Highlights
- Durable, incremental index — SQLite + WAL metadata with warm start; renames tracked by (dev, ino); batched commits; stale-document purge.
- Query language —
type:,ext:,in:,path:,modified:,size:,is:,lang:,name:,content:, negation, phrases. - Hybrid ranking — fuzzy + BM25 + static-vector tiers fused with RRF, frecency, recency and exact-match bonuses.
- Documents — PDF, DOCX/PPTX/XLSX, EPUB/ODT, RTF, encoding detection; streaming XML; extraction cache.
- OCR — image text via tesseract, indexed alongside content.
- UI v2 — paper/void themes, category chips, list + grid (row-first navigation), command palette, preview pane.
- Settings — Doctor (9 checks), Stats for Nerds, model-file download, factory reset with restart-safe wipe.
Known gaps
See docs/REMAINING.md. Biggest ones: the downloaded model table is not yet wired into the ranker, ONNX/image embeddings are pending, there are no tests yet, and packaging/CI is not set up.