Blend AI Documentation

When a published run completes, GET /api/v1/runs/{runId} returns an output object keyed by the normalized label of each User Result node. Each key maps to an array of output items. Each output item preserves one logical unit of work: one normal run, one batch iteration, one selected item, or one element from an LLM structured list.

{
    "output": {
        "caption": [
            {
                "id": "item_1",
                "item_index": 0,
                "status": "completed",
                "parts": [
                    { "type": "text", "text": "A cat on the moon, gazing at earth." }
                ]
            }
        ],
        "image": [
            {
                "id": "item_2",
                "item_index": 0,
                "status": "completed",
                "parts": [
                    { "type": "image", "url": "https://cdn.tryblend.ai/runs/abc/image.png" }
                ]
            }
        ]
    }
}

Output items

Field	Type	Notes
`id`	`string`	Stable item id.
`item_index`	`number`	Ordered item position, starting at `0`.
`status`	`completed` / `failed` / `pending` / `running`	Completed API output usually contains completed or failed items.
`parts`	`OutputPart[]`	The content fragments inside this item.
`error`	`object?`	Present when the item failed. Contains `message`, and may include `type` or `code`.

Why an array of items? A Result can receive multiple logical outputs: a Batch with five rows, a multi-select Choose Output, or an LLM structured list with several elements. Each logical output stays separate. Why can one item contain multiple parts? One provider call can produce several fragments, such as text plus a file, or several files from one generation call.

Part types

Check part.type and branch on it.

text

Plain string output — chat responses, captions, summaries, transcriptions.

{ "type": "text", "text": "A cat on the moon, gazing at earth." }

Field	Type	Notes
`text`	`string`	The content.

image

An image file hosted on Blend AI’s CDN.

{
    "type": "image",
    "url": "https://cdn.tryblend.ai/runs/abc/image.png",
    "name": "image.png"
}

Field	Type	Notes
`url`	`string`	Direct CDN URL.
`name`	`string?`	File name when known.

video

A video file hosted on Blend AI’s CDN.

{
    "type": "video",
    "url": "https://cdn.tryblend.ai/runs/abc/video.mp4",
    "name": "video.mp4"
}

Field	Type	Notes
`url`	`string`	Direct CDN URL.
`name`	`string?`	File name when known.

file

Any non-image, non-video file, including audio, PDFs, documents, or archives.

{
    "type": "file",
    "url": "https://cdn.tryblend.ai/runs/abc/audio.mp3",
    "name": "audio.mp3",
    "mime_type": "audio/mpeg"
}

Field	Type	Notes
`url`	`string`	Direct CDN URL.
`name`	`string?`	File name when known.
`mime_type`	`string?`	MIME type when known.

json

Structured JSON data returned by a workflow output.

{
    "type": "json",
    "data": {
        "title": "Moon Cat",
        "score": 0.92
    }
}

Field	Type	Notes
`data`	`unknown`	The structured value.

Reading parts in your code

type OutputPart =
    | { type: 'text'; text: string }
    | { type: 'image' | 'video'; url: string; name?: string }
    | { type: 'file'; url: string; name?: string; mime_type?: string }
    | { type: 'json'; data: unknown }

type OutputItem = {
    id: string
    item_index: number
    status: 'completed' | 'failed' | 'pending' | 'running'
    parts: OutputPart[]
    error?: { message?: string; type?: string; code?: number }
}

function renderPart(part: OutputPart): string {
    switch (part.type) {
        case 'text': return part.text
        case 'image': return `![](${part.url})`
        case 'video': return `<video src="${part.url}" />`
        case 'file': return part.url
        case 'json': return JSON.stringify(part.data)
    }
}

def render_part(part: dict) -> str:
    kind = part["type"]
    if kind == "text":
        return part["text"]
    if kind in ("image", "video", "file"):
        return part["url"]
    if kind == "json":
        return str(part["data"])
    return ""

Duplicate labels

If two User Result nodes share the same normalized label, publishing asks you to rename one of them. Explicit labels keep your runner and API contract stable across workflow edits.

Get Started

Concepts

Workflow

Output Parts

Output items

Part types

Reading parts in your code

Duplicate labels

What’s next

Lists & single values

API Reference

​Output items

​Part types

​Reading parts in your code

​Duplicate labels

​What’s next

Lists & single values

API Reference

Output items

Part types

Reading parts in your code

Duplicate labels

What’s next