Repository navigation
FAQ
1. Is my data uploaded?
No. Parsing, diffing, embeddings, severity scoring and history all run on your machine. The only network code in the entire app is the Downloads manager (core/src/download/), used only when you explicitly fetch a model or the llama-server sidecar — each SHA-256-verified against a pinned manifest. There is no telemetry, no crash reporting, no analytics, and a CI job (egress-audit) fails the build if any HTTP client code ever appears outside the download module. See Privacy-and-Security.
2. Why is the first alignment slow? Two costs are paid once: the e5 embedding model loads into memory on first use, and the first compare embeds every block of both documents. Subsequent comparisons of similar documents are much faster because only the model load was one-time. The optional LLM has the same pattern — the first explanation after launch includes llama-server spawn + GGUF load; later ones reuse the running sidecar.
3. Which languages does DiffLens support?
The UI is fully bilingual: English and Tiếng Việt (210 i18n keys with automated parity tests against the backend's rule label keys). The alignment engine uses multilingual embeddings chosen specifically for strong Vietnamese quality alongside English; heuristics parse Vietnamese entities natively (triệu/tỷ scales, ngày … tháng … năm … dates). Text in other languages still diffs fine at the word level, but entity/severity intelligence is VN+EN.
4. Does OCR work on handwriting? No. The Tesseract path targets printed text; handwriting recognition is beyond both Tesseract generally and this integration specifically. Even for printed text, see the honest caveats in OCR-experimental — expect diacritic errors on Vietnamese scans.
5. Can I diff two folders? Not in v1. A comparison is exactly two documents of the same format. Folder/multi-file comparison is a conceivable future feature, not a current one.
6. Where is my history stored?
In a local SQLite database — difflens.sqlite3 inside the OS app-data directory (%APPDATA%\app.difflens.desktop on Windows, ~/Library/Application Support/app.difflens.desktop on macOS, ~/.local/share/app.difflens.desktop on Linux). Deleting an entry in the UI removes the row permanently. Nothing syncs anywhere. Downloaded models/sidecar live alongside it in bin/ and models/ (see AI-Explanations-Setup).
7. Why is severity deterministic — why not let the LLM rate it? Because auditability beats cleverness here. Bands come from rules with published point values (money 35, date 30, …) plus a fixed lexicon, so the same documents always produce the same report and every severity can be explained ("why this severity" in the detail drawer). An LLM verdict would be non-reproducible, unauditable, and could hallucinate. The LLM's role is deliberately limited to explaining a change; it can never alter results or severity — that is a hard architectural rule, enforced in review.
8. What licenses do the models have? multilingual-e5-small: MIT. Qwen3-4B-Instruct-2507 and Qwen2.5-3B GGUFs: Apache-2.0. llama-server (llama.cpp b12503): MIT. PDFium: BSD-3-Clause. Everything is permissive; the full inventory with pinned versions is THIRD-PARTY-NOTICES.md. No AGPL/GPL components exist in the stack.
9. Are Intel Macs supported?
No. v1 ships macOS builds for Apple Silicon only (aarch64 .dmg). Windows x64 and Linux x64 are supported.
10. Is there an auto-updater?
No updater is wired in v1 — updates are manual: download the new installer from Releases (and verify its SHA-256). Note the release pipeline already generates Tauri updater signatures infrastructure-wise (TAURI_SIGNING_PRIVATE_KEY in CI), but the app does not phone home or self-update, and there are no plans to add telemetry-style update checks.
11. How do I verify a release?
Every release publishes checksums. Download checksums.txt plus your installer, then sha256sum -c checksums.txt (Linux/macOS) or PowerShell Get-FileHash comparison (Windows). Full walkthrough including per-OS commands: Installation. The installers themselves are unsigned at the OS level (no Developer ID / Authenticode in v1) but are built by a public, SHA-pinned GitHub Actions workflow.
12. How do I cite DiffLens? It's MIT-licensed open source — cite the repository and version:
@software{difflens,
author = {vibecoder880},
title = {DiffLens: local-first document diffing with semantic alignment},
url = {https://github.com/vibecoder880/difflens},
version = {1.0.1},
year = {2026}
}