AI and Performance

AI and Performance

PICAZHU treats AI as an optional subsystem, not as a required part of basic browsing. The default fast path is AI off.

Core Rule

When AI is off:

• no AI queue should process media
• no visual model requests should run
• browsing, indexing, thumbnailing, and preview should stay as lightweight as possible

When AI is on:

• the app should show visible AI status
• the app should show visual-analysis progress
• provider readiness should be clear
• failures should be explicit instead of silent

Supported Provider Surface

• LM Studio local vision models through OpenAI-compatible chat completions.
• Ollama local vision models through Ollama chat API.
• Ollama Cloud through cloud endpoint and bearer API key.
• OpenAI vision through API key and selected vision-capable model.

Current AI Capabilities

• Runtime AI kill switch.
• Provider readiness/status checks.
• OCR extraction for searchable visible text.
• Vision tagging for images.
• Representative-frame analysis for videos.
• Tags tab for selected media.
• Search over filename, folder path, metadata, OCR text, captions, and AI tag text.

Not Complete Yet

• Embeddings generation.
• Vector storage.
• Hybrid semantic ranking.
• Production-grade API-key storage.
• Broad live QA across all cloud and local providers.

LM Studio Notes

Local VLMs are sensitive to context size. For qwen/qwen2.5-vl-7b, earlier failures were traced to requests exceeding the active context window. The app now keeps local vision requests compact by:

• using shorter prompts
• capping output tokens
• sending compact JPEG analysis images
• reducing analysis resolution and quality
• using generated previews for HEIC when needed

If LM Studio reports context-window errors, increase model context in LM Studio or reduce request size further.

Testing AI

1. Start the selected provider.
2. Confirm the model is vision-capable.
3. Open Settings and test the connection.
4. Turn AI on in the header.
5. Add or rescan a small folder first.
6. Watch AI progress and status.
7. Select analyzed media and inspect the Tags tab.
8. Search for terms visible in tags, OCR text, captions, or filenames.