Attachments
You can send images, video, and documents along with a chat message. The agent's LLM reads or analyzes those files and uses them to answer.
Attach a file
- Open a chat.
- Start with
- Click the paperclip icon in the bottom-left icon group of the input area. It is the rightmost icon next to Conversation history (clock) and Agent tools (wrench).
- Pick one or more files from disk.
- Each file appears as a pending attachment chip inside the input area.
- Type a message, or leave the message blank for a file-only send, then click Send (arrow icon).
Heads-up. The paperclip is disabled until your conversation has at least one message. Hovering it shows "Start a conversation first". After the agent acknowledges your first message, the button becomes active without a page reload.
You can also drag and drop files onto the chat input once the conversation has started — same limits apply.
Pending attachments
Before sending, attachments appear as compact chips below the text field and above the toolbar icons, inside the input area.
Each chip shows:
- A file-type icon.
- The filename, truncated with
...if needed. - The file size in KB or MB.
- An × button to remove the file.
Images do not show thumbnails while pending. Image thumbnails appear after the message is sent.
If you attach multiple files, the chips wrap into a small grid. Click × on any chip to remove that file immediately. There is no confirmation prompt; if all chips are removed, the input area returns to its normal state.
With files queued, the Send button appears even if the text field is empty. When there are no files and no text, the same bottom-right position shows Voice mode (waveform icon) instead, if the agent supports voice.
Allowed types and sizes
| Category | Types | Max per file |
|---|---|---|
| Images | JPEG, PNG, WebP, GIF | 5 MB |
| Videos | MP4, WebM, QuickTime (MOV), AVI | 100 MB |
| Documents | PDF, TXT, CSV, HTML, Markdown, Excel (.xlsx and .xls), Word (.docx and .doc) | 10 MB |
Per message: up to 4 files and 150 MB total across all attached files.
When 4 files are pending, the paperclip disables until you remove one. If you select more than 4 files at once from the file picker, only the first 4 are added.
If a file is too big, the wrong type, or pushes you past the total size limit, you'll see a small toast like:
Unsupported file type:
<type>. This model supports: …
File too large:<name>. Maximum size for<category>:<size>.
Total size would exceed 150.0 MB limit.
Remove or swap the offender and try again.
After you send
After sending, attachments are shown inside your message bubble:
| Type | How it appears |
|---|---|
| Images | Inline thumbnail preview. Click it to open the full image. |
| Documents | Clickable filename link. Opens or downloads the file. |
| Video | Attached media link or preview, depending on browser support. |
This lets you confirm the upload was included with the message.
What the agent does with them
- Images — the agent's vision model looks at them. Useful for "what's in this picture", screenshots, photos of devices, hand-drawn diagrams.
- Documents — the agent reads the text. Useful for "summarize this", "find errors", "extract a table".
- Video — the agent's multimodal model can analyze short clips.
For attachments to work, the agent needs an LLM that supports the relevant capability, such as vision for images or multimodal input for video. If the model cannot handle the uploaded file, the dashboard may show an error or the agent may say it cannot read that file. Use an agent with the right model capability, or try a simpler format.
If you create or edit agents, use the admin dashboard's Compare text models table above the Model Selection dropdown to check model capabilities before choosing one. Look for the VISION column: a checkmark means the model can analyze image attachments; a dash means it cannot. You can search or filter the table by model name or provider to find a vision-capable option.
How upload works (under the hood)
When you pick a file, the dashboard:
- Asks the platform for a one-time upload URL (a presigned S3 URL).
- Uploads the file directly to storage.
- Sends a reference to your message when you click Send.
The file never sits in the browser longer than it has to, and it's bound to the conversation it was attached in.
Privacy
- Attachments are scoped to the conversation. Other users can't see them.
- They're stored server-side as long as the conversation is alive. Deleting the conversation removes them.
Tips
- Crop before sending. A tightly-cropped image of the relevant area beats a full-screen screenshot.
- One question at a time. "What's wrong with this code?" + one file works better than five files in one shot.
- If the agent can't read it, try a different format. PDFs of scanned images don't OCR well — re-export as searchable text if you can.