kotoshu-server 0.1.0 → 0.1.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 80541f5f06f9c4f0c41cbdecfbf195fec21d630f9e89711c061f4a3796753cfc
-  data.tar.gz: dae5536d6035f67546136f12ffe1fc52d602bdffc8dd5eeba120fec2313c4cd7
+  metadata.gz: 6fa24971e66e791bd277c8fe878890693c6a624a3de1c488dbb7f5b38f629f0a
+  data.tar.gz: eabc7a8dd8124cf20a1c05b32257e9fdfdcdebeb72ac6aceae249c2aba692976
 SHA512:
-  metadata.gz: b3bcab28c17474fbc07c721bb7686ae6d7505580a70e03042438aef75d032b846efc9576520351f1967a4e5fc04d8e9b21590a4961b7f9bbc420bb3ff10505b6
-  data.tar.gz: 8cf8d948c812075b068479e25737a0d43999831f2dea87a018d6d62fe8c1049d96bb27076f88173ed91d4bd4ae1939669d768402ee9e8ce8e5c215c14014d909
+  metadata.gz: fac8b861344a723119bf84e7c078c772d491d974cb37a9338b2100acb0e28cf9e17d83e06b4f18aad41ecba9d2ec78bc6249575830b485f03f6f1200ab58daa0
+  data.tar.gz: bd7c11018743035167858f90f8a7b70aa29efdfb6313496b3f6d06a1e297931985e6259625d3b43b34a5236d03179f58851a33f388f9f5a6e2cbcb0701ab745b

data/LICENSE

@@ -0,0 +1,26 @@
+BSD 2-Clause License
+
+Copyright (c) 2025-2026, Kotoshu contributors
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+

