Function calling - Yelu AI API

Function calling lets a model request a named application operation with structured arguments. Your code—not the model—decides whether to execute the function and returns the result for a final response.

Define and request a tool

The following request gives the model one read-only weather function. It intentionally stops after the model’s tool request so you can inspect and validate the arguments.

curl https://api.yelu.ai/v1/chat/completions \
  -H "Authorization: Bearer $YELU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o-mini",
    "messages": [{"role":"user","content":"What is the weather in Shanghai?"}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {"type":"string","description":"City name"},
            "unit": {"type":"string","enum":["celsius","fahrenheit"]}
          },
          "required": ["city", "unit"],
          "additionalProperties": false
        }
      }
    }],
    "tool_choice": "auto"
  }'

const body = {
  model: 'gpt-4o-mini',
  messages: [{ role: 'user', content: 'What is the weather in Shanghai?' }],
  tools: [{
    type: 'function',
    function: {
      name: 'get_weather',
      description: 'Get current weather for a city',
      parameters: {
        type: 'object',
        properties: {
          city: { type: 'string', description: 'City name' },
          unit: { type: 'string', enum: ['celsius', 'fahrenheit'] },
        },
        required: ['city', 'unit'],
        additionalProperties: false,
      },
    },
  }],
  tool_choice: 'auto',
};

const response = await fetch('https://api.yelu.ai/v1/chat/completions', {
  method: 'POST',
  headers: { Authorization: `Bearer ${process.env.YELU_API_KEY}`, 'Content-Type': 'application/json' },
  body: JSON.stringify(body),
});
if (!response.ok) throw new Error(await response.text());
console.log((await response.json()).choices[0].message.tool_calls);

import os
import requests

payload = {
    "model": "gpt-4o-mini",
    "messages": [{"role": "user", "content": "What is the weather in Shanghai?"}],
    "tools": [{
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a city",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string", "description": "City name"},
                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
                },
                "required": ["city", "unit"],
                "additionalProperties": False,
            },
        },
    }],
    "tool_choice": "auto",
}
response = requests.post(
    "https://api.yelu.ai/v1/chat/completions",
    headers={"Authorization": f"Bearer {os.environ['YELU_API_KEY']}"},
    json=payload,
    timeout=60,
)
response.raise_for_status()
print(response.json()["choices"][0]["message"].get("tool_calls"))

Complete the tool loop

Inspect the assistant message

If finish_reason is tool_calls, read every item in message.tool_calls. A model can request more than one call.

Validate arguments

Parse function.arguments as JSON and validate it against your own schema, permissions, value ranges, and business rules. Treat it as untrusted input.

Execute with safeguards

Map the exact allowlisted function name to application code. Apply authentication, authorization, deadlines, idempotency, and rate limits.

Return tool results

Append the assistant tool-call message and one role: tool message per call, matching tool_call_id. Then request the final assistant response.

{
  "role": "tool",
  "tool_call_id": "call_01JYELU",
  "content": "{\"city\":\"Shanghai\",\"temperature\":31,\"unit\":\"celsius\",\"condition\":\"cloudy\"}"
}

Security rules

Never use eval, reflection, or arbitrary shell execution on a generated function name.
Require user confirmation for purchases, messages, deletion, permission changes, or other consequential actions.
Bind tools to the authenticated user and re-check authorization at execution time.
Limit argument length and recursive JSON depth.
Remove secrets and unnecessary internal details from tool results.
Cap loop iterations and detect repeated identical calls.
Record an audit event for state-changing tools.

A JSON schema improves argument shape; it does not make the requested action safe or correct. Application-side validation and authorization remain mandatory.

​Define and request a tool

​Complete the tool loop

​Security rules

Define and request a tool

Complete the tool loop

Security rules