Tool Call

A tool call is a structured request to perform an operation through an available tool. In a model-based application, the model can propose a tool name and arguments, while application code decides whether and how to execute them. A proposed call is not an execution result.

Trace a return request

Suppose the model requests a return for order 42. The application checks that the operation exists, the arguments have the expected form, the user may act on that order, and the action is within the authorized task. It then invokes the service and records its response.

Hugging Face's tool-use documentation illustrates this separation: model output requests a call, and application code handles execution and returns the tool result to the conversation. A correctly shaped request can still contain the wrong order identifier, so schema validation is not enough.

A timeout leaves a question to resolve

If the service times out, the return may have failed or succeeded without a received response. Reporting success from the model's wording would be unjustified. Check the operation state or use the service's retry protocol before repeating a state-changing request.

Keep tool results distinct from instructions that grant permissions. Retrieved content cannot authorize an unrelated action. The final user-facing response should reflect confirmed results and unresolved outcomes. One tool call can support a fixed workflow; repeated model-directed tool selection can form part of an agent.

Reference: Hugging Face: Tool use.