data/README.md

@@ -0,0 +1,141 @@
+# kotoshu-server
+
+Self-hostable HTTP API wrapping the [Kotoshu](https://github.com/kotoshu/kotoshu) spell checker.
+
+## Status
+
+MVP. Six endpoints over JSON. Rack/Sinatra + Puma. Pre-warms
+languages on boot. Designed as the deployment surface for non-Ruby
+SDKs (`kotoshu-python`, `kotoshu-js`, `kotoshu-go`).
+
+See `TODO.impl/64-http-api-and-sdks.md` for the full plan.
+
+## Install
+
+**Run from source.** The published `kotoshu-server 0.1.0` gem is
+empty (a gemspec file-list bug, fixed on main). Use source until the
+owner republishes 0.1.1:
+
+```bash
+cd kotoshu-server && bundle install && bundle exec exe/kotoshu-server
+```
+
+## Run
+
+```bash
+# Default: localhost:9292, English pre-warmed
+kotoshu-server
+
+# Custom port + multi-language pre-warm
+KOTOSHU_SERVER_PORT=8080 \
+KOTOSHU_SERVER_LANGUAGES="en de fr" \
+  kotoshu-server
+```
+
+## Semantic models (opt-in)
+
+By default the server is dictionary-only — exactly the 0.1.0
+behavior. Semantic reranking is a boot-time opt-in; the server never
+downloads a model implicitly:
+
+```bash
+KOTOSHU_SERVER_LANGUAGES="en de" \
+KOTOSHU_SERVER_MODEL_LANGS="en" \
+  kotoshu-server
+```
+
+- `KOTOSHU_SERVER_MODEL_LANGS` — languages the pre-warm thread sets
+  up with spelling + semantic model (space separated). Unset means no
+  models, no downloads, no behavior change.
+- `KOTOSHU_SERVER_MODEL_TIER` — model tier to set up and resolve:
+  `fluency` (default), `full`, or `mini`.
+
+**Requires kotoshu >= 0.7.0.** Model tiers, the resource registry,
+and the confidence cascade ship in the 0.7.0 cut. The gemspec keeps
+its `kotoshu ~> 0.6` constraint (dependency floors are the owner's
+decision), so the server checks the *installed* gem instead: setting
+`KOTOSHU_SERVER_MODEL_LANGS` with an older kotoshu fails fast at
+boot with a clear error, and explicit `"model": true` requests
+return 503.
+
+Once a language has a model set up, `POST /v1/check` accepts an
+optional `"model": true|false` flag (default: whether the language
+has a model set up server-side). With it on, each error's
+suggestions are reranked by the semantic analyzer; the gem's
+confidence cascade (`KOTOSHU_SEMANTIC_CASCADE_THRESHOLD`) decides per
+word whether the ONNX rerank actually runs. `GET /v1/languages`
+reports `"model": {"en": true}` per cached language.
+
+Memory and latency: budget roughly 15 MB resident per language at
+the default fluency tier (the full tier is far larger), plus
+one-time model load on the first model-enabled request. Listing the
+language in `KOTOSHU_SERVER_MODEL_LANGS` warms it at boot instead.
+
+## Endpoints
+
+| Method | Path | Body | Returns |
+|---|---|---|---|
+| `GET` | `/` | — | service metadata |
+| `GET` | `/v1/health` | — | `{ status, ready, timestamp }` |
+| `GET` | `/v1/version` | — | `{ server, kotoshu, ruby }` |
+| `GET` | `/v1/languages` | — | `{ cached, supported, model }` |
+| `POST` | `/v1/check` | `{ text, language?, format?, model? }` | `{ file, word_count, errors: [...] }` |
+| `POST` | `/v1/suggest` | `{ word, language?, max? }` | `{ word, suggestions: [...] }` |
+| `POST` | `/v1/detect` | `{ text }` | `{ language, confidence }` |
+
+Each `error` and `suggestion` mirrors the lutaml-model serialization
+(`word`, `distance`, `confidence`, `source`).
+
+## curl examples
+
+```bash
+# Check a document
+curl -X POST http://localhost:9292/v1/check \
+  -H "Content-Type: application/json" \
+  -d '{"text":"helo wrold","language":"en"}'
+
+# Suggestions for one word
+curl -X POST http://localhost:9292/v1/suggest \
+  -H "Content-Type: application/json" \
+  -d '{"word":"helo","language":"en","max":3}'
+
+# Detect language
+curl -X POST http://localhost:9292/v1/detect \
+  -H "Content-Type: application/json" \
+  -d '{"text":"bonjour le monde"}'
+```
+
+## Docker
+
+```bash
+docker build -t kotoshu-server .
+
+docker run --rm -p 9292:9292 \
+  -e KOTOSHU_SERVER_LANGUAGES="en de" \
+  kotoshu-server
+```
+
+Healthcheck probes `/v1/health` every 30s.
+
+## Configuration
+
+| Env var | Default | Purpose |
+|---|---|---|
+| `KOTOSHU_SERVER_PORT` | `9292` | Listen port |
+| `KOTOSHU_SERVER_BIND` | `0.0.0.0` | Bind address |
+| `KOTOSHU_SERVER_LANGUAGES` | `en` | Languages to pre-warm on boot |
+| `KOTOSHU_SERVER_MODEL_LANGS` | unset | Languages to set up with a semantic model (requires kotoshu >= 0.7) |
+| `KOTOSHU_SERVER_MODEL_TIER` | `fluency` | Model tier for `KOTOSHU_SERVER_MODEL_LANGS` |
+| `KOTOSHU_SERVER_LAZY` | `0` | Skip pre-warm; load on first request |
+| `KOTOSHU_SERVER_DEFAULT_LANG` | `en` | When client omits `language` |
+| `KOTOSHU_SERVER_LOG_LEVEL` | `info` | `debug`/`info`/`warn`/`error` |
+| `KOTOSHU_OFFLINE` | `1` (in Docker) | Never trigger downloads |
+
+## OpenAPI
+
+The full OpenAPI 3.1 spec is at [`openapi.yaml`](./openapi.yaml). Use
+`openapi-generator` to spin up SDKs in Rust / .NET / Java.
+
+## License
+
+BSD-2-Clause, same as Kotoshu.

data/exe/kotoshu-server

@@ -0,0 +1,7 @@
+#!/usr/bin/env ruby
+# frozen_string_literal: true
+
+$LOAD_PATH.unshift(File.expand_path("../lib", __dir__))
+require "kotoshu/server"
+
+Kotoshu::Server.run!

