Glossary¶

Agent: A model running in a loop that chooses and calls tools until a goal is met.
AST match: An evaluation method that parses a predicted call and compares the function name and each argument with the expected one, allowing equivalent values.
Automatic function calling: An SDK feature that runs the tool loop for you when you pass real functions (e.g. the Gemini Python SDK, or the Anthropic tool runner).
Chat template: The model-specific text format (often Jinja) that turns messages + tools into tokens. Training and serving must use the same one.
Constrained decoding: Restricting token sampling to a grammar or JSON Schema so that output always parses. It powers strict mode.
Distractor tool: A plausible but wrong tool included in training or eval to test selection.
Function calling: See Tool calling.
Irrelevance detection: Correctly making no call when no tool fits.
MCP: Model Context Protocol: an open standard for exposing tools, resources, and prompts to LLM clients.
Multi-turn: A conversation with more than one user message (Turn).
Parallel calls: Several independent tool calls emitted in one assistant turn.
Consecutive (sequential) calls: Calls where a later one depends on an earlier result (Function Call).
Programmatic tool calling: The model writes code that calls tools inside a sandbox, instead of one call per model turn.
Server / built-in tool: A tool executed by the provider (web search, code execution, file search).
Strict mode: A provider option that guarantees arguments conform to the schema.
Tool calling / tool use: The model emitting a structured request (name + arguments) for the application to run.
tool_choice: Request parameter controlling whether the model may, must, or must not call tools.
Tool result: The application's response to a tool call, linked by the call id.
Tool search / pivoting: Retrieving the relevant tools at runtime from a large set (Tool Search).
Trajectory: The full sequence of messages, calls, and results for one task. It's the unit of training data for agents.