Technology releaseNovember 21, 2026Редакция 105.kz

Gemini API adds Tool Retrieval: tool schemas load on demand

Google has released Tool Retrieval in the Gemini API in preview: the agent sees the names and descriptions of all tools but loads their parameter schemas only on request. We look at what this gives agents over DCS and maintenance-and-repair systems and what stands in the way of applying the approach in a closed perimeter.

ManufacturingPrivate LLMPredictive Maintenance

Gemini API adds Tool Retrieval: tool schemas load on demand

Key takeaway. Google has released Tool Retrieval for the Gemini API in preview: the agent keeps a library of tools “in a warehouse” and loads parameter schemas only on demand, rather than in every request. For a plant this is an architectural technique, not a ready-made solution: the service is cloud-based, so in a perimeter with an on-premise requirement its use requires separate work, and token savings must be measured on your own traffic.

What happened

On 19 November 2026, Google announced the launch of Tool Retrieval in the Gemini API — the feature is available in preview in the Interactions API. The authors of the publication are Lucia Loher, product manager for Gemini API, and Nora Kassner, research scientist at Google DeepMind. The function’s job: not to send the full parameter schemas of all tools the agent is connected to in every request, but to load them as needed.

The stated reason: when an agent is connected to dozens or hundreds of tools, sending all schemas “inflates input token counts, adds latency, and lowers tool-selection accuracy as the candidate list grows.” This is enabled by the tool_search tool and the defer_loading: true flag on the tools that need to be held back: the model sees the name and description of each of them, but the parameter schemas remain outside the context. When a tool is needed, the model requests it by name, the API loads the schema, the model calls the function — all within a single request. All three steps are visible in interaction.steps (tool_search_call, tool_search_result, function_call); the authors note that this simplifies debugging of tool names and descriptions.

Additionally from the publication: there is no need to build and host an index, the search runs on the Gemini server, and only the new tool and the flag change in the code. The solution is compatible with prefix caching and with remote MCP servers: the defer_loading: true flag on an mcp_server defers all tools that this server provides. There is a client-side option ("execution": "client"), when the search is performed by the application itself: one round trip is added for each search, but it becomes possible to connect new tools in the middle of a conversation and take user permissions into account. There is no separate charge for Tool Retrieval — standard pricing applies; tool_search adds a small fixed block of input tokens, and each loaded schema is billed as input tokens until the end of the conversation. Gemini 3.6 and newer models are supported, including all Gemini 4 models, in the Interactions API.

What this means for enterprises in Kazakhstan

For a plant, the topic is practical. An agent for maintenance and repair almost never starts with a single tool: it is connected to the DCS and SCADA archive, operating-hours and failure history, vibration diagnostics results for pump and induced-draft fan bearings, the MRO catalogue, procedures and work permits. The problem described by the source is familiar even without Gemini: the longer the candidate list, the lower the accuracy of selecting the right tool. Tool Retrieval offers an organizational technique — keep the library “in a warehouse” and load schemas on demand, while keeping frequently called tools loaded.

Agent tool class Data What matters with deferred loading
Reading parameter history SCADA/DCS archive, tags, operating hours tag names in the tool description must not disclose more than necessary
Diagnostics vibration spectra, temperatures, currents the parameter schema is large, calls are rare — a candidate for deferred loading
Documents procedures, datasheets, operating instructions frequent calls, better kept loaded
Maintenance and repair operations requests, work permits, MRO depends on user permissions — an argument for client-side search

Two limitations for the Kazakhstan context. First: the Gemini API is a cloud service, while for national companies and subsoil users the norm is on-premise and personal data localization; connecting an external API to the DCS perimeter is possible only after a separate cybersecurity review (assessment). Second: with deferred loading, the model judges a tool only by its name and description — which means descriptions in Russian, and where required also in Kazakh, must be unambiguous, without synonyms or internal jargon. This is separate work, not a side effect (assessment).

For those building an agent in a closed perimeter, the value of the publication is not the service itself but the pattern: a tool registry, deferred schema loading, and a log of search steps as a debugging aid. Reproducing such a pattern on-premise is a separate engineering task (assessment); legacy DCS and heterogeneous equipment catalogues make it non-trivial.