data/lib/kotoshu/server/app.rb

@@ -0,0 +1,510 @@
+# frozen_string_literal: true
+
+require "json"
+require "logger"
+require "sinatra/base"
+require "kotoshu"
+
+module Kotoshu
+  module Server
+    # Raised at boot when KOTOSHU_SERVER_MODEL_LANGS / _TIER are set
+    # but cannot be honored (kotoshu gem too old, unknown tier).
+    class ModelConfigError < StandardError; end
+
+    class App < Sinatra::Base
+      VERSION = "0.1.0".freeze
+
+      # Semantic models (tiers, registry, confidence cascade) ship in
+      # kotoshu 0.7.0. The gemspec keeps its `kotoshu ~> 0.6`
+      # constraint (dependency floors are the owner's call), so the
+      # server validates the *installed* gem at boot instead.
+      MODEL_MIN_KOTOSHU = Gem::Version.new("0.7.0")
+
+      # Languages to set up at boot, from KOTOSHU_SERVER_LANGUAGES
+      # (space separated). Single source of truth for the env var: the
+      # boot pre-warm and /v1/health both read it here.
+      #
+      # @return [Array<String>] Language codes to pre-warm
+      def self.configured_languages
+        ENV.fetch("KOTOSHU_SERVER_LANGUAGES", "en").split
+      end
+
+      # Languages to set up with spelling + semantic model at boot,
+      # from KOTOSHU_SERVER_MODEL_LANGS (space separated, e.g. "en de").
+      # Empty when unset — boot-time opt-in only, never an implicit
+      # download (the gem's two-stage promise).
+      #
+      # @return [Array<String>] Language codes to set up with a model
+      def self.model_languages
+        ENV.fetch("KOTOSHU_SERVER_MODEL_LANGS", "").split
+      end
+
+      # Model tier used for setup and resolution, from
+      # KOTOSHU_SERVER_MODEL_TIER. Defaults to "fluency", the
+      # ecosystem default (owner decision 2026-09-04).
+      #
+      # @return [String] "full", "fluency", or "mini"
+      def self.model_tier
+        ENV.fetch("KOTOSHU_SERVER_MODEL_TIER", "fluency")
+      end
+
+      # Whether the installed kotoshu gem carries the semantic-model
+      # surface the server wires against (0.7.0+).
+      #
+      # @return [Boolean]
+      def self.semantic_models_supported?
+        Gem::Version.new(Kotoshu::VERSION) >= MODEL_MIN_KOTOSHU
+      end
+
+      # Whether a semantic model is actually set up server-side for a
+      # language (cache-only lookup; false on gems without model
+      # support, so the /v1/check default flag never turns models on
+      # underneath an old install).
+      #
+      # @param language [String] Language code
+      # @return [Boolean]
+      def self.model_available?(language)
+        return false unless semantic_models_supported?
+
+        Kotoshu.setup?(language.to_sym, :model)
+      rescue StandardError
+        false
+      end
+
+      # Validate the semantic-model environment before anything is
+      # set up. Raises {ModelConfigError} with an actionable message
+      # when KOTOSHU_SERVER_MODEL_LANGS is set but the installed
+      # kotoshu gem predates model support, or the configured tier is
+      # unknown. A no-op when the vars are unset — boot behavior is
+      # then exactly today's.
+      #
+      # @raise [ModelConfigError] on an unsatisfiable configuration
+      # @return [void]
+      def self.validate_model_config!
+        return if model_languages.empty?
+
+        unless semantic_models_supported?
+          raise ModelConfigError,
+                "KOTOSHU_SERVER_MODEL_LANGS requires kotoshu >= 0.7.0 " \
+                "(installed: #{Kotoshu::VERSION}). Semantic models ship in " \
+                "kotoshu 0.7.0; unset KOTOSHU_SERVER_MODEL_LANGS or upgrade " \
+                "the kotoshu gem."
+        end
+
+        begin
+          Kotoshu::Cache::ModelCache.normalize_tier(model_tier)
+        rescue ArgumentError => e
+          raise ModelConfigError, "KOTOSHU_SERVER_MODEL_TIER: #{e.message}"
+        end
+      end
+
+      # Synchronously set up the given languages (downloads on a cold
+      # or expired cache). Runs on the pre-warm thread; also usable
+      # directly by embedders that want a blocking warm-up.
+      #
+      # Languages listed in KOTOSHU_SERVER_MODEL_LANGS are set up with
+      # spelling + model (at KOTOSHU_SERVER_MODEL_TIER) in one setup
+      # call; the rest get spelling only, as before. Both lists are
+      # unioned, and setup is idempotent, so a language in both is
+      # set up once with the model.
+      #
+      # @param languages [Array<String>] Language codes
+      # @return [void]
+      def self.prewarm!(languages)
+        logger = Logger.new($stderr)
+        model_langs = model_languages
+        tier = model_tier
+        (languages | model_langs).each do |lang|
+          with_model = model_langs.include?(lang)
+          logger.info("pre-warming #{lang}#{with_model ? " (spelling + model, #{tier} tier)" : ""}")
+          begin
+            if with_model
