Function calling guides

A practical recipe for getting tool calling right, from designing schemas to choosing a model, with every step linked to the in-depth docs on toolaugmented.ai.

Open toolaugmented.ai

The recipe

Six steps for shipping reliable tool calling. Try each one in the workspace as you go.

  1. 1Understand the loop
    What is function calling

    How the model, your tools and the tool-result turn fit together.

  2. 2Design your tools
    Schema best practices

    Names, descriptions, types, enums and distractor tools that models get right.

    Paste them in the Tools editor
  3. 3Write the system instruction
    System instruction

    Policies, defaults and scope: when to call, when to ask, when to answer directly.

  4. 4Shape tool outputs
    Tool output

    Return success, errors, timeouts and large results in a way the model can use.

    Add a Tool result turn
  5. 5Cover the call patterns
    Function call patterns

    Single, parallel, sequential and no-call cases, plus the tricky ones.

    Try them in the Arena
  6. 6Pick a model and measure
    Evaluation

    Metrics and per-slice evaluation on held-out tools, before you commit to a model.

    Compare models blind

All guides

Guides are hosted on toolaugmented.ai, our open documentation on tool-augmented AI.