@atlaskit/editor-plugin-autocomplete 6.0.0 → 8.0.0

package/dist/esm/pm-plugins/text-predictor.js

@@ -1,5 +1,5 @@
-import _asyncToGenerator from "@babel/runtime/helpers/asyncToGenerator";
 import _typeof from "@babel/runtime/helpers/typeof";
+import _asyncToGenerator from "@babel/runtime/helpers/asyncToGenerator";
 import _slicedToArray from "@babel/runtime/helpers/slicedToArray";
 import _createClass from "@babel/runtime/helpers/createClass";
 import _classCallCheck from "@babel/runtime/helpers/classCallCheck";
@@ -29,9 +29,10 @@ function _arrayLikeToArray(r, a) { (null == a || a > r.length) && (a = r.length)
 
 import { EXPERIENCE_NAME, failExp, startExp, succeedExp } from '../analytics/ufo';
 
-// The vocabulary, L3 and word-index JSON payloads are dynamically imported in
-// loadDefaultVocabulary / loadVectorsAsync so their (large) contents stay out of
-// the editor's main chunk and only load when autocomplete is initialised.
+// The vocabulary, L3 and word-index payloads are fetched from the CDN in
+// loadDefaultVocabulary / loadVectorsAsync so their (large) contents never enter
+// the editor's bundle and only load when autocomplete is initialised.
+import { ARTIFACT_NAME, fetchArtifactJson, getArtifactUrl } from './artifacts-manifest';
 import { isAutocompleteDebugEnabled } from './debug-mode';
 import { rankCandidates, STAGE1_WEIGHT, STAGE2_WEIGHT, MIN_STAGE1_SCORE } from './scoring-pipeline';
 import { getStoredContextVector, getStoredLmLogits } from './slow-lane-client';
@@ -685,56 +686,10 @@ export var predict = function predict(textBefore) {
 
