Google Gemini API Launches Tool Retrieval Preview: Multi-Tool Agents No Longer Need to Handle Full Schema
Google has launched a preview version of the Tool Retrieval feature in the Gemini API, integrating tool filtering logic into the API layer. It enables agents to keep a full set of tools on standby and load parameter rules on demand, addressing pain points of excessive context occupation and redundant development processes for large-scale tool libraries. Currently, the feature is not yet officially commercialized.
Google has recently launched a preview version of the Tool Retrieval feature in the Gemini API.
Those who develop AI agents are familiar with this logic: To allow a large language model (LLM) to call external tools—whether web search, database queries, or access to internal enterprise systems—you must first provide the model with each tool's parameter schema, i.e., the rules for input and output parameters. The LLM uses this format to confirm specific calling methods.
Previously, when calling the Gemini API, developers needed to include the full parameter schemas of all tools available to the agent in every single request context. This was barely noticeable when there were only three or five tools, but once the tool library expanded to dozens or hundreds, the rule descriptions alone would occupy a large number of tokens. This directly increased call costs and easily took up space in the normal conversation context, slowing down response speeds. To implement agents with large-scale tool libraries, developers often had to build an additional tool retrieval layer by themselves, first screening out tools that might be used before passing them to the model.
The newly launched preview version of Tool Retrieval integrates this screening logic directly into the Gemini API itself. The agent can keep its full tool library on standby, and the entire process of judgment and loading is completed within a single Interactions API request. The model only loads the corresponding tool's parameter schema when it determines that a particular tool needs to be called, eliminating the need to carry all tools' rules in full every time.
For developers, this means no need to maintain an additional intermediate tool retrieval layer, and the token consumption per request will decrease, greatly lowering the barrier to implementing large-scale tool libraries. For end users, future multi-functional assistants built on Gemini will support a much larger number of tools without experiencing slowdowns due to increased tool count.
Currently, this feature is still in the preview stage. Relevant information comes from technology media MarkTechPost, written by Michal Sutter. You can view the original article to learn more.
发布时间: 2026-11-22 18:03