# AI & LLM uploads (/docs/guides/ai-uploads)



## Chat attachments [#chat-attachments]

<ComponentPreview name="chat-attachment" />

A compact composer for chat interfaces:

* Attach by clicking, dragging onto the composer, or **pasting screenshots**.
* Files upload while the user types; sending waits for them.
* Images are downscaled to 1568px WebP, about what vision models work with, so
  attachments upload faster without changing answers.

Pass the stored objects to your model instead of base64 payloads:

```ts
onSend={async ({ text, attachments }) => {
  await fetch("/api/chat", {
    method: "POST",
    body: JSON.stringify({ text, files: attachments.map((item) => item.result) }),
  })
}}
```

On the server, give the model a signed URL (or fetch the object and pass its bytes) for
each file.

## Document ingestion (RAG) [#document-ingestion-rag]

<ComponentPreview name="ai-document-upload" />

`DocumentIngest` follows each document through `upload → parse → chunk → embed → index`.
The stages are driven by your backend through `process`:

```ts
const ingest: IngestProcess<StoredObject> = async (item, { setStage, signal }) => {
  const { jobId } = await api.ingest(item.result!.key, { signal })
  for await (const event of api.watch(jobId, { signal })) {
    setStage(event.stage) // "parse" | "chunk" | "embed" | "index"
  }
}
```

Throw to fail the item; it can be retried from the list.