 // ─── Data Loading ────────────────────────────────────────────────────────────
 
-/**
- * Unwrap a dynamically imported JSON module to its parsed value, handling both
- * interop modes AFM's bundler chain emits: a `.default`-wrapped namespace
- * (classic webpack) and a named-exports namespace (webpack 5 / atlaspack JSON
- * modules, where `default` can be a misleading scalar). Named exports are
- * preferred when present. The caller declares the JSON `shape` because a dense
- * array and a sparse numeric-keyed object are emitted identically as named
- * exports. Kept in lock-step with the matching helper in local-slow-lane-client.ts.
- */
-function unwrapJsonModule(mod, shape) {
-  if (mod == null || _typeof(mod) !== 'object') {
-    return null;
-  }
-  var namespace = mod;
-  var ownKeys = Object.keys(namespace).filter(function (k) {
-    return k !== 'default' && k !== '__esModule';
-  });
-  if (ownKeys.length > 0) {
-    if (shape === 'array') {
-      var len = ownKeys.length;
-      var arr = new Array(len);
-      for (var i = 0; i < len; i++) {
-        arr[i] = namespace[String(i)];
-      }
-      return arr;
-    }
-    var obj = {};
-    var _iterator1 = _createForOfIteratorHelper(ownKeys),
-      _step1;
-    try {
-      for (_iterator1.s(); !(_step1 = _iterator1.n()).done;) {
-        var k = _step1.value;
-        obj[k] = namespace[k];
-      }
-    } catch (err) {
-      _iterator1.e(err);
-    } finally {
-      _iterator1.f();
-    }
-    return obj;
-  }
-  if ('default' in namespace && namespace.default != null) {
-    return namespace.default;
-  }
-  return null;
-}
 export var loadVectorsAsync = /*#__PURE__*/function () {
   var _ref6 = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee(options) {
     var _options$isLocalLLM;
-    var isLocalLLM, surface, url, _wordIndexOuter$index, res, buffer, float32, wordIndexModule, wordIndexOuter, wordIndex, nWords, dim, _t, _t2;
+    var isLocalLLM, surface, url, res, buffer, float32, wordIndexPayload, wrappedIndex, wordIndex, nWords, dim, _t, _t2;
     return _regeneratorRuntime.wrap(function (_context) {
       while (1) switch (_context.prev = _context.next) {
         case 0:
@@ -744,14 +699,6 @@ export var loadVectorsAsync = /*#__PURE__*/function () {
           }
           return _context.abrupt("return");
         case 1:
-          if (options !== null && options !== void 0 && options.getBinaryUrl) {
-            _context.next = 2;
-            break;
-          }
-          // eslint-disable-next-line no-console
-          console.warn('[text-predictor] loadVectorsAsync called without a getBinaryUrl — vectors will not load. Pass getVectorsBinaryUrl via plugin options.');
-          return _context.abrupt("return");
-        case 2:
           isLocalLLM = (_options$isLocalLLM = options === null || options === void 0 ? void 0 : options.isLocalLLM) !== null && _options$isLocalLLM !== void 0 ? _options$isLocalLLM : false;
           surface = options === null || options === void 0 ? void 0 : options.surface;
           vectorsLoadStarted = true;
@@ -760,16 +707,16 @@ export var loadVectorsAsync = /*#__PURE__*/function () {
           }, surface ? {
             surface: surface
           } : {}));
-          _context.prev = 3;
-          _context.next = 4;
-          return options.getBinaryUrl();
-        case 4:
+          _context.prev = 2;
+          _context.next = 3;
+          return getArtifactUrl(ARTIFACT_NAME.WORD_VECTORS);
+        case 3:
           url = _context.sent;
-          _context.next = 6;
+          _context.next = 5;
           break;
-        case 5:
-          _context.prev = 5;
-          _t = _context["catch"](3);
+        case 4:
+          _context.prev = 4;
+          _t = _context["catch"](2);
           vectorsLoadStarted = false;
           failExp(EXPERIENCE_NAME.LOAD_VECTORS, 'singleton', _objectSpread({
             isLocalLLM: isLocalLLM,
@@ -780,14 +727,14 @@ export var loadVectorsAsync = /*#__PURE__*/function () {
           // eslint-disable-next-line no-console
           console.warn('[text-predictor] Failed to resolve vectors URL:', _t);
           return _context.abrupt("return");
-        case 6:
-          _context.prev = 6;
-          _context.next = 7;
+        case 5:
+          _context.prev = 5;
+          _context.next = 6;
           return fetch(url);
-        case 7:
+        case 6:
           res = _context.sent;
           if (res.ok) {
-            _context.next = 8;
+            _context.next = 7;
             break;
           }
           vectorsLoadStarted = false;
@@ -801,24 +748,26 @@ export var loadVectorsAsync = /*#__PURE__*/function () {
           // eslint-disable-next-line no-console
           console.warn("[text-predictor] Failed to load vectors: ".concat(res.status));
           return _context.abrupt("return");
+        case 7:
+          _context.next = 8;
+          return res.arrayBuffer();
         case 8:
+          buffer = _context.sent;
+          float32 = new Float32Array(buffer); // The word index is published either bare or wrapped as `{ "index": {…} }`
+          // (the wrapper exists so no real entry, e.g. the word "default", can shadow
+          // a bundler-synthesised ESM export). Detect the wrapper by value type, not
+          // by key presence — "index" is itself a plausible vocabulary word, and in a
+          // bare index every value is a number.
           _context.next = 9;
-          return res.arrayBuffer();
+          return fetchArtifactJson(ARTIFACT_NAME.WORD_INDEX);
         case 9:
-          buffer = _context.sent;
-          float32 = new Float32Array(buffer); // word_index_10k.json is wrapped as `{ "index": {…} }` so no real entry
