Attaching files and images

Short answer

Click the ”+” button on the input pill to attach files or images. Supported: PDF, TXT, MD, DOCX for text; PNG, JPG, WEBP for images. Images are OCR’d locally into text (they’re not “seen” by a vision model). Each attachment shows a 72×72 chip in a horizontal strip, in the order you added them. Max 5 attachments, 50 MB per file.

Detailed explanation

Attaching. Click the ”+” button (glyph +) on the input pill, then choose “Add Files” or “Choose Image” from the sheet that opens. This routes to the Windows file picker.

Drag-and-drop of files onto the chat area is not supported - file attachment happens only through the ”+” button. (Drag-and-drop of GGUF model files onto the Models screen for adding models is a separate feature.)

File types supported for text extraction:

• PDF - extracted using local PdfPig parser
• TXT / MD - read directly
• DOCX - extracted with OpenXml parser

For images (PNG, JPG, WEBP), the image is passed through Windows OCR to extract any text visible in the image, then that text is sent to the model along with your message. There is no vision model - a model that doesn’t understand images natively will only “see” whatever text OCR pulled out. If OCR finds no text (e.g. a photograph of a landscape), you’ll see “No text found in image … nothing to send to model”. OCR needs a Windows language pack installed for languages other than the system default.

Limits:

• Maximum 5 attachments per message
• Maximum 50 MB per file

Unified chip strip. Every attachment renders as a 72×72 chip in a single horizontal row above the input, in the order you attached them. Image chips show a thumbnail; document chips show a bordered box with the file type label (PDF, TXT, DOCX, MD) in the top-left corner. Text label color adapts to theme (dark on light, light on dark).

Removing. Click the ✕ on any chip to remove it before sending.

Token cost display. Under the strip, an aggregate badge shows total attachment tokens as a percentage of context (e.g. Attachments: ≈2,400 tokens (30% of context)). Individual chip borders change color based on how much of your context that single attachment eats:

• Default border (under 50% of context) - normal
• Amber border with ⚠ icon (50-90%)
• Red border with ⚠ icon (over 90%)

The exact per-chip token count and warning appear when you hover over the chip (tooltip). See Attachment traffic light for what the colors mean.

Where attachments live in the chat. After sending, attachments appear as chips inside your message bubble (same 72×72 layout, same colors). The extracted text is included in the prompt sent to the model, not shown separately in the UI.

Why it matters

Attachments make Abliteration useful beyond pure chat - you can point the model at a document and ask questions. But attachments consume context aggressively - a 40 KB PDF can eat 10K tokens, which fills a 32K context 30% in one attachment.

In Abliteration specifically

Text extraction is local - no cloud upload. Everything happens on your PC.

Token counters use rough conversions (PDF: extracted characters / 4; MD: characters / 3.9; DOCX: characters / 4). Not exact but calibrated to be within useful range.

If you attach a very large file that would exceed context even at 90%, Abliteration won’t block sending - you’ll see the red warning and can decide to trim or proceed. Text beyond the model’s context window will be truncated when context compression fires.

Multiple attachments stack in one row. If they exceed the input width, the strip scrolls horizontally.

Related

• Attachment traffic light - the color-coded context indicator
• Context length - the ceiling attachments push against
• Attachment too big - what to do when red