Public developments and original analysis · Announcement date: · Sources
A customer asks a support agent whether an invoice has been issued. The agent may already have received the parameter schemas for dozens of operations across ordering, shipping, invoicing, returns and warehousing. On November 19, Google announced Tool Retrieval in preview for the Gemini Interactions API. With server-side retrieval, the model sees the deferred tools' names and descriptions. When it chooses one, the API loads its full parameter schema within the same request and the model emits a function call. The application still runs its own function and returns the result. The schema tells the model what fields to send, their types and which ones are required, details it may rarely need for a given conversation.

The documentation requires the v1beta Interactions API for this preview and excludes generateContent. Google lists Gemini 3.6 and later models as supported. Server-side retrieval also supports deferred loading for tools from a remote MCP server. If a company maintains its own registry, its application can search for candidates and send them back using client-side retrieval. That adds a round trip per search in exchange for control over what the application returns.
The cost of tools over a full conversation
Suppose most customers in that support queue ask for order or invoice status. The team can keep those two tools available upfront and defer the schemas for less common operations, such as investigating a carrier exception or resending an electronic invoice. When an intake rule already identifies a class of requests, routing them to a small fixed set can work well. It limits competition between similarly named tools and makes a wrong-system call easier to trace.
Costs depend on where the conversation goes next. A customer who checks an invoice might then mention an address change and ask about an older order. Several sets of parameters can accumulate over those turns. Google says it charges no separate fee for retrieval, but enabling it adds a small fixed amount of input, and loaded schemas are billed as input for the rest of the conversation. The technical documentation adds that every server-side search triggers another model call within the request. Its extra input is metered as total_tool_use_tokens. Unused schemas have to outweigh these costs for deferral to pay off. If most conversations eventually use the same few interfaces, declaring them upfront can also make a common prompt prefix easier to reuse. A saving measured on one short request is a poor monthly estimate.
Replay the same historical requests by intake route, number of turns and number of required tools. Compare a fixed shortlist with server-side deferral and include client-side retrieval if it serves a real need. Capture token usage and actual charges, including cached and retrieval input, empty searches, retries and handoffs. Measure wait time up to a usable business result, including the extra request and the time spent in business systems. A smaller input can cost more overall if the agent confuses similarly named functions and somebody has to redo the work. Cost per correctly completed case and the wait for slower cases capture both effects.
What the agent can discover and what it can execute
The tool descriptions are visible before server-side retrieval happens. Suppose a catalog includes an operation for checking public stock by store and another for reading internal inventory by bin. Deferring the second schema still lets the model see that operation's name and purpose. If store staff should only see the public tool, the application has to filter candidates by identity before exposing them to the model. A team with a tenant-specific registry can use it to produce a fixed list or return candidates through client-side retrieval. The extra trip makes more sense when the allowed set changes frequently by user or stage of work.
Discovery and execution have separate checks. Suppose somebody uses an old page to request that an invoice be resent. After the model selects the right tool, the application still has to check the order's current state, the operator's authority and whether the request has already been processed, then obtain any required confirmation. Those checks affect the outcome even when Gemini loads the right schema. Test with similarly named tools, customers with several orders and already-processed requests. The trace should distinguish a failed search, a wrong choice and a call the application correctly declined. Each result points to a different fix.
A fixed shortlist saves preview integration and debugging work when the tickets already fall into clear groups and each group needs only a few stable operations. On-demand loading is more promising when follow-up questions regularly cross systems and a typical conversation uses only a fraction of a large tool library. Apply the same authorization and accuracy requirements to each option, then compare what it costs to finish the same cases and how soon the users with the longest waits get a result.
Sources
- Google, Introducing Tool Retrieval in the Gemini API
Source published: · Accessed:
- Google AI Studio, Tool retrieval
Accessed:
- Google AI Studio, Function calling with the Gemini API
Accessed:
Mokyun Editorial · neo@mokyun.com
