Created: 02-07-2026 · Last updated: 06-07-2026
Status: pre-alpha — API runs locally, nothing is deployed yet. Phase 1 (data + API against real MW/PWG/AP90 data, 115/115 tests green) is complete; the P2 static-cache generator for the public GitHub Pages tier is built (released as v0.5.0). P3 (evidence layer) and P4 Wave K1 (Cologne inflection-table ingest) are also done, released together as v0.7.0. What remains for a first public URL is MG's deploy step — see PHASE1_PLAN.md and .ai_state.md for the exact resume point; no file in this repo claims to be "ready" unless its check actually runs.
A translator-first Sanskrit dictionary: every major dictionary on one page, every sense anchored to the scanned print, every word findable as it actually appears in a text — open, API-first, versioned, citable.
kosha is the working codename and deployment path (samskrtam.ru/kosha);
the public brand is Gasuns Sanskrit Dictionary — which also sidesteps the
name collision with Ambuda's vidyut-kosha. Naming and identity review:
POSITIONING.md ·
one-page distillation:
POSITIONING_SUMMARY.md.
Sanskrit NLP data & tools directory — the project's public entry point for Sanskrit computational linguistics: our openly-licensed derived datasets (downloadable) plus the external stacks it builds on, with what-each-does / how-to-call / license. Live at gasyoun.github.io/kosha/directory, rendered by scripts/build_directory.py from data/manifest/datasets.json
- data/manifest/external_tools.json (the single sources — no facts hand-copied into the page).
Fast lookup over the Cologne Digital Sanskrit Dictionaries — Monier-Williams (MW), the large Petersburg lexicon (PWG), and Apte 1890 (AP90) first, with the rest of the collection to follow. Primary audience: translators, then learners, then scholars.
- Multi-dictionary view — one headword, all dictionaries collapsed onto one page.
- Scan-anchored print truth — per-entry links to the scanned page of the printed edition, so the digital text is always one click from its source.
- Meet the form, not the lemma — diacritic-free typing (
krsnafinds kṛṣṇa), inflected-form lookup, and eventually paste-anything: a sandhied, compounded verse segmented at query time. - Evidence-graded entries — frequency bands and attestation from the Digital Corpus of Sanskrit: what a second-year should memorise vs look up and forget.
- Generated paradigms + a compact grammar token modelled on Zaliznyak's index from the Russian lexicographic tradition.
- Trilingual glosses — English (MW) · German (PWG) · Russian (pwg_ru translation layer) side by side; unique worldwide.
- Offline PWA — the static cache tier makes offline lookup nearly free.
- Citable, not just browsable — persistent sense-level IDs
(
dict.L.senseN@version), versioned data releases mirrored on Zenodo with DOIs; a citation from 2026 still resolves in 2036.
The formula, distilled from a twelve-platform live survey (COMPARISON.md, 02-07-2026): meyer's collapse + Heritage's morphology + DCS's evidence + Logeion's sidebar + Cologne's corrections loop + the trilingual DE/EN/RU layer. No existing project occupies that intersection.
Full plan with exit checks: IMPLEMENTATION_PLAN.md · scenarios it must serve: USE_CASES.md (UC1–UC13).
| Phase | Deliverable | State |
|---|---|---|
| P1 Data + API | kosha.db (MW+PWG+AP90) + API v1 running locally, tests green |
done — 115/115 tests green (see D5_MEASUREMENTS.md) |
| P2 Public alpha | first live URLs: static lookup on GitHub Pages + API on samskrtam.ru | generator built (v0.5.0) — next: MG deploy |
| P3 Evidence layer | DCS frequency badges, corpus example per lemma | done (v0.7.0) |
| P4 Forms & grammar | paste-anything segmentation, paradigm tables, grammar token | Wave K1 done (nominal inflection ingest + analysis API, v0.7.0) — K2/K3 + verb conjugations still open, see ROADMAP_INFLECT_2026_2027.md |
| P5 Advanced UI | the project's own UI at samskrtam.ru/kosha | gated on P2 deploy (K2/K3 in progress) |
| P6 Trilingual RU | Russian gloss layer beside DE/EN | gated on human review + rights |
| P7 v1.0 citable | DOI'd data release, dumps, "Cite" everywhere | gated on P2+P3 |
Optimistic elapsed total to v1.0 ≈ 4–6 weeks, dominated by human-side gates (deploys, design sign-off, translation review), not coding time.
| What | License | Why |
|---|---|---|
| Code | CC BY-NC 4.0 — LICENSE.md | attribution required, non-commercial |
| Data releases | CC BY-SA 4.0 — LICENSE-DATA.md | inherited from Cologne's ShareAlike, which does not permit adding a non-commercial restriction to derived data |
Which dictionaries are included? MW, PWG, and AP90 first — deliberately, because their page-reference formats are the three hardest cases, so the pipeline generalises. The lemma spine (the union headword index, 323,426 rows) already spans the whole Cologne collection, so further dictionaries are an ingestion task, not a redesign.
