Run Tasks

Interfaze architecture allows you to programmatically run parts of the model or built-in tools without activating the full model making it significantly faster and cheaper.

Available tasks

Limits

• Only one task can be run at a time.
• The structured output is fixed for the task and cannot be customized.

How to run a task

• The system system prompt has to contain the task name in the format <task>task_name</task>
• The structured output response format has to be a type of any ot empty schema.

Example of system prompt:

<task>web_search</task>

Example of running a task

import { z } from "zod";
import { zodResponseFormat } from "openai/helpers/zod";

const response = await interfaze.chat.completions.create({
	model: "interfaze-beta",
	messages: [
		{
			role: "system",
			content: "<task>speech_to_text</task>",
		},
		{
			role: "user",
			content: [
				{ type: "text", text: "Transcribe the audio file https://r2public.jigsawstack.com/interfaze/examples/stt_long_audio_sample_3.mp3" },
			],
		},
	],
	response_format: zodResponseFormat(z.any(), "empty_schema"),
});

console.log(response.choices[0].message.content);

Output

• The output will always be a structured output with name of the task and the raw result
• The result schema is different depending on the task
• The result is the raw result of the specific model layer or tool
• Each task will have a consistent structure on evert run

The output is truncated for this example.

Common issues faced

• Only one <task> tag is parsed from the system message (the first match). One task can only be run at a time.
• If a non-empty schema is provided alongside a <task> tag, the schema takes priority — the run task execution is skipped and the full model is activated to generate a structured response following the schema.
• The response is the raw task output, not a natural language summary — plan your downstream processing accordingly.