> ## Documentation Index
> Fetch the complete documentation index at: https://docs.inb.cc/llms.txt
> Use this file to discover all available pages before exploring further.

# Multimodal Input

> Passing images to chat completions

# Multimodal Input

Vision-capable chat models accept images inside a user message. The content becomes an array of typed parts instead of a single string. This guide covers the two ways to attach an image — by URL or as a base64 data URI — and how to mix text and images in a single turn.

## Two ways to attach an image

<CardGroup cols={2}>
  <Card title="Public URL" icon="link">
    Reference an image that the gateway can fetch over HTTPS. Best for static assets, CDN-hosted images, and pre-signed URLs.
  </Card>

  <Card title="Base64 data URI" icon="file-binary">
    Inline the image bytes as a base64 string. Best for ephemeral content, user uploads, and private assets.
  </Card>
</CardGroup>

<Tip>
  Prefer URLs when possible. The gateway can cache and reuse them across requests. Base64 is best reserved for images that change per request.
</Tip>

## URL example

Pass the image as a content part with `type: image_url`:

```bash theme={null}
curl https://api.getinfinityblue.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.4","messages":[{"role":"user","content":[{"type":"text","text":"What is in this image?"},{"type":"image_url","image_url":{"url":"https://example.com/cat.jpg"}}]}]}'
```

The URL must be reachable from the gateway. If your bucket is private, generate a short-lived pre-signed URL or switch to base64.

## Base64 example

Inline the bytes as a data URI:

```python theme={null}
import base64
from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.getinfinityblue.com/v1")
encoded = base64.b64encode(open("cat.jpg", "rb").read()).decode("utf-8")

response = client.chat.completions.create(
    model="gpt-5.4",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What is in this image?"},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded}", "detail": "high"}},
        ],
    }],
)
print(response.choices[0].message.content)
```

<Info>
  The `detail` parameter controls how much compute the model spends on the image. `low` is faster and cheaper; `high` reads small text and dense details; `auto` lets the model pick.
</Info>

## Multiple images in one turn

Add as many content parts as you need. The model reads them in order:

```json theme={null}
{
  "model": "gpt-5.4",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "Compare these two receipts and tell me which has the larger tip."},
        {"type": "image_url", "image_url": {"url": "https://example.com/receipt-a.jpg"}},
        {"type": "image_url", "image_url": {"url": "https://example.com/receipt-b.jpg"}}
      ]
    }
  ]
}
```

## Limits and gotchas

* Each model has a maximum number of images per request and a maximum resolution. Excess is rejected with `400 invalid_request_error`.
* Total request size — text plus encoded images — must stay under the per-request byte cap. For base64, encode the image first and check `len(b64) * 3 / 4` against the limit.
* When you mix text and images, keep the text part concise. The model reads both, but verbose prompts dilute the focus on the visual.

## Streaming and tools

Vision works with streaming and with tool calls. The same content array works in `stream: true` requests and in requests that include `tools`. The model emits text and tool deltas the same way it does for a text-only message, with the image already understood as part of the context.

<Warning>
  Never put sensitive images in a public URL. If the asset is private, use base64 or a pre-signed URL with a short expiration.
</Warning>