Repository navigation
Home
At a glance: 100% local · MIT licensed · 5 formats (PDF · DOCX · XLSX · PPTX · TXT) · English + Tiếng Việt
DiffLens is a local-first desktop app that compares two documents — PDF, DOCX, XLSX, PPTX or TXT — and produces a severity-banded change report with a full, human-readable audit trail. Paragraphs are matched by meaning, using multilingual embeddings that run entirely on your machine, so moved clauses are detected as moves instead of masquerading as mass delete-plus-insert noise. Money, date, percent and duration changes are extracted by deterministic rules that show which rule fired and the before/after values — the same two inputs always produce the same report, with no hallucination possible. An optional local LLM (Qwen, via llama.cpp) can explain any change in plain English or Vietnamese — but it can never alter the diff or the severity, and nothing you open is ever uploaded.
| Platform | Windows 10/11 x64 · macOS 14+ Apple Silicon · Linux x64 |
| Current version | 1.0.1 (see the Changelog) |
| License | MIT — third-party notices |
| Docs site | https://vibecoder880.github.io/difflens/ |
| Feature | What it does | Where |
|---|---|---|
| 5-format extraction | PDF (PDFium), DOCX (raw OOXML incl. tracked changes), XLSX (cells + formulas), PPTX (slide runs), TXT/Markdown | How-Diffing-Works |
| Semantic alignment | 3-pass: deterministic anchors → e5 embeddings + Needleman–Wunsch → move repair | How-Diffing-Works |
| Word-level diff | Inline added/removed spans, block text on every change | User-Guide |
| Deterministic intelligence | Money/date/duration/percent entities with VN scale words (triệu/tỷ), delta math, banded severity, risk lexicon | How-Diffing-Works |
| Audit trail | Every change records extraction provenance, alignment decision and rule evidence | User-Guide |
| AI explanations (optional) | Streaming natural-language summaries from a local Qwen GGUF via a loopback-only llama-server | AI-Explanations-Setup |
| OCR fallback (experimental) | Scanned PDFs through your locally installed Tesseract (vie language pack) |
OCR-experimental |
| Report export | Standalone HTML or Markdown with hostile-content escaping, golden-tested | Privacy-and-Security |
| Session history | Recent comparisons in a local SQLite store — list, re-open, delete | User-Guide |
| EN + VI interface | Full i18n with parity tests (210 keys) | User-Guide |
- Installation — per-OS installers, Gatekeeper/SmartScreen notes, SHA-256 verification.
- User-Guide — the four screens in depth: Compare wizard, Diff view, Downloads, Settings.
- How-Diffing-Works — extraction, normalization, 3-pass alignment, diff, heuristic rules.
- AI-Explanations-Setup — llama-server sidecar, model downloads, disk/RAM needs, troubleshooting.
- OCR-experimental — when it triggers, what to install, honest accuracy caveats.
- Privacy-and-Security — the local-first threat model and download integrity chain.
- Architecture — workspace layout, tech stack table, Tauri command/event surface.
- Building-from-Source — toolchain, feature flags, test tiers, CI structure.
- FAQ — twelve common questions.
- Not a cloud service — there is no server component and nothing to sign up for.
- Not a legal reviewer — severity bands come from rules, not judgment; see the disclaimer in How-Diffing-Works.
- Not able to diff folders or convert formats — exactly two documents of the same format per comparison.
None are shipped in-repo yet: the app requires a graphical environment and the documentation was finalized on a headless machine. Every screen is described in words (with ASCII layout sketches in the User-Guide) so you know exactly what to expect before you download a build.