retrieval in preview for the Gemini Interactions API on Nov. 19, 2026, as described in the original analysis, giving developers a way to load full function parameter schemas only when a model requests a tool. The feature targets agents connected to large tool libraries, where sending every schema with each request can use input tokens even when most tools are not needed. Google says any savings depend on which schemas are loaded, and it has not published a general estimate, a consideration also relevant to Gemini API decisions.
With Tool Retrieval, developers add a tool_search tool and mark selected functions with defer_loading: true. The model can still see the deferred tools’ names and descriptions, but their full parameter schemas are withheld until the model asks for a particular tool. The API then makes that schema available so the model can issue the function call within the same request. The application handles the function and returns its result through the usual process.
Google says the sequence appears in the interaction steps as a tool search call, a search result containing the schema, and the resulting function call. Developers can keep frequently used functions loaded from the start and defer less common ones. The server-side feature does not require developers to build or host an index, according to Google. It also supports remote MCP servers, whose tools can be made available on demand.
Applications with their own tool registry, embedding index or user-permission filters can instead use client-side retrieval. In that setup, the application searches its own system and returns matching schemas; Google says this adds a round trip for each search. The company says Tool Retrieval is available through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models, and is compatible with prefix caching.
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.
How Deferred Schemas Could Change Costs
Tool libraries can become a source of prompt overhead when agents receive many function definitions even though a task may require only one. Deferring full schemas could keep unused parameter details out of the initial request, potentially helping teams that connect agents to dozens or hundreds of tools. Keeping names and descriptions visible gives the model information to identify a candidate before the API supplies its full schema.
The cost effect is not automatic. Google says there is no separate charge for the retrieval feature, but the search itself adds a small fixed block of input tokens. Each schema loaded is billed as input at the model’s standard rate for the rest of the conversation. Savings occur only when the schemas left unloaded outweigh those retrieved. A system that uses most of its tools often may save little, while client-side retrieval also brings an additional round trip.
For developers, the practical question is whether deferred loading improves total usage and request behavior for their own traffic. Google recommends measuring the feature against real workloads. The announcement does not include independent evaluations or figures for token savings, latency or tool-selection accuracy, so those outcomes should not be assumed from the feature description alone.
Products Worth Considering

Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD — Midnight

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

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)
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 those details are supplied: the model sees the deferred tool’s name and description first, then receives its full schema after identifying a function it may need. The function still runs through the application’s ordinary execution and result-return process.
Google presents the feature as a way to manage the growing context demands of agents with large tool collections, rather than as a replacement for function calling. Teams can choose which functions to defer, leaving common ones ready in advance. The company also says that adding schemas later in the interaction can preserve the same cached prompt prefix across turns, supporting prefix caching.
The announcement describes both server-side retrieval managed through the API and client-side retrieval managed by an application’s existing systems. That distinction matters for teams with custom search or permission controls: they can keep their own retrieval logic, but must account for the added search round trip. Google has invited developers to consult its setup documentation and test the preview with supported models.
“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 Still Unknown
Google has not supplied a general token-savings estimate or benchmark results for latency or tool-selection accuracy. The actual effect will depend on the size and usage pattern of a tool library, how many schemas are retrieved, and how requests are structured. The announcement recommends measuring costs against application traffic but does not provide independent results from real deployments.
The feature remains in preview, and Google has not announced a general-availability date. Its announcement does not detail all preview limitations, nor does it establish how accurately the model will find relevant tools across different library sizes. Whether server-side or client-side retrieval is preferable will also depend on each application’s infrastructure and permission requirements.
Teams Can Test Their Own Workloads
Developers can test Tool Retrieval with supported Gemini models through the Interactions API while the feature is in preview. The next useful step is to compare requests with and without deferred schemas using representative traffic, tracking input tokens, added search activity and latency, as well as whether the model selects appropriate functions.
Google has invited feedback from users, but it has not set out a timeline for general availability. Any decision to adopt the feature at scale will depend on preview behavior and workload-specific measurements; claims of savings should remain tied to those results rather than treated as guaranteed.
Key Questions
What does Tool Retrieval do?
It lets developers defer a function’s full parameter schema and make that schema available when the model requests the tool. The tool’s name and description remain visible beforehand.
Does Tool Retrieval guarantee lower API costs?
No. Google says retrieval has no separate charge, but the search adds a small fixed block of input tokens and loaded schemas are billed at the model’s standard input rate. The cost outcome depends on how many schemas are retrieved versus left unloaded.
Can developers use their own tool search system?
Yes. Google’s client-side approach lets an application search its own registry, index or permission-filtered system and return matching schemas. Google says this approach adds a round trip for each search.
Which Gemini models support Tool Retrieval?
Google says the preview works through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models.
When will Tool Retrieval be generally available?
Google has not announced a general-availability date. The feature is currently described as in preview.
Primary source: Google AI · via ThorstenMeyerAI.com