+              Kotoshu.setup(lang.to_sym, want: %i[spelling model], tier: tier)
+            else
+              Kotoshu.setup(lang.to_sym)
+            end
+            logger.info("pre-warm #{lang} complete")
+          rescue StandardError => e
+            logger.warn("pre-warm #{lang} failed: #{e.message}")
+          end
+        end
+      end
+
+      # Start pre-warming in a detached background thread and return
+      # immediately. The server must bind and serve within seconds of
+      # boot, while resource setup can block on the network
+      # (download retries, and DNS resolution that no Net::HTTP
+      # timeout bounds) — so setup never runs on the boot path.
+      # Progress and completion are logged; /v1/health reports
+      # per-language readiness while it runs.
+      #
+      # @param languages [Array<String>] Language codes
+      # @return [Thread] the detached pre-warm thread
+      def self.prewarm_async!(languages)
+        validate_model_config! # fail fast, on the caller's thread
+        Thread.new do
+          Thread.current.name = "kotoshu-server-prewarm"
+          prewarm!(languages)
+        end
+      end
+
+      # ---- Semantic analyzers (memoized per language + model file) ----
+
+      @semantic_analyzers = {}
+      @semantic_analyzers_mutex = Mutex.new
+
+      # A memoized {Kotoshu::Analyzers::SemanticAnalyzer} for the model
+      # file the ResourceManager resolved for `language`. Loading a
+      # model (vocab + ONNX session) is expensive; one analyzer per
+      # language is kept for the process lifetime. A changed model
+      # file (different path) loads a fresh analyzer.
+      #
+      # @param language [String] Language code
+      # @param model_info [Hash] bundle.model from ResourceManager
+      #   (needs :model_path)
+      # @return [Kotoshu::Analyzers::SemanticAnalyzer]
+      def self.semantic_analyzer_for(language, model_info)
+        path = model_info[:model_path]
+        @semantic_analyzers_mutex.synchronize do
+          @semantic_analyzers[[language.to_s, path]] ||= begin
+            model = Kotoshu::Models::OnnxModel.from_file(path, language_code: language.to_s)
+            Kotoshu::Analyzers::SemanticAnalyzer.new(model)
+          end
+        end
+      end
+
+      configure do
+        set :logging, false
+        set :show_exceptions, false
+
+        app_logger = Logger.new($stderr)
+        app_logger.level = ENV.fetch("KOTOSHU_SERVER_LOG_LEVEL", Logger::INFO)
+        set :app_logger, app_logger
+      end
+
+      # ---- Endpoints ----
+
+      get "/" do
+        content_type :json
+        {
+          name: "kotoshu-server",
+          version: VERSION,
+          kotoshu_version: Kotoshu::VERSION,
+          docs: "/v1/health, /v1/version, /v1/languages, /v1/check, /v1/suggest, /v1/detect"
+        }.to_json
+      end
+
+      get "/v1/health" do
+        content_type :json
+        ready = self.class.configured_languages.map { |l| [l, Kotoshu.setup?(l.to_sym, :spelling)] }.to_h
+        { status: "ok", ready: ready, timestamp: Time.now.utc.iso8601 }.to_json
+      end
+
+      get "/v1/version" do
+        content_type :json
+        { server: VERSION, kotoshu: Kotoshu::VERSION, ruby: RUBY_VERSION }.to_json
+      end
+
+      get "/v1/languages" do
+        content_type :json
+        cached = Kotoshu.languages_setup
+        model = cached.to_h { |lang| [lang, self.class.model_available?(lang)] }
+        { cached: cached, supported: cached, model: model }.to_json
+      end
+
+      post "/v1/check" do
+        body = parse_json_body(request.body.read)
+        text = body["text"]
+        language = body["language"] || ENV.fetch("KOTOSHU_SERVER_DEFAULT_LANG", "en")
+        format_hint = body["format"] || "full"
+        want_model = model_request_flag(body["model"], language)
+
+        halt_with_error(400, "missing 'text'") unless text.is_a?(String)
+
+        want = want_model ? %i[spelling model] : %i[spelling]
+        result = with_resource(language, want: want) do |checker, bundle|
+          if want_model && bundle.model.nil?
+            # The gem resolves a nil model for languages that cannot
+            # have one; an explicit request must not silently degrade.
+            halt 422, { "Content-Type" => "application/json" },
+                 {
+                   error: "resource_not_setup",
+                   message: "no semantic model set up for language '#{language}'",
+                   hint: "list it in KOTOSHU_SERVER_MODEL_LANGS and restart"
+                 }.to_json
+          end
+          check_result = checker.check(text)
+          want_model ? rerank_semantic(language, check_result, text, bundle.model) : check_result
+        end
+
+        content_type :json
+        case format_hint
+        when "errors" then serialize_errors(result).to_json
+        else serialize_full(result).to_json
+        end
+      end
+
