Skip to content
Decisio

DocsGuides

2 min read · updated

Images

Send up to two images with a request; a second engine on the same card reads them, and the answers have the same shape.

A request can carry images beside its text. Image input is served by a second engine on the same card, so it needs a server started for it, and it was measured with the Qwen base only.

Figure · Image input

Images ride in the request; a second engine on the same card reads them.

ImajevBench v2.0-lite, the 230 answerable items, on the Qwen base: 0.791. It was measured with the Qwen base only.

Source: the README and docs/api.md at the pinned release.

TL;DR

Section titled “TL;DR”

On a GPU server started with --image-model, this script sends one image as a data URL with a question about it. It is a shape to adapt, not a command to run: it needs that server and an image of yours.

import base64
import json
import urllib.request


with open("photo.jpg", "rb") as f:
    image = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode()


body = {
    "state": "A photo of a delivered parcel.",
    "images": [image],
    "questions": {"damaged": {"type": "noul", "instructions": "Is the parcel visibly damaged?"}},
}
request = urllib.request.Request(
    "http://127.0.0.1:8000/v1/systemone",
    data=json.dumps(body).encode(),
    headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(request) as response:
    print(json.load(response)["answers"]["damaged"]["noul"])
# the probability of yes, as for a text-only request

A shape to fill in, not a command to run.

WordMeaning
image enginea second engine on the same card, started with --image-model, that serves the requests that carry images
data URLan image written into the request as data:image/...;base64,...
routewhich engine served the request; x-decisio-route is image for these

1. Start a server for images

Section titled “1. Start a server for images”

--image-model names the full multimodal checkpoint and starts the second engine, with its share of the card (--image-gpu-memory-utilization, 0.5 by default; the measured shares were 0.49 beside the text engine at 0.47). --one-engine loads only the image engine and serves text requests on it too, for a card that cannot hold both, and changes the text route’s class. A Mac’s MLX backend refuses the image route.

2. Send the images

Section titled “2. Send the images”

Up to 2 images per request, PNG, JPEG or WebP, up to 20 MB each, as data URLs in images, as data:image/... URIs anywhere inside the state, or as a multipart/form-data request whose request field holds the JSON body and whose image, images or image[] fields hold the files. x-decisio-route: text or image chooses the route explicitly.

3. What was measured

Section titled “3. What was measured”

On ImajevBench v2.0-lite, 254 labelled items, the Qwen base scored 0.717 on all items and 0.791 on the 230 answerable ones; with a “can’t tell” option offered, 0.756 and 0.813, with 5 of the 24 unanswerable items right and 2 false abstentions; the median request took 166 ms. The image route is not tempered, and it keeps the compact prompt layout.

What next

Section titled “What next”