Do I need to type diacritics?
No. krsna finds kṛṣṇa. IAST, SLP1, and Devanagari input are all planned,
with auto-detection of the input scheme — the bar set by
sanskritdictionary.com.
Can I paste a whole sandhied verse? That is the P4 flagship: query-time segmentation of sandhied, compounded, inflected text, turning the dictionary into a reading companion. Until P4, lookup is by headword and inflected form.
Will it work offline? Yes — from P2 the static cache tier ships as a progressive web app: repeat lookups work with no connection.
What about Russian? P6 adds a Russian gloss layer beside the German and English, built on the pwg_ru translation project. It is explicitly gated on human review of the translated cards and on rights clarification for the Kochergina layer — it ships when those clear, not before.
Can I cite an entry in a paper?
That is a founding requirement, not an afterthought: every sense has a
persistent ID (dict.L.senseN@version), data releases are versioned and
mirrored on Zenodo with DOIs from the first citable release, and an old
citation resolves to the exact version it named — in a browser, no tooling.
Why do entries link to page scans? Because the digital text is a transcription and transcriptions drift. Every entry links to the scanned page of the printed edition, so the print remains the court of appeal — the same discipline the Cologne project itself follows.
Why another Sanskrit dictionary site? Because the composite doesn't exist. Each ingredient exists somewhere — multi-dictionary collapse (michaelmeyer.fr), form lookup (sanskritdictionary.com), morphology (Heritage/Ambuda), corpus evidence (DCS) — but no site combines them, and none is simultaneously open-source, API-first, versioned, and citable. The full survey is in COMPARISON.md.
How is this different from michaelmeyer.fr/sanskrit? Meyer's site is the closest thing to "Logeion for Sanskrit" — 41 dictionaries on one page, per-sense scan links for 19 of them, remarkable speed. It is also read-only, closed-source, and single-maintainer. This project adds the open API, morphology, corpus evidence, the Russian layer, and citability — and treats meyer's collapse as the bar to clear, not a rival to displace (see RELATIONS.md).
How is this different from sanskritdictionary.com?
That site is the feature benchmark: input auto-detection, sandhi:/root:
operators, per-sense permalinks. But it is closed to programmatic use
(Cloudflare-walled) and unversioned. This project is API-first and open by
construction.
Is this competing with Cologne / CDSL? No — Cologne is the substrate, not a competitor. The work runs on three tracks: (1) correcting the source texts upstream in csl-orig, (2) improving the shared Cologne tooling for everyone, (3) this project — the opinionated integration layer that doesn't belong in Cologne's conservative frontend. Corrections found here flow back upstream.
What about Ambuda / vidyut?
The strongest modern neighbour, and a dependency rather than a rival: this
project consumes the vidyut toolkit
(paradigms, form lookup fallback) and never re-implements it. The public
brand "Gasuns Sanskrit Dictionary" exists partly because vidyut-kosha
already owns the kosha name in that ecosystem.
Can I use it today? Not at a public URL yet. The API runs locally against real MW/PWG/AP90 data (Phase 1 complete, 115/115 tests green) and the P2 static-cache generator is built; the first public URL arrives once MG runs the deploy step.
Why is the repo all Markdown? Deliberately. An audit of this project's earlier planning corpus found fabricated figures (invented latency numbers, wrong attributions), and the lesson became policy: plan first, write the quality gates before the code (EVAL_PLAN.md), enumerate the risks (RISKS.md), and never mark anything done without a runnable artifact. The planning layer was finished and released as v0.2.0 before the first real line of application code.
When will something be usable? P1 (data + API, local) is done. The P2 static-cache generator is built; the first public URLs come as soon as MG runs the deploy step. The optimistic path to a citable v1.0 is 4–6 weeks from there, and the long poles are human gates, not code.
Can I use the data? Yes — data releases are CC BY-SA 4.0: use them, including commercially, with attribution and share-alike. This is inherited from Cologne's own licensing; a non-commercial restriction legally cannot be attached to data derived from BY-SA sources.
Why is the code non-commercial but the data isn't? Because the two have different ancestors. The data derives from Cologne (BY-SA, share-alike is contagious); the code is original work and carries CC BY-NC 4.0 — free for scholarship and teaching, commercial use by permission.
Can I self-host it? Yes — the architecture is local-first by contract (everything builds and runs on one machine; deployment is a copy step). The server reference (systemd/nginx/cron) is KOSHA_DEPLOYMENT.md.