+      post "/v1/suggest" do
+        body = parse_json_body(request.body.read)
+        word = body["word"]
+        language = body["language"] || ENV.fetch("KOTOSHU_SERVER_DEFAULT_LANG", "en")
+        max = body["max"]&.to_i
+
+        halt_with_error(400, "missing 'word'") unless word.is_a?(String)
+
+        result = with_resource(language) do |checker|
+          checker.suggest(word, max_suggestions: max)
+        end
+
+        content_type :json
+        { word: word, suggestions: serialize_suggestions(result) }.to_json
+      end
+
+      post "/v1/detect" do
+        body = parse_json_body(request.body.read)
+        text = body["text"]
+        halt_with_error(400, "missing 'text'") unless text.is_a?(String)
+
+        lang, confidence = Kotoshu.detect_language_with_confidence(text)
+        content_type :json
+        { language: lang, confidence: confidence }.to_json
+      end
+
+      # ---- Error handling ----
+
+      error Kotoshu::Server::ModelConfigError do
+        status 500
+        content_type :json
+        { error: "model_config", message: env["sinatra.error"].message }.to_json
+      end
+
+      error Kotoshu::Models::OnnxModel::OnnxUnavailable do
+        status 503
+        content_type :json
+        { error: "onnx_unavailable", message: env["sinatra.error"].message,
+          hint: "install the onnxruntime gem, or stop setting model langs" }.to_json
+      end
+
+      error Kotoshu::ResourceNotSetupError do
+        e = env["sinatra.error"]
+        status 422
+        content_type :json
+        { error: "resource_not_setup", message: e.message,
+          hint: "POST /v1/admin/setup with {language: \"...\"} (admin only)" }.to_json
+      end
+
+      error JSON::ParserError do
+        status 400
+        content_type :json
+        { error: "invalid_json", message: env["sinatra.error"].message }.to_json
+      end
+
+      error StandardError do
+        e = env["sinatra.error"]
+        settings.app_logger&.error("unhandled: #{e.class}: #{e.message}")
+        status 500
+        content_type :json
+        { error: "internal", message: e.message }.to_json
+      end
+
+      # ---- Helpers ----
+
+      helpers do
+        def parse_json_body(body)
+          return {} if body.nil? || body.empty?
+
+          JSON.parse(body)
+        end
+
+        def halt_with_error(code, message)
+          halt code, { "Content-Type" => "application/json" },
+               { error: "invalid_request", message: message }.to_json
+        end
+
+        def with_resource(language, want: %i[spelling])
+          Kotoshu.reset_spellchecker if Kotoshu.instance_variable_get(:@spellcheckers).nil?
+          bundle = Kotoshu::ResourceManager.resolve(language: language, want: want)
+          checker = Kotoshu::Spellchecker.new(resource_bundle: bundle)
+          yield checker, bundle
+        end
+
+        # Resolve the effective "model" flag for /v1/check: an explicit
+        # request value wins; otherwise the default is whether the
+        # language has a model set up server-side (cache-only). An
+        # explicit true on a kotoshu install without model support is
+        # a 503 naming the requirement, not a silent dictionary-only
+        # answer.
+        #
+        # @param value [Boolean, nil] the request's "model" field
+        # @param language [String] resolved request language
+        # @return [Boolean]
+        def model_request_flag(value, language)
+          case value
+          when nil then self.class.model_available?(language)
+          when true
+            unless self.class.semantic_models_supported?
+              halt 503, { "Content-Type" => "application/json" },
+                   {
+                     error: "model_unsupported",
+                     message: "semantic models require kotoshu >= 0.7.0 " \
+                              "(installed: #{Kotoshu::VERSION})"
+                   }.to_json
+            end
+            true
+          when false then false
+          else halt_with_error(400, "'model' must be true or false")
+          end
+        end
+
+        # Rerank a traditional check result with the gem's semantic
+        # analyzer (dictionary verdict first, neural rerank for
+        # uncertain candidates — the gem's confidence cascade decides
+        # per word whether the ONNX rerank runs at all). Semantic
+        # candidates lead the merged suggestion list; traditional
+        # candidates follow, deduplicated by word. Any per-word
+        # analyzer failure keeps the traditional suggestions for that
+        # word — one bad word never fails the request.
+        #
+        # @param language [String] Language code
+        # @param result [Kotoshu::Models::Result::DocumentResult]
+        # @param text [String] the checked text (context source)
+        # @param model_info [Hash, nil] bundle.model from the resolve
