🔍 Read the full analysis: Introducing 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 keeps tool names and descriptions visible while loading full parameter schemas only when requested; Google says the feature has no separate charge, but savings depend on use and have not been quantified.
Google announced Tool Retrieval in preview for the Gemini Interactions API on Nov. 19, 2026, giving developers a way to load full tool parameter schemas only when a model requests a function, as detailed in the original analysis. The feature is intended for agents using large tool libraries, where sending every schema in each request can add input tokens; Google has not published a general savings figure, a concern related to calibrated multimodal decisions.
With the server-side option, developers add a tool_search tool and set defer_loading: true for functions whose parameter schemas should be withheld initially. The model can still see those tools’ names and descriptions. When it 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 the function and returns its result as it normally would.
Google says developers can leave frequently used tools loaded and defer less common ones. The interaction steps show the search call, a result containing the requested schema, and the subsequent function call. Google also says there is no index to build or host for server-side retrieval. Tool Retrieval can be used with tools exposed by remote MCP servers; marking those tools for deferred loading makes them available on demand.
Developers with their own registry, embedding index, or per-user permission filters can instead use client-side retrieval. In that setup, the application searches its own system and returns matching schemas. Google notes that this approach adds a round trip for each search. The announcement says the feature works through the Interactions API with Gemini 3.6 and later, including all Gemini 4 models.
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 Larger Gemini Tool Libraries
Tool Retrieval addresses a scaling problem for agents connected to many functions: a request may need only one tool, but conventional function calling can put definitions for the whole library into the prompt. Keeping unused parameter details out of the initial context could reduce repeated input and leave more room for task-specific information. It may also help developers organize access to tools without presenting every full schema at once.
Those benefits remain a design aim, not a measured result in the announcement. Google says the retrieval feature carries no separate charge, but search adds a small fixed block of input tokens, and each loaded schema is billed as input at the model’s standard rate for the rest of the conversation. The net cost depends on how many schemas are ultimately loaded. Teams whose agents use most tools often may see less benefit than teams with large libraries and many rarely used functions.
Google also says deferred schemas are compatible with prefix caching, because they enter the interaction later rather than changing the initial prompt prefix. That could matter for applications reusing a stable prompt, though the announcement supplies no latency or cost measurements to show the effect in practice.
Products Worth Considering

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W

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

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, developers provide function definitions, including parameter schemas, with a request. For a large tool collection, this can mean repeatedly including details that are irrelevant to the current task. Tool Retrieval changes when those details are supplied: names and descriptions remain available for discovery, while full schemas are returned after a tool is selected.
The feature is a preview within the Interactions API, Google’s interface for building model interactions that can include tools. The announcement describes both a managed server-side search path and a client-side option for applications that already maintain their own search or permissions system. It does not report independent testing or benchmark comparisons against conventional tool definitions.
““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 Remain Unknown
Google has not provided benchmark results for token savings, latency, or tool-selection accuracy. It says results depend on which schemas are loaded and recommends developers measure costs against their own traffic. The announcement does not establish how retrieval performs across different library sizes or task patterns, or whether any token reduction changes the accuracy of tool selection.
The company has not specified a general-availability date or detailed the preview’s limitations. It is also unclear how much the fixed input-token cost of searching affects different workloads. For client-side retrieval, the additional round trip is a stated trade-off; its effect on end-to-end response time is not quantified.
Measure Results During the Preview
Developers can consult Google’s setup documentation and test the feature with supported models in the Interactions API. Google has invited users to discuss Tool Retrieval and share feedback. For teams evaluating it, the next step is to compare requests with and without deferred schemas using representative workloads, tracking total input tokens, response time, tool-selection behavior, and how often deferred tools are actually loaded.
Those measurements will help determine whether retrieval benefits a particular application before teams rely on it more broadly. Google has not announced when the preview will end or when the feature may become generally available.
Key Questions
What is Tool Retrieval in the Gemini API?
It is a preview feature for the Gemini Interactions API that lets developers defer full function parameter schemas and load them when the model requests a tool. Tool names and descriptions remain visible before the schema is retrieved.
How does server-side Tool Retrieval work?
Developers add a tool_search tool and mark selected functions with defer_loading: true. When the model requests one of those tools, the API returns its schema, allowing the model to make the function call.
Does Tool Retrieval reduce costs?
It may reduce input-token use when many schemas are not needed, but Google has not published a general savings figure. Search adds a small fixed block of input tokens, and loaded schemas are billed at the model’s standard input-token rate.
Which models support Tool Retrieval?
Google’s announcement says it 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 announced the feature in preview but has not provided a general-availability date.
Primary source: Google AI · via ThorstenMeyerAI.com
