Intelligence Engine Overhaul
Major AI reliability improvements — smarter model selection, template echo protection, and real-time thumbnail updates for nested maps.
Consistent Model Selection
Your chosen AI model is now respected across all generation modes — single, compare, and sub-maps. No more silent overrides.
Template Echo Protection
New validation layer detects when AI echoes prompt templates instead of generating real content. Auto-retries with model rotation for guaranteed results.
Live Thumbnail Refresh
Nested map thumbnails now update in real-time when generated — no more manual page refresh needed.
Codebase Cleanup
Removed experimental Warp and Mentor features. Deprecated deepseek-chat model removed with automatic fallback for users with stale preferences.
This release focuses on the invisible but critical infrastructure that powers every mind map you generate. We've rebuilt the model selection pipeline to be more predictable, more reliable, and fully respectful of your preferences.
Smarter Model Selection
Previously, when you selected a specific AI model in your profile, the system would silently override your choice for certain generation modes like Compare or deep-dive maps. This happened because the internal 'capability' system would discard your model when it decided a different capability (like 'reasoning') was needed.
Now, your model choice is always passed through first. The capability hint is still sent for auto-select fallback, but it never overrides an explicit user selection. This means consistent output quality across single mode, compare mode, and all sub-map generations.
Template Echo Protection
We discovered a rare but frustrating bug where some AI models would echo back the JSON template from our prompt instead of generating real content. The result was a mind map filled with placeholder text like 'Subtopic Name' and 'Category Name'.
The new template detection system checks every generation against a set of known placeholder markers. If 2+ markers are found, the generation is automatically retried with a different model — up to 3 attempts with automatic model rotation. The system prompt was also enhanced with explicit anti-template instructions.
Real-Time Thumbnail Updates
The nested maps dialog (Knowledge Navigator) now reflects thumbnail changes instantly. Previously, when you generated a new thumbnail for a sub-map, the card in the navigator would still show the old image until you refreshed the browser. The fix was a priority resolution change — the dialog now reads from the live `thumbnailOverrides` state before falling back to persisted Firestore data.
Cleanup & Deprecations
The experimental **Warp** (Perspective Warp) and **Mentor Roleplay** features have been fully removed. These were partially wired handlers with no clear product surface. The deprecated `deepseek-chat` model was also removed from the available models list, with a sanitization layer added to gracefully handle users whose saved preferences reference it.
