A New Way To Retrieve Tools With The Gemini API
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: A New Way To Retrieve Tools With The Gemini API on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

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 a model requests a tool, while leaving names and descriptions available; Google has not published general savings or performance benchmarks.

Google announced Tool Retrieval in preview for its Gemini Interactions API on Nov. 19, 2026, as detailed in the original announcement, giving developers a way to load a function’s full parameter schema only after a model identifies that tool as relevant. The change targets agents connected to large tool libraries, where sending every schema with every request can use input tokens even when most tools are not needed, a concern related to fast, calibrated Gemini decisions.

With the server-side setup, developers add a tool_search tool and mark selected functions with defer_loading: true. The model can still see a deferred tool’s name and description, but its full parameter schema is held back. When the model requests a tool by name, the API returns the schema, after which the model can make the function call in the same request. The application then handles that call and returns its result through the usual process.

Google says the interaction’s steps show the search call, the search result containing the schema, and the subsequent function call. Developers can leave frequently used functions loaded and defer less common ones. Google also says server-side retrieval requires no index for developers to build or host, and that deferred loading is supported for tools exposed by remote MCP servers.

Developers that already maintain a tool registry, embedding index or per-user permission filter can instead use client-side retrieval: their application searches its own system and returns matching schemas. Google says this option involves an extra round trip for each search. The feature is available through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models, according to the announcement.

At a glance
announcementWhen: Announced Nov. 19, 2026; available in p…
The developmentGoogle has introduced a preview feature that retrieves deferred tool schemas on demand in the Gemini Interactions API.
AI tools · Briefing

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.

Nov. 19, 2026
Preview announced
Gemini 3.6+
Supported models
Gemini 4
All models supported

What changes for agents

Tool names and descriptions remain visible; full parameter schemas load when requested.

1
Keep schemas deferred

Developers mark selected functions for deferred loading.

2
Find a relevant tool

The model requests a tool by name and receives its schema.

3
Call the function

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.

Server-side
  • Uses the tool_search tool.
  • No index to build or host.
  • Supports remote MCP server tools.
Client-side
  • 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.

Potential saving

Skip unused schemas

Savings depend on whether schemas left unloaded outweigh schemas retrieved.

Added usage

Search and loaded schemas

Search adds a small fixed block of input tokens; loaded schemas count as input for the rest of the conversation.

Preview caveatGoogle has not provided a general savings figure or benchmark results. Measure token use and request behavior against your own traffic.
The takeaway
Tool Retrieval may help large-library agents avoid sending unused schemas, but teams should measure its costs and performance on their own workloads.

Less Schema Data in Each Request

Tool Retrieval changes when an agent receives tool definitions, rather than how the function is ultimately executed. For developers working with dozens or hundreds of tools, deferring parameter details could keep unused schemas out of the initial request and reduce repeated prompt overhead. It may also make a large set of candidate tools easier for a model to work through, Google says.

The cost effect is not guaranteed. Google says the feature has no separate charge, but retrieval adds a small fixed block of input tokens, and schemas that are loaded are billed as input at the model’s standard rate for the rest of the conversation. Savings depend on how many schemas are never loaded compared with the retrieval overhead. A system that uses most of its tools regularly may see less benefit than one with many rarely used functions.

For teams, the announcement is a new option to test, not evidence of a quantified improvement. Google has not released a general savings figure or benchmark results for token use, latency or tool-selection accuracy. The practical effect will depend on the size and usage pattern of each library and how the application makes requests.

Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight

Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB-C MagSafe Charging Case

Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB-C MagSafe Charging Case

As an affiliate, we earn on qualifying purchases.

From Full Schemas to On-Demand Loading

In conventional function calling, developers send tool definitions, including parameter schemas, along with a request. That gives the model the information needed to call functions, but can add to the request even when many available tools have no bearing on the current task. Tool Retrieval separates the brief information used to identify a tool from the full schema needed to call it.

Google says the deferred schemas are compatible with prefix caching: because those schemas arrive later rather than at the start of the prompt, the company says the cached prefix can remain unchanged between turns. This is a statement about the design; the announcement does not provide independent measurements of caching effects or overall performance.

Tool Retrieval is also distinct from an application’s own retrieval system. Server-side retrieval lets the Gemini API search the available deferred tools, while client-side retrieval leaves search with the developer’s system and returns selected schemas to the API. Google describes the latter as useful for teams that need their own registry or access filters, with the trade-off of an additional round trip.

““Tool Retrieval lets your agent keep its full tool library on standby and load parameter schemas only when the current task needs them.””

— Google

Savings and Preview Limits Remain Open

The announcement does not quantify token savings, latency changes or tool-selection accuracy, and it does not report independent evaluations across different library sizes. Google advises developers to measure the effects with their own traffic. Results may differ depending on how often tools are used, how many schemas are deferred and whether retrieval is server-side or client-side.

Google has not specified a date for general availability or described all limitations that may apply during the preview. It is also unclear how performance changes for agents that frequently need many deferred tools in a single interaction. Until those details or broader measurements are available, cost and latency benefits should be treated as workload-dependent rather than established outcomes.

Test the Preview Against Real Workloads

Developers can consult Google’s setup documentation and try Tool Retrieval with supported models through the Interactions API while the feature is in preview. Google has invited users to discuss the feature and share feedback. Teams assessing it can compare requests with and without deferred schemas, tracking total input tokens, added search steps, latency and whether the model selects the appropriate function.

Google has not announced a general-availability timetable. For now, the next concrete step is testing against each application’s own tool usage rather than assuming a fixed reduction in cost or delay. Further documentation, preview changes or measurements may clarify the feature’s limits, but none were specified in the announcement.

Key Questions

What is Tool Retrieval?

Tool Retrieval is a Gemini Interactions API preview that lets developers defer full function parameter schemas until a model requests a tool. The model can still see the deferred tool’s name and description.

How does the server-side setup work?

Developers add a tool_search tool and set defer_loading: true for functions whose schemas should be held back. After the model requests one by name, the API returns its schema so the model can proceed with the function call.

Does Tool Retrieval guarantee lower costs?

No. Google says there is no separate feature charge, but retrieval adds a small fixed block of input tokens, and loaded schemas are billed at the model’s standard input rate. Whether overall use falls depends on how many schemas remain unused.

Which Gemini models support the preview?

Google says Tool Retrieval 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 remains in preview, and the announcement did not specify all preview limitations.

Primary source: Google AI · via ThorstenMeyerAI.com

You May Also Like

Anthropic’s Claude: Building The Future Of Artificial Intelligence

Anthropic claims its AI model, Claude, is currently assisting in building the next version of itself through AI-generated code and research, marking a step toward self-improving AI.

Nvidia Nemotron 3.5 Lightning And NeMo Switchyard

Nvidia announced the Nemotron 3.5 Lightning and NeMo Switchyard, expanding its AI hardware ecosystem. Details remain limited, with official specs pending.

AI Innovation At Its Early Stage: How To Experiment With SpaceXAI’s Grok Bot

SpaceXAI’s Grok bot is now in early beta, offering limited user access to test its conversational AI capabilities amid fierce industry competition.

Dessa: The Most Authentic Full-duplex Speech Model Ever Built

Block’s Dessa page calls its offering an authentic full-duplex speech model, but the supplied source contains no technical evidence or launch details.