+        # @return [Kotoshu::Models::Result::DocumentResult]
+        def rerank_semantic(language, result, text, model_info)
+          return result if model_info.nil? || model_info[:model_path].nil?
+          errors = Array(result.errors)
+          return result if errors.empty?
+
+          analyzer = self.class.semantic_analyzer_for(language, model_info)
+          cascade = Kotoshu::Suggestions::SemanticCascade.from_configuration(Kotoshu.configuration)
+
+          reranked = errors.map { |error| rerank_error(analyzer, cascade, text, error) }
+
+          Kotoshu::Models::Result::DocumentResult.new(
+            file: result.file,
+            errors: reranked,
+            suppressed_errors: result.suppressed_errors,
+            word_count: result.word_count
+          )
+        end
+
+        # Rerank one error's suggestions, or return the error unchanged
+        # when the cascade skips it or the analyzer has nothing better
+        # (including an analyzer failure on this word — traditional
+        # suggestions survive).
+        #
+        # @param analyzer [Kotoshu::Analyzers::SemanticAnalyzer]
+        # @param cascade [Kotoshu::Suggestions::SemanticCascade]
+        # @param text [String] the checked text
+        # @param error [Kotoshu::Models::Result::WordResult]
+        # @return [Kotoshu::Models::Result::WordResult]
+        def rerank_error(analyzer, cascade, text, error)
+          traditional = error.suggestions.to_a
+          return error if traditional.empty? || cascade.skip?(traditional)
+
+          semantic = begin
+            analyzer.suggest_corrections(error.word, context: semantic_context(text, error))
+          rescue StandardError
+            []
+          end
+          return error if semantic.empty?
+
+          Kotoshu::Models::Result::WordResult.new(
+            word: error.word,
+            correct: false,
+            position: error.position,
+            suggestions: merge_suggestions(semantic, traditional),
+            suppressed: error.suppressed,
+            suppressed_by: error.suppressed_by
+          )
+        end
+
+        # Merge analyzer suggestions (semantic, context-ranked) ahead
+        # of the traditional ones, deduplicated by word, mapped to the
+        # wire Suggestion shape /v1/check already serializes.
+        #
+        # @param semantic [Array<Kotoshu::Models::Suggestion>]
+        # @param traditional [Array<Kotoshu::Suggestions::Suggestion>]
+        # @return [Array<Kotoshu::Suggestions::Suggestion>]
+        def merge_suggestions(semantic, traditional)
+          merged = []
+          seen = {}
+          semantic.each do |s|
+            next if seen[s.word]
+
+            seen[s.word] = true
+            merged << Kotoshu::Suggestions::Suggestion.new(
+              word: s.word,
+              distance: s.metadata[:distance].to_i,
+              confidence: s.confidence,
+              source: :semantic
+            )
+          end
+          traditional.each do |s|
+            next if seen[s.word]
+
+            seen[s.word] = true
+            merged << s
+          end
+          merged
+        end
+
+        # The text window around an error, in the shape the analyzer's
+        # own context ranking expects ({Models::Context} before /
+        # current / after slices; nil when the error has no position).
+        #
+        # @param text [String] the checked text
+        # @param error [Kotoshu::Models::Result::WordResult]
+        # @return [Kotoshu::Models::Context, nil]
+        def semantic_context(text, error)
+          pos = error.position
+          return nil unless pos.is_a?(Integer)
+
+          window = 32
+          word_end = pos + error.word.length
+          Kotoshu::Models::Context.new(
+            before: text[[pos - window, 0].max...pos] || "",
+            current: text[pos...word_end] || "",
+            after: text[word_end...(word_end + window)] || "",
+            location: nil
+          )
+        end
+
+        def serialize_full(result)
+          {
+            file: result.file,
+            word_count: result.word_count,
+            errors: serialize_errors(result)
+          }
+        end
+
+        def serialize_errors(result)
+          (result.errors || []).map do |err|
+            {
+              word: err.word,
+              position: err.position,
+              suggestions: err.suggestions.map { |s| serialize_suggestion(s) }
+            }
+          end
+        end
+
+        def serialize_suggestions(set)
+          return [] unless set.respond_to?(:suggestions)
+
+          set.suggestions.map { |s| serialize_suggestion(s) }
+        end
+
+        def serialize_suggestion(s)
+          {
+            word: s.word,
+            distance: s.distance,
+            confidence: s.confidence,
+            source: s.source
+          }
+        end
+      end
+    end
+  end
+end
+
+require "time"

