Phase 4: Fix critical bugs, improve document parsing, add vision OCR
- Fix SSE stream 500: use async_session_maker inside StreamingResponse generator
(Depends session closes when endpoint returns, before streaming starts)
- Fix template application: store template_name in prepare endpoint so worker
uses the selected custom template instead of defaulting to "general"
- Fix OverlayLoader: replace loading.gif with HamsterLoader component
- Fix parse_mode default: change from "slides" to "layouts" to avoid 70+ layouts
- Update Gemini Flash model to gemini-3.1-flash-image-preview
- Improve DOCX parsing: python-docx for structured table extraction, OCR enabled
- Add vision-based image text extraction via Gemini for uploaded images
- Add LayoutParser integration for slide layout structure analysis
- Add Phase 4 MVP features: transfer ownership, URL input, follow-up questions,
attachment-to-slide mapping, content router
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
e8295d6e71