-          // (e.g. the word "default") can shadow the synthetic ESM `default` export
-          // the bundler creates for dynamically-imported JSON.
-          _context.next = 10;
-          return import( /* webpackChunkName: "@atlaskit-internal_editor-plugin-autocomplete-word-index-10k" */'./data/word_index_10k.json');
-        case 10:
-          wordIndexModule = _context.sent;
-          wordIndexOuter = unwrapJsonModule(wordIndexModule, 'object');
-          wordIndex = (_wordIndexOuter$index = wordIndexOuter === null || wordIndexOuter === void 0 ? void 0 : wordIndexOuter.index) !== null && _wordIndexOuter$index !== void 0 ? _wordIndexOuter$index : {};
+          wordIndexPayload = _context.sent;
+          wrappedIndex = wordIndexPayload === null || wordIndexPayload === void 0 ? void 0 : wordIndexPayload.index;
+          wordIndex = _typeof(wrappedIndex) === 'object' && wrappedIndex !== null ? wrappedIndex : wordIndexPayload !== null && wordIndexPayload !== void 0 ? wordIndexPayload : {};
           nWords = Object.keys(wordIndex).length;
           if (nWords === 0) {
             // eslint-disable-next-line no-console
-            console.warn('[text-predictor] word_index_10k.json missing its `index` wrapper — wordIndex is empty, semantic scoring will be a no-op.');
+            console.warn("[text-predictor] ".concat(ARTIFACT_NAME.WORD_INDEX, " was empty \u2014 semantic scoring will be a no-op."));
           }
           dim = float32.length / nWords;
           vectorStore = {
@@ -842,11 +791,11 @@ export var loadVectorsAsync = /*#__PURE__*/function () {
               sizeBytes: float32.byteLength
             });
           }
-          _context.next = 12;
+          _context.next = 11;
           break;
-        case 11:
-          _context.prev = 11;
-          _t2 = _context["catch"](6);
+        case 10:
+          _context.prev = 10;
+          _t2 = _context["catch"](5);
           vectorsLoadStarted = false;
           failExp(EXPERIENCE_NAME.LOAD_VECTORS, 'singleton', _objectSpread({
             isLocalLLM: isLocalLLM,
@@ -856,11 +805,11 @@ export var loadVectorsAsync = /*#__PURE__*/function () {
           } : {}));
           // eslint-disable-next-line no-console
           console.warn('[text-predictor] Failed to load vectors:', _t2);
-        case 12:
+        case 11:
         case "end":
           return _context.stop();
       }
-    }, _callee, null, [[3, 5], [6, 11]]);
+    }, _callee, null, [[2, 4], [5, 10]]);
   }));
   return function loadVectorsAsync(_x) {
     return _ref6.apply(this, arguments);
@@ -881,7 +830,7 @@ export var loadDefaultVocabulary = function loadDefaultVocabulary(options) {
   var isLocalLLM = (_options$isLocalLLM2 = options === null || options === void 0 ? void 0 : options.isLocalLLM) !== null && _options$isLocalLLM2 !== void 0 ? _options$isLocalLLM2 : false;
   var surface = options === null || options === void 0 ? void 0 : options.surface;
   vocabularyLoadPromise = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee2() {
-    var _yield$Promise$all, _yield$Promise$all2, vocabularyModule, l3VocabularyModule, vocabularyData, l3VocabularyData, terms, _t3;
+    var _yield$Promise$all, _yield$Promise$all2, vocabularyData, l3VocabularyData, terms, _t3;
     return _regeneratorRuntime.wrap(function (_context2) {
       while (1) switch (_context2.prev = _context2.next) {
         case 0:
@@ -892,19 +841,17 @@ export var loadDefaultVocabulary = function loadDefaultVocabulary(options) {
           } : {}));
           _context2.prev = 1;
           _context2.next = 2;
-          return Promise.all([import( /* webpackChunkName: "@atlaskit-internal_editor-plugin-autocomplete-vocabulary-10k" */'./data/vocabulary_10k.json'), import( /* webpackChunkName: "@atlaskit-internal_editor-plugin-autocomplete-l3-vocabulary" */'./data/l3_vocabulary.json')]);
+          return Promise.all([fetchArtifactJson(ARTIFACT_NAME.VOCABULARY), fetchArtifactJson(ARTIFACT_NAME.L3_VOCABULARY)]);
         case 2:
           _yield$Promise$all = _context2.sent;
           _yield$Promise$all2 = _slicedToArray(_yield$Promise$all, 2);
-          vocabularyModule = _yield$Promise$all2[0];
-          l3VocabularyModule = _yield$Promise$all2[1];
-          vocabularyData = unwrapJsonModule(vocabularyModule, 'object');
-          l3VocabularyData = unwrapJsonModule(l3VocabularyModule, 'array');
+          vocabularyData = _yield$Promise$all2[0];
+          l3VocabularyData = _yield$Promise$all2[1];
           if (!((vocabularyData === null || vocabularyData === void 0 ? void 0 : vocabularyData.words) == null || !Array.isArray(l3VocabularyData))) {
             _context2.next = 3;
             break;
           }
-          throw new Error('[text-predictor] vocabulary JSON modules could not be unwrapped');
+          throw new Error("[text-predictor] unexpected vocabulary payload \u2014 " + "".concat(ARTIFACT_NAME.VOCABULARY, " ").concat((vocabularyData === null || vocabularyData === void 0 ? void 0 : vocabularyData.words) == null ? 'is missing its `words` wrapper' : 'is valid', ", ") + "".concat(ARTIFACT_NAME.L3_VOCABULARY, " ").concat(Array.isArray(l3VocabularyData) ? 'is valid' : 'is not an array'));
         case 3:
           terms = Object.entries(vocabularyData.words).map(function (_ref8) {
             var _ref9 = _slicedToArray(_ref8, 2),
@@ -937,7 +884,7 @@ export var loadDefaultVocabulary = function loadDefaultVocabulary(options) {
           _t3 = _context2["catch"](1);
           failExp(EXPERIENCE_NAME.LOAD_VOCABULARY, 'singleton', _objectSpread({
             isLocalLLM: isLocalLLM,
-            errorType: 'parse_error'
+            errorType: 'load_error'
           }, surface ? {
             surface: surface
           } : {}));

