Bug Report
Description
I’ve been testing Cotabby side-by-side with Cotypist using the exact same local Gemma model.
There are two noticeable differences:
Inline ghost text rendering: Cotabby’s suggestions often look like a floating overlay rather than text that naturally continues from the cursor. The baseline, spacing, font sizing, and/or positioning don’t seem to match the host text field as closely. Cotypist’s inline suggestions look almost indistinguishable from native text except for the lighter color.
Suggestion quality: Even with the same Gemma GGUF model, Cotabby’s completions are noticeably less natural and useful than Cotypist’s. This makes me wonder whether the difference could be related to prompt construction, context selection, sampling parameters, stopping rules, or post-processing rather than the model itself.
I’ve tried adjusting ghost text size, opacity, color, and display settings, but the inline positioning still feels noticeably less polished.
Steps to Reproduce
- Run Cotabby with Gemma as the local model.
- Enable inline suggestions.
- Open a normal text field in Chrome, Notes, or another supported application.
- Type several sentences and allow Cotabby to generate completions.
- Compare the alignment of the ghost text with normally typed text.
- Run Cotypist with the same Gemma model and type the same text.
- Compare both the visual alignment and the usefulness/naturalness of the generated suggestions.
Expected Behavior
Inline suggestions should appear as though they are a natural continuation of the host application’s text: same baseline, font metrics, spacing, and cursor position, with opacity being the main visual difference.
When using the same underlying model, I’d also expect suggestion quality to be reasonably comparable. It would be great to investigate whether Cotabby’s prompting/context/sampling pipeline could be tuned further for short autocomplete suggestions.
Environment
Screenshots
Submitted via tabby feedback form
Bug Report
Description
I’ve been testing Cotabby side-by-side with Cotypist using the exact same local Gemma model.
There are two noticeable differences:
Inline ghost text rendering: Cotabby’s suggestions often look like a floating overlay rather than text that naturally continues from the cursor. The baseline, spacing, font sizing, and/or positioning don’t seem to match the host text field as closely. Cotypist’s inline suggestions look almost indistinguishable from native text except for the lighter color.
Suggestion quality: Even with the same Gemma GGUF model, Cotabby’s completions are noticeably less natural and useful than Cotypist’s. This makes me wonder whether the difference could be related to prompt construction, context selection, sampling parameters, stopping rules, or post-processing rather than the model itself.
I’ve tried adjusting ghost text size, opacity, color, and display settings, but the inline positioning still feels noticeably less polished.
Steps to Reproduce
Expected Behavior
Inline suggestions should appear as though they are a natural continuation of the host application’s text: same baseline, font metrics, spacing, and cursor position, with opacity being the main visual difference.
When using the same underlying model, I’d also expect suggestion quality to be reasonably comparable. It would be great to investigate whether Cotabby’s prompting/context/sampling pipeline could be tuned further for short autocomplete suggestions.
Environment
Screenshots
Submitted via tabby feedback form