data/lib/kotoshu/server/version.rb

@@ -0,0 +1,7 @@
+# frozen_string_literal: true
+
+module Kotoshu
+  module Server
+    VERSION = "0.1.1"
+  end
+end

data/lib/kotoshu/server.rb

@@ -0,0 +1,27 @@
+# frozen_string_literal: true
+
+require_relative "server/version"
+require_relative "server/app"
+
+module Kotoshu
+  module Server
+    # Boot the server.
+    #
+    # Pre-warm starts in a detached thread and never blocks: Puma binds
+    # and serves within seconds regardless of how long resource setup
+    # (cold-cache downloads) takes. KOTOSHU_SERVER_LAZY=1 skips the
+    # pre-warm entirely; /v1/health reports readiness while it runs.
+    #
+    # Semantic-model env vars (KOTOSHU_SERVER_MODEL_LANGS / _TIER) are
+    # validated before the bind, so an unsatisfiable model config
+    # fails fast with a clear error instead of surprising per-request
+    # failures. Unset, boot is exactly as before.
+    def self.run!(port: ENV.fetch("KOTOSHU_SERVER_PORT", 9292), bind: ENV.fetch("KOTOSHU_SERVER_BIND", "0.0.0.0"))
+      App.validate_model_config!
+      unless ENV.fetch("KOTOSHU_SERVER_LAZY", "0") == "1"
+        App.prewarm_async!(App.configured_languages)
+      end
+      App.run!(port: port.to_i, bind: bind)
+    end
+  end
+end