How to apply: an 8–12 week pilot

  1. Tool and owner registry (1–2 weeks). List which systems the agent must call: SCADA archive, vibration diagnostics, MRO catalogue, procedures, maintenance and repair requests. For each tool — owner, data class, required permissions, call frequency. Without this registry, deferred loading becomes a formality.
  2. Data classification and choice of perimeter (1 week). Separately mark tools that work with personnel and trade secrets, and with process parameters. This determines whether an external API is acceptable or a closed perimeter with a local model is needed — see Enterprise LLMs.
  3. Baseline on your own traffic (2 weeks). The source directly advises measuring the effect on your own traffic. Collect 50–100 real requests from engineers and dispatchers, record token consumption, latency and the share of correctly selected tools. This is the basis for acceptance.
  4. Pilot on one domain (4–6 weeks). Limit the scenario, for example to diagnostics of pumping and draft-and-fan equipment. Use the search step log as a working tool for error analysis: it shows which tools the model considered. An adjacent class of tasks is Predictive maintenance.
  5. Acceptance (1 week). Metrics: tool-selection accuracy, share of successful calls, cost per request, latency, share of correct “no data” refusals. Decide on scaling by numbers, not by a demonstration.

Limitations and risks

  • Preview status. The source says nothing about SLA, stability or timelines for general availability. For an industrial perimeter this is a blocking factor (assessment).
  • Not a single measured figure. Google claims lower input tokens and higher selection accuracy, but provides no measurements either for savings or for latency. Everything stated is a vendor claim, not a measurement.
  • Savings are conditional. The source itself stipulates: tokens are saved when unloaded schemas outweigh loaded ones. A loaded schema is billed until the end of the conversation, and tool_search adds a fixed block of tokens on top.
  • Client-side search adds a round trip for each search — this is stated directly and affects dialogue latency.
  • Description quality is the weak point. The model sees only the name and description; with imprecise wording, selection accuracy falls rather than rises.
  • Cloud nature of the service. The requirement for on-premise and data localization for Kazakh national companies and subsoil users is not removed. The publication contains no data on service availability in Kazakhstan or on support for Russian and Kazakh.
  • Industrial specifics are not covered. The source has no information on working with OPC servers, industrial protocols and legacy DCS; the examples are MCP servers and return-processing functions. Applicability to production systems is a hypothesis to be tested by a pilot, not a fact.

Sources

  1. Introducing Tool Retrieval in the Gemini API · 2026-11-19T18:00:00.000Z
    Load tool parameter schemas on demand instead of sending your full tool library in every prompt, designed to reduce input tokens and keep tool selection accurate as your agent scales.
  2. Фото: Linda Spashett Storye_book — Bradford Industrial Museum 030.jpg (CC BY 3.0, Wikimedia Commons)

This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.

FAQ

Frequently asked questions

Short, straight answers to the questions we hear most often.

Can Tool Retrieval be used in a closed perimeter without internet access?

The source says nothing about this; the Gemini API is a Google cloud service. The technique itself — keeping the tool library on your side and loading schemas on demand — is reproducible in your own perimeter, but this is separate development (assessment), not a flag setting.

Will this save tokens on our agent?

Not necessarily. The source indicates that savings arise when schemas that are not loaded outweigh loaded ones, and advises measuring the effect on your own traffic. There are no guaranteed percentages in the publication.

What exactly does the model know about deferred tools?

Only the name and description. The parameter schema remains outside the context and is loaded at the moment the tool is accessed — within the same request.

Do we need to build and host a search index?

In the server-side option — no, Gemini performs the search. The client-side option (`"execution": "client"`) is needed if the enterprise has its own tool registry or user permission filter; then one round trip is added for each search.

Next step

Discuss a pilot project

Tell us about your task — we'll come back with a data audit plan and an effect estimate within two business days.

  1. 1A 30-minute call: the task, your current systems, who makes the decision
  2. 2A review of data and processes, a staged plan with time and cost ranges
  3. 3A 2–4 week prototype or an 8–12 week pilot with success criteria agreed upfront

NDA from the first contact

Reply within one business day