Attaching files
On the Mac, click the attach button in the composer and choose Add photos & files, or drop files anywhere on the chat. Recent files lists what you attached recently. On iPhone, tap the attach button and choose Photo Library, Take Photo or Files.
Extraction happens when you attach, before you send. Everything uses Apple's frameworks on the device: PDFKit for PDFs, Vision for text recognition, and a built-in reader for Office files.
Supported files
| Kind | Extensions | What Kalon reads |
|---|---|---|
| Images | png, jpg, jpeg, gif, webp, heic, heif, tiff, tif, bmp | The image itself for vision models, and its text through OCR. |
| The text layer of each page. Pages with no text are rendered and read with OCR. | ||
| Word processing | docx, rtf | Document text. |
| Spreadsheets | xlsx, csv | Cell text. |
| Presentations | pptx | Slide text. |
| Markdown and JSON | md, markdown, json, jsonl, geojson, ndjson | The file as written. |
| Code and plain text | txt, log, html, css, xml, yaml, toml, sql, sh, swift, py, js, ts, rs, go, java, kt, c, cpp, rb, php and more | The file as written. |
Older binary formats such as doc, xls, ppt and pages attach, but no text is extracted from them. The iPhone reads PDF, docx, xlsx, pptx and text files, and png, jpg, jpeg, heic, gif and webp images.
Limits:
- Extracted text is capped at 120,000 characters per file.
- Images are scaled down before they are sent to a model: to 1,568 pixels on the long edge on the Mac and 1,024 on iPhone.
- Kalon stores small previews of attached images with the chat, not the original files.
Images and vision models
Models marked with the image glyph can see pictures. When one is selected, attached images go to it directly, on the device. With any other model, Kalon sends the text it recognized in the image instead, and tells you it did.
To force text recognition with a vision model, include the word "ocr" in your message.
Long documents
Short documents are included whole. When the attached text is over 8,000 characters, Kalon splits it into overlapping passages and sends the five that best match your question. A note under your message reports how many passages were searched and how many were used, and the reply cites them by number.
By default, passages are matched by keyword. For matching by meaning, turn on Semantic search for documents in Settings › Controls on the Mac, or Search documents by meaning in Settings › Documents on iPhone. It needs the 35 MB bge-small:en embedding model, which Kalon offers to download from that row. Both methods run on the device.
Files the agent creates
In agent mode, documents the agent writes appear in a tray at the end of the reply, with options to view the file, open it in its app, or show it in its folder.