simonandCursor
1e1111c90f
Safety snapshot of mobile work before merging upstream chat-v2: - Qortino AI: QDN pack discovery/download (Q-Share ids), external app storage, install validation, teach/report Q-Mail, PDF thumbs+zoom - Android share target "Quitter" with native ShareReceiver plugin - Categorized device saves (Images/Videos/Audio/Documents/Apps/GO state) - createNamedFile helper: cordova-plugin-file clobbers global File Co-authored-by: Cursor <cursoragent@cursor.com>
31 lines
1.3 KiB
Markdown
31 lines
1.3 KiB
Markdown
# Qortino vs reference answers
|
||
|
||
## What was wrong before
|
||
|
||
Offline Qortino used a small Qwen GGUF + hash RAG over ~223 chunks, with Fast mode truncating context to ~220 characters and ~80 output tokens. That made answers thin, sometimes confused “Playing Qortal” metaphor with “Qortal is a game,” and missed practical detail.
|
||
|
||
## What we did
|
||
|
||
1. Authored **1600** high-value Q&As (200 × 8 categories) as golden reference answers.
|
||
2. Imported them into:
|
||
- `docs/faq/*-faq.md` → rebuilt knowledge pack (**1823** chunks, 1600 FAQ units)
|
||
- `src/qortinoAgent/goldenAnswers.json` → fuzzy exact-match short-circuit
|
||
3. Raised offline RAG budgets (Fast 2×900 chars, Max 3×1400) and maxTokens (Fast 160 / Max 280).
|
||
4. Prompt tweak: stay faithful to retrieved `A:` FAQ lines.
|
||
|
||
## How answers improve
|
||
|
||
| Path | Behavior |
|
||
|------|----------|
|
||
| High fuzzy match (≥0.58 Fast / ≥0.64 Max) | Return golden answer directly (same quality as reference) |
|
||
| No match | RAG over FAQ + handbooks with much longer chunks → model paraphrases stronger notes |
|
||
| QDN mode | Unchanged live scan; pack still available as backup |
|
||
|
||
## Categories
|
||
|
||
qortal · programming · survival · health · gardening · life-skills · philosophy · earth-changes
|
||
|
||
## Rebuild
|
||
|
||
See `golden/README.md`.
|