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

# File attachments in chat

> PDFs, images, audio, video — inline or by file_id.

Chat completions accept rich content blocks: text, images, audio, files. You can pass them inline (URL or base64) or by reference to a [Files API](/api-reference/files) `file_id`.

## When to inline vs. upload

| You...                                              | Use                                        |
| --------------------------------------------------- | ------------------------------------------ |
| Send a one-off image/PDF you already have at a URL  | Inline `image_url` / `file` (data URI)     |
| Re-use the same document across many calls          | Upload once, reference the `file_id`       |
| Need workspace-wide access (multiple keys, members) | Upload — files are workspace-scoped        |
| Care about idempotency on retries                   | Upload + `Idempotency-Key`                 |
| Pass >20 MB                                         | Upload — gateway request body cap is 20 MB |

## Inline content blocks

`messages[i].content` can be an array of typed blocks:

```json theme={null}
{
  "role": "user",
  "content": [
    {"type": "text", "text": "What's in this image?"},
    {"type": "image_url", "image_url": {"url": "https://example.com/cat.jpg"}}
  ]
}
```

### Image (URL or base64)

```json theme={null}
{"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0..."}}
```

`detail` (optional): `"low"` (faster, cheaper, \~85 tokens) or `"high"` (default for most models).

### Audio (inline base64 only)

```json theme={null}
{
  "type": "input_audio",
  "input_audio": {"data": "<base64>", "format": "wav"}
}
```

`format`: `wav`, `mp3`, `pcm16`, `webm`. The model must support audio input — check `supportsAudioInput` on `GET /v1/models`.

### File — inline

```json theme={null}
{
  "type": "file",
  "file": {"data": "<base64>", "mime_type": "application/pdf", "filename": "report.pdf"}
}
```

### File — by `file_id`

```json theme={null}
{"type": "file", "file_id": "file_abc123..."}
```

The gateway resolves the id to bytes server-side, injects them into the provider call, and returns 400 if the id doesn't exist or belongs to another workspace.

## PDFs

Models with `supportsPdf: true` (Anthropic Claude, Google Gemini, OpenAI gpt-4o) read PDFs natively. For others, the gateway transparently converts each page to an image and prepends the extracted text — **you don't change a thing**, you just see a small `pdf_processing` line item on the next invoice.

```json theme={null}
{
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "Summarise the financials"},
      {"type": "file", "file_id": "file_pdf_xyz..."}
    ]
  }]
}
```

## Vision

Models with `supportsVision: true` accept arbitrary images. URL fetches happen on the gateway with a 10-second timeout — if your URL is slow or behind auth, prefer base64 or upload.

## Quick recipes

### Multi-image diff

```python python theme={null}
client.chat.completions.create(
    model="gpt-4o",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What's different between these two screenshots?"},
            {"type": "image_url", "image_url": {"url": "https://app.com/v1.png"}},
            {"type": "image_url", "image_url": {"url": "https://app.com/v2.png"}},
        ],
    }],
)
```

### Reusable contract

```python python theme={null}
contract = client.files.create(file=open("master_contract.pdf", "rb"), purpose="assistants")

def ask(question: str):
    return client.chat.completions.create(
        model="claude-sonnet-4-5",
        messages=[{
            "role": "user",
            "content": [
                {"type": "text", "text": question},
                {"type": "file", "file_id": contract.id},
            ],
        }],
    )

ask("What's the termination notice period?")
ask("List every penalty clause.")
```

The same `file_id` is referenced from many calls; you upload once.

## Limits

* Per-call inline payload: **20 MB** (sum of all base64 blocks)
* Per-file upload: plan-based (see [Plans](/billing/subscription-plans))
* Image dimensions: rescaled by the provider — no need to pre-resize
* PDF pages: practical cap \~100 (model context window limits dominate)

See [Files API](/api-reference/files) for upload, list, delete, download, and quotas.