Function calling & tool integration with the innoGPT API

With the Responses API, you can define your own tools (functions) that the AI model can call during a conversation. This forms the basis for agents, workflows with n8n, database queries, external API calls, and much more.

💡 Also works with n8n, Make, Zapier, and LangChain. Since the innoGPT API is OpenAI-compatible, you can use it in any tool that supports OpenAI as a provider. The complete interactive documentation for live testing is available when logged in under Settings → API → Documentation.

How Function Calling works

Define tools — you describe functions using JSON schema
Model decides — it analyzes the request and calls a tool if necessary
You execute — your code calls the actual function with the arguments
Return result — the model incorporates the result into its response

Example: Weather Tool

curl -X POST https://app.innogpt.de/api/ext/v1/responses \
  -H "Authorization: Bearer $INNOGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5",
    "input": "Wie ist das Wetter in Berlin?",
    "tools": [
      {
        "type": "function",
        "name": "get_weather",
        "description": "Aktuelles Wetter für einen Ort abrufen",
        "parameters": {
          "type": "object",
          "properties": {
            "location": {
              "type": "string",
              "description": "Stadtname, z. B. Berlin, Paris"
            },
            "units": {
              "type": "string",
              "enum": ["celsius", "fahrenheit"]
            }
          },
          "required": ["location"]
        }
      }
    ]
  }'

Tool definition: Schema

			FieldTypeRequiredDescription
`type`	string	✓	Must be "`"function"`"
`name`	string	✓	a-z, 0-9, underscores, max. 64 characters
`description`	string	✓	Clear description — determines when the model selects the tool
`parameters`	object	✓	JSON schema for the arguments

Control Tool Choice

With `tool_choice`, you define how aggressively the model should use tools:

	ValueBehavior
`"auto"` (Default)	Model decides on its own
`"none"`	Do not use any tools
`"required"`	At least one tool must be called
`{"type": "function", "name": "..."}`	Force specific tool

Parallel tool calls

By default, the model can execute multiple tools in parallel:

{
  "model": "gpt-5",
  "input": "Wie ist das Wetter in Paris, Tokio und New York?",
  "tools": [...],
  "parallel_tool_calls": true
}

⚠️ Known behavior: Anthropic and Google models do not always strictly adhere to the OpenAI convention when dealing with complex tool arguments (e.g., escaping in arrays). For purely parallel multi-tool scenarios in n8n workflows, GPT and Mistral models are generally the most reliable choice. Please contact the support team if you encounter any issues.

With `parallel_tool_calls: false`, tools are executed sequentially.

Limit Tool Loop

Prevents "tool runaways" caused by faulty model logic:

{
  "model": "gpt-5",
  "input": "Recherchiere zu Thema X",
  "tools": [...],
  "max_tool_calls": 5
}

The default is 10 tool calls per response.

Practical examples

Database query

{
  "type": "function",
  "name": "query_orders",
  "description": "Bestellungen aus der Datenbank abfragen",
  "parameters": {
    "type": "object",
    "properties": {
      "customer_id": {"type": "string"},
      "status": {
        "type": "string",
        "enum": ["pending", "shipped", "delivered", "cancelled"]
      },
      "limit": {"type": "integer"}
    },
    "required": ["customer_id"]
  }
}

Multi-Tool: Book a meeting + send an invitation

{
  "tools": [
    {
      "type": "function",
      "name": "book_room",
      "description": "Meeting-Raum buchen",
      "parameters": {
        "type": "object",
        "properties": {
          "room_id": {"type": "string"},
          "date": {"type": "string", "format": "date"},
          "time": {"type": "string"},
          "duration_minutes": {"type": "integer"}
        },
        "required": ["room_id", "date", "time"]
      }
    },
    {
      "type": "function",
      "name": "send_calendar_invite",
      "description": "Kalender-Einladung versenden",
      "parameters": {
        "type": "object",
        "properties": {
          "attendees": {"type": "array", "items": {"type": "string"}},
          "title": {"type": "string"},
          "datetime": {"type": "string", "format": "date-time"}
        },
        "required": ["attendees", "title", "datetime"]
      }
    }
  ]
}

Best practices

Clear tool descriptions

descriptions are crucial in determining when the model selects a tool:

{
  "name": "search_knowledge_base",
  "description": "Durchsucht die interne Wissensdatenbank nach Dokumentation, Richtlinien und Prozessen. Nutze dies bei Fragen zu Firmenprozessen, HR-Richtlinien oder internen Tools."
}

Precise parameter descriptions

{
  "date": {
    "type": "string",
    "description": "Datum im ISO-8601-Format (YYYY-MM-DD), z. B. 2026-03-15"
  }
}

Return errors cleanly

{"error": "Kunde nicht gefunden", "customer_id": "12345"}

This allows the model to respond to the error instead of hallucinating.

Webhooks for real-time events

If you want to respond to API events (e.g., a completed background job), use webhooks:

# Webhook anlegen
curl -X POST https://app.innogpt.de/api/ext/v1/webhooks \
  -H "Authorization: Bearer $INNOGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://deine-app.de/webhook", "events": ["response.completed"]}'

Additional endpoints:

GET /webhooks — list
POST /webhooks/{id}/test — Send test delivery
POST /webhooks/{id}/secret — Rotate secret
GET /webhooks/{id}/deliveries — Check delivery history

Was this helpful?

Still need help? Ask the team