AI Completion

Run LLM completions from multiple providers within a flow. Supports tool calling, streaming, and sandboxed execution.

Configuration

- ai_completion:
    name: analyze
    provider: google
    model: gemini-2.5-flash-lite
    credentials_path: /path/to/credentials.json
    prompt: "Analyze this data: {{event.data}}"

Fields

Field	Type	Default	Description
`name`	string	required	Task name.
`provider`	string	required	LLM provider (`google`, `anthropic`, `openai`).
`model`	string	required	Model identifier.
`credentials_path`	string	required	Path to provider credentials.
`prompt`	string/resource	required	Prompt template. Supports templating and resource files.
`system_prompt`	string/resource		System prompt.
`max_tokens`	int		Maximum tokens in response.
`temperature`	float		Sampling temperature.
`tools`	list		Tool definitions for function calling.
`sandbox`	object		Sandbox configuration for tool execution.
`depends_on`	list		Upstream task names.
`retry`	object		Retry configuration.

Sandbox

Optional sandboxing for tool execution via nsjail. Rhai scripts do not need sandboxing (safe by design).

- ai_completion:
    name: agent
    provider: google
    model: gemini-2.5-flash-lite
    prompt: "{{event.data}}"
    sandbox:
      memory_limit_mb: 512
      time_limit_seconds: 30
      max_pids: 10
      allow_network: false