A free, private reverse dictionary using sentence embeddings, 1-bit quantization, and static model inference - zero server-side compute.
- Full mode (desktop): Asymmetric retrieval: definitions encoded offline by mxbai-embed-large-v1 (1024d, MRL-truncated to 384d), queries encoded at runtime by mdbr-leaf-mt (22M params) via Transformers.js. Two-stage scoring: 1-bit binary ITQ Hamming first-pass (~13ms) + int3 reranking of top-500 candidates.
- Full mode (mobile): Same model, pure 1-bit binary ITQ scoring only. ~30 MB total download.
- Lite mode: Knowledge-distilled static embeddings (256d, fine-tuned from mxbai-embed-large via Model2Vec). Sub-1ms queries via WASM or pure JS. Automatic fallback when ONNX model can't load (e.g. iOS).
| Mode | Config | MRR | Hit@1 | Hit@6 |
|---|---|---|---|---|
| Full | binary + int3 rerank (desktop) | 0.644 | 37/67 | 52/67 |
| Full | pure binary ITQ (mobile) | 0.563 | 30/67 | 52/67 |
| Lite | distilled-mxbai fine-tuned, int4 | 0.566 | 33/67 | 42/67 |
370,000+ visible definitions (576,000 total entries incl. hidden embedding-only) from six public-domain / open sources: Open English WordNet 2025 (CC BY 4.0), Webster's 1913, GCIDE Webster portion, Century Dictionary (1889-1911), Chambers's Twentieth Century Dictionary (1908), and curated LLM-augmented entries (CC0). Enriched with Moby Thesaurus synonyms. Wiktionary (CC BY-SA 3.0) and ConceptNet 5.7 (CC BY-SA 4.0) used at build time for quality signals and embedding enrichment only; not redistributed.
WordFor keeps three license lanes strictly separated (full table: LICENSE_SOURCES.md):
- Clean core — only public-domain / CC0 / CC BY definition text ships in
data/words.json(with attribution on about.html). - Build-time signals — share-alike / copyleft / proprietary sources (Wiktionary, ConceptNet, full GCIDE, Open Roget's, …) shape quality scores and hidden embeddings but are never shipped verbatim.
- Optional share-alike / GPL packs — anything copyleft could only ever be shipped as a separate, correctly-licensed artifact, not merged into core.
The build fails (build/audit_sources.py, build/audit_modern_sources.py) if any non-redistributable source leaks visible text, and build/verify_sources.py records a SHA-256 provenance snapshot of every ingested source. Search-time word-family grouping (run/ran/running → run) comes from build-time, license-audited lemma maps (build/build_lemma_families.py), not runtime guessing; negative prefixes are never collapsed (unhappy ≠ happy).
Static files served from GitHub Pages through Cloudflare CDN. GoatCounter for cookie-free analytics. No personal data collected.
© 2025 Zeeshan Khan Suri. Licensed under CC-BY-NC-ND-4.0.