🔍 Read the full analysis: What Is 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. The feature keeps tool names and descriptions available while loading full parameter schemas only when requested; Google says cost and performance effects depend on actual usage.
Google announced Tool Retrieval in preview for the Gemini Interactions API on Nov. 19, 2026, as detailed in the original analysis, giving developers a way to hold back function parameter schemas until a model requests a tool. The approach is intended for agents with large tool libraries, where sending every schema in each request can add input tokens; Google has not published a general savings estimate.
With the server-side setup, developers add a tool_search tool and mark functions for deferred loading with defer_loading: true. The model can initially see a deferred function’s name and description, but not its full parameter schema. If it requests that function, the API returns the schema, after which the model can make a function call in the same request. The application handles the call and returns its result as it would with ordinary function calling.
Google says the interaction steps expose this sequence: a tool search call, a search result containing the schema, and then the function call. Developers can keep frequently used tools loaded while deferring less common ones. The company also says server-side retrieval does not require developers to build or host an index, and deferred loading is supported for tools exposed through remote MCP servers.
Applications with their own registry, embedding index or permission filters can instead use client-side retrieval. In that setup, the application searches its own system and returns relevant schemas, but the search adds a round trip. Google says Tool Retrieval has no separate feature charge: the search adds a small fixed block of input tokens, while each loaded schema is billed as input at the model’s standard rate for the rest of the conversation.
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 when tool definitions enter a model request, rather than changing what a function does after it is selected. For agents connected to dozens or hundreds of tools, deferring parameter details could reduce the amount of schema text sent when a task needs only a small subset. Google also says this design is intended to make tool selection easier and reduce latency as libraries grow, though the announcement gives no benchmark results to quantify those effects.
The practical result will depend on how often tools are used. Schemas that are never loaded may save input tokens; schemas that are frequently needed are still sent and billed, and retrieval itself adds tokens. Teams therefore need to compare the full cost and request behavior on their own traffic rather than assume that a larger library automatically means savings. In client-side setups, the extra search round trip is another factor to measure.
Google says deferred schemas are compatible with prefix caching because they are added later rather than included in the initial prompt. That may help preserve a shared cached prefix across turns, but the announcement does not provide independent measurements showing how caching, token use or response time change in production.
Products Worth Considering

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

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

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 generally send tool definitions, including parameter schemas, with a request. That gives the model those details up front, but can add context even when a particular task will use only one function. Tool Retrieval offers a different arrangement for large tool collections: retain tool names and descriptions for discovery, then provide full parameter details after the model identifies a candidate.
The announcement places the feature in the Interactions API and says it works with Gemini 3.6 and later, including all Gemini 4 models. It is in preview, not a general-availability release. Google directs developers to its setup documentation and invites feedback during testing. The announcement describes both managed, server-side retrieval and an option for applications that want to run retrieval through their own systems.
“Tool Retrieval lets your agent keep its full tool library on standby and load parameter schemas only when the current task needs them.”
— Google, in its Tool Retrieval announcement
Savings and Preview Limits Remain Unclear
Google has not published a general token-savings figure, benchmark results for latency, or measurements of tool-selection accuracy. The size of any benefit will depend on library size, how frequently tools are called, and how requests are structured. The company recommends measuring costs against actual traffic; the announcement does not establish that every workload will use fewer tokens or respond faster.
The announcement also does not specify a general-availability date or detail all preview limitations. Independent evaluations of whether models reliably find relevant tools as libraries grow were not provided. It remains unclear how the feature performs across different applications and workloads beyond Google’s description.
Measure Results During the Preview
Developers can test Tool Retrieval with supported models in the Interactions API while it remains in preview. The practical next step is to compare requests with deferred schemas against the team’s existing setup, tracking total input tokens, latency, and tool-selection behavior under representative traffic. Client-side implementations should also account for the added search round trip.
Google has invited users to share feedback, but has not announced when the feature will reach general availability. Until further measurements or release details are published, cost and performance gains should be treated as workload-dependent rather than guaranteed.
Key Questions
What does Tool Retrieval do?
It lets a Gemini API application defer full function parameter schemas. The model can see a tool’s name and description first, then request the schema when it needs that function.
Does Tool Retrieval reduce API costs?
It may reduce input tokens when schemas are deferred and never loaded, but Google gives no general savings estimate. Retrieval adds a small fixed block of input tokens, and loaded schemas are billed at standard input rates.
Does it work with MCP servers?
Google says deferred loading is supported for tools exposed through remote MCP servers. The announcement also describes client-side retrieval for applications using their own registries or search systems.
Which Gemini models support the feature?
Google says Tool Retrieval works through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models. The feature is currently in preview.
When will Tool Retrieval be generally available?
Google has not announced a general-availability date. The feature remains in preview, and the company is inviting developers to test it and share feedback.
Primary source: Google AI · via ThorstenMeyerAI.com