Everything below is the engineering spine. All product and architecture decisions are locked — do not re-litigate; the documents are the contract.
kosha is glue over existing ecosystem assets, not a rebuild. Canonical inputs — consume, never regenerate:
| Concern | Owned by | What we consume |
|---|---|---|
| Lemma spine | union_headwords.tsv | 323,426 rows, slp1/iast/n_dicts/dicts/gender |
| Dict entry text | csl-orig / CDSL downloads | per-dict source, <L>/<k1>/<pc> keyed; csl-sqlite releases are the primary entry source, text parse is fallback |
| Existing dict API | csl-apidev (C-SALT Kosh API, api1/salt_*.php) + csl-websanlexicon (getword/servepdf/serveimg) |
endpoints + patterns; kosha builds on, not beside — Salt reuse is required (facade REST in D4) |
| Scan links | ls_resolver.py | Cologne scan URL resolution (port of csl-app ls_service.dart) |
| Form → lemma | RussianTranslation/glossary/ | DCS + vidyut.kosha fallback, 86.6 % token coverage — no live Sanskrit Heritage calls |
| Transliteration | sanskrit-util (Py + JS) | IAST ⇄ SLP1 ⇄ Devanagari — no transcoder #63 |
| Corpus evidence | corpus_lexicon.jsonl (RussianTranslation/src) |
1.09 M aligned Sa→Ru pairs (Phase 4) |
Per-dict <pc> ground truth (02-07-2026 audit — a naive comma-split parser
fails on 2 of 3): MW 720,1 = page,column, single volume · PWG 1-0001 =
volume-page, hyphen, 7 volumes · AP90 0001-a = page-column-letter.
- API tier: FastAPI + SQLite (single
kosha.db) on samskrtam.ru (/kosha/api/…). - Static tier: precomputed top-N lemma cache on this repo's GitHub Pages, cache-first JS with API fallback.
- A1–A4 contract (ARCHITECTURE.md):
raw-markup storage + ported transforms ·
dict.L.senseN@versionsense IDs · local-first, M.G. deploys · execution by a Sonnet/Opus session. - Builds never call live third-party services (RISKS R12); Heritage integration, when it comes, uses the UoHyd mirror (Inria host is bot-walled).
- Repo
vX.Y.Ztags anddata_versionare separate tracks; sense citations pin to the latter.
- No ✅ without an artifact. A phase or task is done when its exit check runs green, not when its code exists. Gates live in EVAL_PLAN.md (start at §0, the anti-gaming rules).
- Model provenance: every session records the model tier + exact version that did the work.
- Changelog discipline: SemVer, Keep-a-Changelog; every changelog edit
promotes
[Unreleased]to the next version + tag + GitHub release in the same pass. - Session state lives in .ai_state.md; open decisions (update cadence, latency SLO, etymology depth) get settled with real measurements at the end of Phase 1, recorded in KOSHA_DECISIONS_NEEDED.md (D5 settled the first three — see D5_MEASUREMENTS.md).
| In this repo | What it holds |
|---|---|
| ARCHITECTURE.md | A1–A4 contract, SQLite DDL, API v1, Salt max-reuse rules |
| PHASE1_PLAN.md | D1–D5 work order — the next thing to execute |
| IMPLEMENTATION_PLAN.md | P1–P7 full plan with exit checks |
| EVAL_PLAN.md | quality gates, anti-gaming rules, T-UC1…T-UC13 acceptance tests |
| USE_CASES.md | UC1–UC13 scenarios mapped to phases |
| RISKS.md | R1–R12 pre-mortem, citability commitments |
| RELATIONS.md | ecosystem diplomacy — Meyer, Cologne, Ambuda, C-SALT |
| POSITIONING.md · POSITIONING_SUMMARY.md | identity, audience, the UI decision |
| COMPARISON.md | 12-platform live survey (02-07-2026) with feature matrix |
| D5_MEASUREMENTS.md · KOSHA_DECISIONS_NEEDED.md | D5 measurement report + the SLO/cadence/cache-N decisions it settled |
| CHANGELOG.md | SemVer history |
| In SanskritLexicography | What it holds |
|---|---|
| KOSHA_FOLDER_SETUP.md | post-audit truth doc, meta-decisions M1–M4 |
| KOSHA_REFERENCE_ANALYSIS.md | feature-by-feature reference-site analysis |
| KOSHA_DEPLOYMENT.md | samskrtam.ru deployment reference (systemd/nginx/cron) |
The application deliberately lives outside SanskritLexicography (a data/research workspace whose charter forbids source code — meta-decision M2). This repo's GitHub Pages will host the static cache tier (M4). Treat the older planning docs' timeline/latency/cadence figures as historical; each carries a triage banner saying so — trust only the dated live survey.
Böhtlingk–Roth → Monier-Williams → Cologne digitization (1994–) → correction infrastructure (csl-orig) → API-ification (Kosh/C-SALT) → integration = this project. It writes no dictionary content and corrects nothing at source; it adds the join + access layer — the public face of a decade of private plumbing (union headwords, pwg_ru, frequency bands, the Zaliznyak index, the scan resolver).
Dr. Mārcis Gasūns