Skip to main content

Model API Reference

Authoritative reference for model integration interfaces and configuration.

Interfaces​

In the MVP, model integration is internal to the Runtime (there is no separate “model service” API).

When running the Runtime in openai mode, the Planner, Tool Args Generator, and Chat roles call an LlmClient:

  • jarvis_runtime.llm.client.LlmClient.responses(system_prompt, user_prompt, output_schema?, model_override?) -> dict

The shipped implementation is OpenAIResponsesClient, which uses the OpenAI Python SDK to call the OpenAI Responses API and (optionally) request JSON Schema structured outputs.

Configuration​

OpenAI mode configuration:

  • Required:
    • OPENAI_API_KEY (or JARVIS_OPENAI_API_KEY)
  • Optional:
    • OPENAI_BASE_URL (or JARVIS_OPENAI_BASE_URL)
    • JARVIS_MODEL_DEFAULT
    • JARVIS_MODEL_PLANNER, JARVIS_MODEL_TOOL_ARGS, JARVIS_MODEL_CHAT

Runtime selection:

  • --mode openai (or JARVIS_RUNTIME_MODE=openai)

Extensibility​

To add a new provider in the MVP:

  • Implement LlmClient.
  • Add a new runtime mode that instantiates your client and wires it into LlmPlanner, LlmToolArgsGenerator, and LlmChat.
  • Validate behavior via trace output (llm_call / llm_result events) and keep stub mode as a fallback.