Images and vision

Keiro accepts image input on Responses, Chat Completions, and Messages when the requested public model and account path support vision. Image input is for analysis and understanding; this page does not describe image generation.

On this page 1 of 6

Send an image with Responses#

Run a curl request Shell

printf 'Keiro API key: '
IFS= read -rs KEIRO_BEARER
printf '\n'

curl -sS https://api.keirolabs.ai/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "eb1-preview",
    "input": [
      {
        "role": "user",
        "content": [
          {
            "type": "input_text",
            "text": "Describe this image and transcribe any visible text."
          },
          {
            "type": "input_image",
            "image_url": "https://keirolabs.ai/assets/og-card.png",
            "detail": "auto"
          }
        ]
      }
    ]
  }' \
  -H @- <<<"Authorization: Bearer $KEIRO_BEARER"

Remote images must be reachable through an HTTPS URL. For private data, use a short-lived access URL approved by your organization's data policy or an inline base64 data URL. Never put credentials in the image URL.

Surface shapes#

Surface	User image content
Responses	`{"type":"input_image","image_url":"https://..."}`
Chat Completions	`{"type":"image_url","image_url":{"url":"https://..."}}`
Messages	`{"type":"image","source":{"type":"url","url":"https://..."}}` or a base64 source

For Responses, detail may be auto, low, or high. Use auto unless your workload has measured a need for another value.

Request limits#

Image requests fail with a caller-fixable 400 when they exceed these ingress bounds:

Limit	Maximum
Image parts per request	8
Remote image URL length	8 KiB
Encoded data per inline image	5 MiB
Total encoded inline image data per request	10 MiB

Inline images must be valid base64. Non-data image URLs must use HTTPS and must not contain surrounding whitespace.

Placement rules#

Keiro validates and routes the image reference but does not turn it into a public file store. Retention and sensitive-workload approval still follow the data-retention and subprocessor postures.

Put images in user message content.
Do not put image content in a tool message or function output.
Messages tool_result blocks support textual results, not image blocks.
Keep the text instruction next to the image so the requested analysis is explicit.

Capability errors#

A model that cannot preserve image input fails closed instead of dropping the image. If a request is rejected, confirm that the model is available to the key, reduce the image payload if it exceeds a bound, and contact support before changing a sensitive workload's model selection.