Attachment context indicator (traffic light)
Each attachment chip has a color-coded border showing how much of your context it eats.
Short answer
Each attachment chip has a color-coded border showing how much of your context it eats. Neutral (default): fine. Amber: uses more than half of context. Red: over 90%. The exact percent and warning text appear on hover; below the strip, an aggregate badge sums all attachments.
Detailed explanation
Every attachment gets a rough token count on add (documents by their extracted text length, images by the OCR’d text length). This count is compared against your current model’s context length setting.
Thresholds:
| Attachment tokens / context | Chip border | Corner icon | Hover tooltip warning |
|---|---|---|---|
| Under 50% | Neutral (theme border) | none | just percent + tokens |
| 50-90% | Amber | ⚠ (amber) top-right | ”⚠ Model may forget earlier messages” |
| Over 90% | Red | ⚠ (red) top-right | ”⚠ Exceeds context - some content will be truncated” |
Under the whole attachment strip (not per chip), an aggregate badge shows the sum: Attachments: ≈N tokens (X% of context).
Per-chip vs aggregate. Each chip is colored based on its own share of context, not the combined share. Two 30% attachments each stay neutral individually, but the aggregate badge will show 60% of context in use - that’s when you notice you’re stacking too much.
What “truncated” means. If your total prompt (system + previous messages + this attachment + your new message) exceeds the model’s context window, older chat history gets compressed (see context compression), and if that’s still not enough, content gets dropped. Very long attachments push more history off, or get partially dropped themselves.
Why it matters
Without the traffic light you’d only find out about a too-big attachment after generation goes weird or the model complains. Border color + aggregate badge give instant visual feedback before you hit Send.
In Abliteration specifically
Border colors match the same amber/red palette used in the Hardware card’s “You can run” line and the fit-check dialog - consistent visual language: neutral = fine, amber = attention, red = problem.
The estimated token count uses:
- PDF: extracted characters / 4.0 (fallback: file bytes / 8 if extraction fails)
- TXT / MD: characters / 3.9
- DOCX: characters / 4.0
- Images: OCR’d text length / 4.0 (fallback: file bytes / 8 if OCR is empty)
- Additional rates for JSON (÷2.4) and CSV/TSV (÷1.75); a Cyrillic multiplier of 1.8 applies for those languages
Estimates are approximate - good enough for warning purposes.
Related
- Attachments - the main attachment feature
- Context length - the number thresholds are based on
- Attachment too big - options when red