Attachments
Files are read when you attach them and added to your message when you send. The most common surprises are size caps, file types Loach can’t decode, and images the model doesn’t see. See Attachments for how it all works.
“… is larger than 20 MB” when attaching a file
The message box refuses the file with that message under the input.
There’s a 20 MB cap per file, whether you drop, paste or pick it. Split large logs, or paste the relevant section as text instead.
“Couldn’t read …” when attaching a file
The file doesn’t attach, and the message box shows “Couldn’t read name:” followed by a reason (in Private Chat, just “Failed to read file”).
Loach couldn’t parse the file. The usual causes are a damaged file, a password-protected PDF or DOCX, or a file whose extension doesn’t match its contents. Remove the password or re-export the document (for example with Print → Save as PDF), then attach it again.
The PDF came in empty or with garbled text
The model says it can’t see the document, even though you attached a PDF.
Loach extracts text from PDFs, but scanned PDFs (image pages with no text layer) have no text to extract - the model only gets a note saying so. Run the PDF through an OCR tool first, paste the relevant pages as plain text, or export the pages as PNG or JPEG images and attach those to a vision-capable model.
The model can’t read my Word document
A .doc file attached, but the model says it only knows the file’s name.
Only .docx is extracted, not legacy .doc. A .doc still attaches, but as a file the model is told about by name only. Open it in Word or LibreOffice and Save As → Word Document (.docx).
A text file’s accented letters come through garbled
A text or code file attaches, but the model sees accented or non-Latin letters as the wrong characters or as �.
Loach reads text files as UTF-8 or UTF-16, with or without a byte-order mark. A file in neither is taken to be in the legacy Windows code page of your system language: Windows-1250 for Polish, Czech, Slovak, Hungarian, Croatian, Slovenian and Romanian, Windows-1252 for every other language. A file saved in some other encoding - Cyrillic Windows-1251, Shift-JIS, or a Central European file on an English system - comes out with the wrong characters. Re-save it as UTF-8 (most editors offer that under Save As or an encoding menu) and attach it again.
A long document’s chip says “truncated”
A long PDF or text file was attached, its chip carries a truncated pill, and only part of it reached the model.
Per-file cap is 200,000 characters; total inlined content per message is 500,000 characters, counting your prompt and fetched pages too. Files past that per-message budget are clipped or left out without a truncated pill - only the model is told. Either:
- Trim the document to the parts you actually need, or
- Send several focused messages, each with a different excerpt.
A PDF I sent shows “Preview not available”
Clicking a PDF chip on a sent message opens a file card instead of the pages, and Save gives you a .txt file.
In a regular chat, Loach keeps a PDF’s original bytes only while it sits in the message box; the sent message stores the extracted text. Preview the PDF from its chip before you send, and keep your original file - Save on a sent PDF (or DOCX) writes the extracted text.
Dropping a file onto Private Chat does nothing
The cursor shows a no-entry sign and nothing attaches.
Expected: dropping files is switched off while Private Chat is open, so a file can’t land in the regular chat behind it. Use the + button in Private Chat’s message box instead. Drops are refused the same way while a dialog - Settings, the snippet editor, Import context - is open over the chat; close it first.
Images don’t seem to be working
You attached a PNG/JPEG but the model can’t describe it.
The model needs vision capability - check the model’s page in the Ollama library or the capabilities that ollama show <model> lists. Switch to a vision-capable model (Llava, Llama 3.2 Vision, GPT-4o, etc.) and re-send. Also check that:
- The image is PNG, JPEG, WEBP or GIF. Other formats (BMP, TIFF, HEIC, SVG) attach as files the model only knows by name; convert them first.
- You attached it to this message. Images go to the model only with the message they are attached to, so for a follow-up question about the same picture, attach it again.
The model misses details in an image
A vision model can’t read small text in a screenshot, or overlooks fine detail in a photo.
Loach scales images down before it sends them - by default to at most 1024 px on the longest side, since a full-size photo costs thousands of image tokens and many local models read it no better. The original stays in your chat. To send more detail, open the chat’s Parameters panel and, under Images, pick Scale down with a larger size, or Original. A Space sets this for its chats on its Settings tab, and the app-wide default - the one Private Chat uses - is Settings → Features → Image scaling; Library images can also follow Settings → Library → Images. See Image scaling.
Pasting from Word or Excel gives text, not a picture
You copied cells from Excel or a paragraph from Word, and the message box got the text instead of an image of it.
Office apps put a picture of the selection on the clipboard next to its text. When the clipboard holds text alongside nothing but images, Loach pastes the text - what the model reads best - and ignores the picture. To send the picture, take a screenshot, or save it as an image file and attach that.