lllin000_PaperForge/CHANGELOG.md

Changelog

All notable changes to PaperForge.

---

1.5.8 (2026-05-19)

Memory Layer (New)

• SQLite + FTS5 database (paperforge.db) for fast paper metadata lookup
• paperforge memory build / memory status / search
• paperforge agent-context — structured bootstrap for AI agents
• paperforge dashboard — backed by SQLite (with file scan fallback)
• Automatic rebuild after sync (fast-path when index hash unchanged)
• Runtime health computation with degraded-status reporting
• JS-native memory state reader (file-based, no Python exec polling)

Vector Embedding (New)

• ChromaDB-backed vector storage for OCR fulltext
• paperforge embed build / embed status / embed stop
• paperforge retrieve — semantic search across OCR chunks
• API-only embedding (OpenAI-compatible), no local models
• Configurable API key/base/model via env, plugin settings, or .env
• --resume / --force build modes with persistent build state
• Auto-trigger after OCR via background subprocess

Embedding Package Refactoring

• Extracted vector logic from memory/vector_db.py into dedicated embedding/ package
• Old modules converted to deprecated forwarding shims
• Removed sentence-transformers dependency (API-only)
• Model change auto-detection: resume mode detects model switch and triggers full rebuild

Orphan Paper Cleanup (New)

• paperforge prune — detect and delete orphan workspaces (dry-run by default)
• sync --prune / sync --prune-force — integrate into sync pipeline
• Automatic orphan detection after every sync with Obsidian modal popup
• Click-to-toggle selection with paper metadata (title, authors, year, PDF, collection)

Sync Performance

• Export hash check skips rebuild when BBT files unchanged (~24s → 0.1s)
• canonical_index_hash in memory DB meta for fast rebuild skip
• Removed redundant refresh_paper calls from build loop

Plugin UI

• Tabbed settings: Installation + Features
• Runtime Health panel with manual sync button
• Memory Layer and Vector Embedding state machine UI
• Discussion card renders from discussion.md with MarkdownRenderer
• i18n: full Chinese/English translation for Features tab + orphan modal
• Vector DB config labeled as Experimental

Discussion System

• Changed from JSON to plain Markdown (discussion.md)
• Plugin renders last 3 Q&A pairs with collapsible long answers

Bug Fixes

• get_collection() no longer silently deletes collection on error
• nopdf OCR status resets to pending when PDF re-appears
• Embed chunk_count now reports actual DB count, not incremental count
• Plugin: embed build survives settings tab close
• Plugin: fix runtime health sync path for first-launch migration

Infrastructure

• release.yml / publish.yml — fixed tag pattern to match [0-9]* (no v prefix)
• Pre-commit JS syntax hook added

---

1.5.7 (2026-05-19)

Recovery release after 1.5.6 PyPI name conflict. All features listed under 1.5.8.

• Model change auto-detection in embed resume mode
• Orphan modal i18n completed (Chinese/English)
• CHANGELOG.md added

---

1.5.9 (2026-06-17)

OCR Tail Rendering Fixes

• _reorder_tail_run: Added skip_section_grouping parameter for tail pages with reference items but no explicit "References" heading. These pages now emit blocks in column-sorted order with refs grouped at end, preserving natural reading order.
• _order_tail_blocks: Per-page detection of ref heading presence to activate the new path.
• _normalize_backmatter_roles_after_boundary: Removed forced body_paragraph → backmatter_body conversion inside backmatter region (body paragraphs attach naturally via geometric ownership).
• Synthetic ref section (Phase 2.5): Creates synthetic ref_section when ref_items exist but no ref_heading, preventing ref scattering as orphans.

For upgrade: pip install --upgrade paperforge