Tool calling

Tool calling is OpenAI’s tools API, on yahoobot-1 only: israelbot-1 answers 400 tools_not_supported to any request that uses it. Tools that put their instructions in the prompt (Cline and Roo Code by default) don’t need any of this and work on text alone.

1. Describe your tools and ask:

{
  "model": "yahoobot-1",
  "messages": [{ "role": "user", "content": "What's the weather in Haifa?" }],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Current weather for a city",
      "parameters": {
        "type": "object",
        "properties": { "city": { "type": "string" } },
        "required": ["city"]
      }
    }
  }],
  "tool_choice": "auto"
}

2. The model answers with a call instead of text:

{
  "choices": [{
    "index": 0,
    "finish_reason": "tool_calls",
    "message": {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "call_9a3f",
        "type": "function",
        "function": { "name": "get_weather", "arguments": "{\"city\":\"Haifa\"}" }
      }]
    }
  }]
}

3. Run it, and send the result back with the assistant message exactly as it came, then one tool message for each call:

{
  "model": "yahoobot-1",
  "tools": [ ...the same tools... ],
  "messages": [
    { "role": "user", "content": "What's the weather in Haifa?" },
    { "role": "assistant", "content": null, "tool_calls": [
      { "id": "call_9a3f", "type": "function", "function": { "name": "get_weather", "arguments": "{\"city\":\"Haifa\"}" } }
    ] },
    { "role": "tool", "tool_call_id": "call_9a3f", "content": "31C and sunny" }
  ]
}

The model then answers in words (finish_reason: "stop"), or asks for another call. When streaming, a call arrives as delta.tool_calls pieces: the first has the id and function.name, and the rest add to function.arguments, which is a JSON string in fragments, as OpenAI’s does. They are passed through unchanged.

The arguments are the model’s. They are not checked against your schema: parse them, validate them, and handle bad JSON, as with any model. Tool schemas, calls and results count toward your prompt size and your daily tokens.

With "stream": true a call arrives as delta.tool_calls pieces, as OpenAI's do. The first piece has the index, id, type and function.name; the rest add to function.arguments, a JSON string in fragments that you join before parsing. The last chunk has finish_reason: "tool_calls". They are passed through as the model wrote them.

data: {"choices":[{"index":0,"delta":{"role":"assistant","content":null,"tool_calls":[{"index":0,"id":"call_9a3f","type":"function","function":{"name":"get_weather","arguments":""}}]},"finish_reason":null}], ...}

data: {"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"city\":"}}]},"finish_reason":null}], ...}

data: {"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\"Haifa\"}"}}]},"finish_reason":null}], ...}

data: {"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}], ...}

data: [DONE]

• The arguments are the model’s. They are not checked against your schema. Parse them, validate them, and handle bad JSON, as with any model.
• A model will sometimes answer in words when you expected a call, or call a tool you didn’t offer. Use tool_choice to require one, and ignore a call to a name you don’t have.
• Tool definitions, calls and results all count toward your prompt size and your daily tokens, every turn. Keep descriptions short and results small.
• Keep every assistant message that has tool_calls in the conversation you send back, with its tool results after it; a tool message with no matching call is refused.
• Limits above (128 tools, 32,768-byte schemas) are checked before the model runs, so a refusal costs nothing.

Errors are listed under Errors: invalid_tools for a malformed tool or call, tools_not_supported for israelbot-1.