data/openapi.yaml

@@ -0,0 +1,244 @@
+openapi: 3.1.0
+info:
+  title: Kotoshu HTTP API
+  version: 0.1.0
+  description: |
+    Self-hostable HTTP API wrapping the Kotoshu spell checker.
+    Used by the Python, JavaScript, and Go SDKs.
+  license:
+    name: BSD-2-Clause
+    url: https://github.com/kotoshu/kotoshu/blob/main/LICENSE
+servers:
+  - url: http://localhost:9292
+    description: Local server
+paths:
+  /:
+    get:
+      summary: Service metadata
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Metadata"
+  /v1/health:
+    get:
+      summary: Liveness probe + cache state
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Health"
+  /v1/version:
+    get:
+      summary: Server + kotoshu versions
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Version"
+  /v1/languages:
+    get:
+      summary: Languages with cached dictionaries and model availability
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Languages"
+  /v1/check:
+    post:
+      summary: Check text for spelling errors
+      description: |
+        Optional `"model": true` reranks each error's suggestions with
+        the semantic model set up server-side for the language
+        (requires KOTOSHU_SERVER_MODEL_LANGS at boot, kotoshu >= 0.7).
+        The default is whether a model is set up for the language.
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/CheckRequest"
+      responses:
+        "200":
+          description: Result
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/DocumentResult"
+        "400":
+          $ref: "#/components/responses/BadRequest"
+        "422":
+          $ref: "#/components/responses/ResourceNotSetup"
+        "503":
+          $ref: "#/components/responses/ModelUnsupported"
+  /v1/suggest:
+    post:
+      summary: Get suggestions for a single word
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/SuggestRequest"
+      responses:
+        "200":
+          description: Suggestions
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Suggestions"
+        "400":
+          $ref: "#/components/responses/BadRequest"
+  /v1/detect:
+    post:
+      summary: Detect the language of text
+      requestBody:
+        required: true
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/DetectRequest"
+      responses:
+        "200":
+          description: Detected language
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Detection"
+        "400":
+          $ref: "#/components/responses/BadRequest"
+components:
+  responses:
+    BadRequest:
+      description: Malformed request
+      content:
+        application/json:
+          schema:
+            $ref: "#/components/schemas/Error"
+    ResourceNotSetup:
+      description: Language not set up on the server
+      content:
+        application/json:
+          schema:
+            $ref: "#/components/schemas/Error"
+    ModelUnsupported:
+      description: Model requested but the installed kotoshu gem is older than 0.7.0
+      content:
+        application/json:
+          schema:
+            $ref: "#/components/schemas/Error"
+  schemas:
+    Metadata:
+      type: object
+      required: [name, version, kotoshu_version]
+      properties:
+        name: { type: string, example: kotoshu-server }
+        version: { type: string, example: 0.1.0 }
+        kotoshu_version: { type: string, example: 0.6.0 }
+        docs: { type: string }
+    Health:
+      type: object
+      required: [status, ready]
+      properties:
+        status: { type: string, example: ok }
+        ready:
+          type: object
+          additionalProperties: { type: boolean }
+          example: { en: true }
+        timestamp: { type: string, format: date-time }
+    Version:
+      type: object
+      required: [server, kotoshu]
+      properties:
+        server: { type: string }
+        kotoshu: { type: string }
+        ruby: { type: string }
+    Languages:
+      type: object
+      required: [cached, model]
+      properties:
+        cached:
+          type: array
+          items: { type: string }
+        supported:
+          type: array
+          items: { type: string }
+        model:
+          type: object
+          description: Whether a semantic model is set up per cached language
+          additionalProperties: { type: boolean }
+          example: { en: true, de: false }
+    CheckRequest:
+      type: object
+      required: [text]
+      properties:
+        text: { type: string }
+        language: { type: string, default: en }
+        format: { type: string, enum: [full, errors], default: full }
+        model:
+          type: boolean
+          description: |
+            Rerank suggestions with the language's semantic model.
+            Default is whether a model is set up server-side. true on a
+            language without a cached model returns 422.
+    SuggestRequest:
+      type: object
+      required: [word]
+      properties:
+        word: { type: string }
+        language: { type: string, default: en }
+        max: { type: integer, minimum: 1, maximum: 50 }
+    DetectRequest:
+      type: object
+      required: [text]
+      properties:
+        text: { type: string }
+    DocumentResult:
+      type: object
+      properties:
+        file: { type: string, nullable: true }
+        word_count: { type: integer }
+        errors:
+          type: array
+          items: { $ref: "#/components/schemas/WordError" }
+    WordError:
+      type: object
+      properties:
+        word: { type: string }
+        position: { type: integer, nullable: true }
+        suggestions:
+          type: array
+          items: { $ref: "#/components/schemas/Suggestion" }
+    Suggestions:
+      type: object
+      properties:
+        word: { type: string }
+        suggestions:
+          type: array
+          items: { $ref: "#/components/schemas/Suggestion" }
+    Suggestion:
+      type: object
+      properties:
+        word: { type: string }
+        distance: { type: integer }
+        confidence: { type: number, format: float }
+        source: { type: string }
+    Detection:
+      type: object
+      properties:
+        language: { type: string, nullable: true }
+        confidence: { type: number, format: float }
+    Error:
+      type: object
+      properties:
+        error: { type: string }
+        message: { type: string }
+        hint: { type: string }

metadata

@@ -1,13 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: kotoshu-server
 version: !ruby/object:Gem::Version
-  version: 0.1.0
+  version: 0.1.1
 platform: ruby
 authors:
 - Ribose Inc.
+autorequire:
 bindir: exe
 cert_chain: []
-date: 1980-01-02 00:00:00.000000000 Z
+date: 2026-09-06 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: kotoshu
@@ -69,10 +70,18 @@ description: Self-hostable HTTP API that exposes Kotoshu's check / suggest / det
   over JSON. Designed as the deployment surface for non-Ruby SDKs (Python, JS, Go).
 email:
 - open.source@ribose.com
-executables: []
+executables:
+- kotoshu-server
 extensions: []
 extra_rdoc_files: []
-files: []
+files:
+- LICENSE
+- README.md
+- exe/kotoshu-server
+- lib/kotoshu/server.rb
+- lib/kotoshu/server/app.rb
+- lib/kotoshu/server/version.rb
+- openapi.yaml
 homepage: https://github.com/kotoshu/kotoshu-server
 licenses:
 - BSD-2-Clause
@@ -80,6 +89,7 @@ metadata:
   homepage_uri: https://github.com/kotoshu/kotoshu-server
   source_code_uri: https://github.com/kotoshu/kotoshu-server/tree/main
   rubygems_mfa_required: 'true'
+post_install_message:
 rdoc_options: []
 require_paths:
 - lib
@@ -94,7 +104,8 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 4.0.16
+rubygems_version: 3.5.22
+signing_key:
 specification_version: 4
 summary: HTTP API server wrapping the Kotoshu spell checker
 test_files: []