package/dist/types/analytics/ufo.d.ts

@@ -30,6 +30,7 @@ export declare const EXPERIENCE_NAME: {
     readonly SLOW_LANE_FETCH: 'slow-lane-fetch';
     readonly LOAD_VOCABULARY: 'load-vocabulary';
     readonly LOAD_VECTORS: 'load-vectors';
+    readonly LOAD_GRAMMAR: 'load-grammar';
 };
 export type AutocompleteExperienceName = (typeof EXPERIENCE_NAME)[keyof typeof EXPERIENCE_NAME];
 export declare const startExp: (name: AutocompleteExperienceName, id: string, metadata?: CustomData) => void;

package/dist/types/pm-plugins/artifacts-manifest.d.ts

@@ -0,0 +1,55 @@
+/**
+ * Artifacts Manifest Client: resolves the autocomplete model artifacts hosted on
+ * the CDN.
+ *
+ * `GET /gateway/api/v1/autocomplete/artifacts` returns one presigned CDN URL per
+ * artifact for the tenant's active model version:
+ *
+ *   { activeVersion, artifacts: [{ name, url, contentType, sizeBytes, checksum }] }
+ *
+ * Every loader in the plugin resolves its payload through here, so no model data
+ * is bundled with the package.
+ */
+/**
+ * Manifest artifact names. These are the manifest's lookup keys and are
+ * hyphenated, unlike the underscored filenames the artifacts are generated from.
+ */
+export declare const ARTIFACT_NAME: {
+    readonly FIRST_TOKEN_TO_WORDS: 'first-token-to-words.json';
+    readonly GHOST_POS_TAGS: 'ghost-pos-tags.json';
+    readonly GRAMMAR_TRANSITIONS: 'grammar-transitions-10k.json';
+    readonly L3_VOCABULARY: 'l3-vocabulary.json';
+    readonly POS_TAGS: 'combined-l2-l3-pos-tags.json';
+    readonly VOCABULARY: 'vocabulary-10k.json';
+    readonly WORD_INDEX: 'word-index-10k.json';
+    readonly WORD_VECTORS: 'word-vectors-10k.bin';
+};
+export type ArtifactName = (typeof ARTIFACT_NAME)[keyof typeof ARTIFACT_NAME];
+/** A single entry in the artifacts manifest. */
+export interface AutocompleteArtifact {
+    checksum: string;
+    contentType: string;
+    name: string;
+    sizeBytes: number;
+    url: string;
+}
+/** Response body of the artifacts endpoint. */
+export interface AutocompleteArtifactsResponse {
+    activeVersion: string;
+    artifacts: AutocompleteArtifact[];
+}
+/** Test-only: drops the memoised manifest so the next lookup re-requests it. */
+export declare const resetArtifactsManifestCache: () => void;
+/**
+ * Resolve the CDN URL for an artifact.
+ *
+ * :raises Error: when the manifest cannot be fetched or omits the artifact.
+ */
+export declare const getArtifactUrl: (name: ArtifactName) => Promise<string>;
+/**
+ * Download and parse a JSON artifact. The presigned URL carries its own auth
+ * token, so the request is sent without Atlassian headers or credentials.
+ *
+ * :raises Error: when the manifest lookup or the download fails.
+ */
+export declare const fetchArtifactJson: <T>(name: ArtifactName) => Promise<T>;

package/dist/types/pm-plugins/autocomplete-plugin.d.ts

@@ -29,22 +29,11 @@ export interface AutocompleteContext {
     siblingCommentsContents?: string[];
 }
 export interface AutocompletePluginOptions {
-    /**
-     * Product/editor surface where autocomplete runs (e.g. "comment", "chat").
-     * Added to analytics and UFO metadata for cross-surface reporting.
-     */
-    surface?: string;
     /**
      * Async function called once on first editor focus to retrieve context for
      * word-frequency boosting. Called lazily so the preset can remain synchronous.
      */
     getContext?: () => Promise<AutocompleteContext | undefined>;
-    /**
-     * Async function that resolves to a URL for the word vectors binary file.
-     * When provided, this takes precedence over the bundled asset URL.
-     * Use this to serve vectors from a CDN or media service in production.
-     */
-    getVectorsBinaryUrl?: () => Promise<string>;
     /**
      * User locale used to determine whether autocomplete should run.
      * Defaults to browser locale when omitted.
@@ -56,6 +45,11 @@ export interface AutocompletePluginOptions {
      * Only meaningful alongside `getContext`; without it each notification is a no-op.
      */
     subscribeToContextUpdates?: (onContextUpdated: () => void) => () => void;
+    /**
+     * Product/editor surface where autocomplete runs (e.g. "comment", "chat").
+     * Added to analytics and UFO metadata for cross-surface reporting.
+     */
+    surface?: string;
     /**
      * When true, uses on-device inference via WebGPU (MLC WebLLM) instead of
      * the network-based slow-lane backend. Defaults to false (network client).

package/dist/types/pm-plugins/scoring-pipeline.d.ts

@@ -3,6 +3,9 @@
  *
  * Operates synchronously on pre-loaded data. Each stage gracefully degrades
  * when its required data isn't available (cold → warm → full warm).
+ *
+ * The grammar/POS payloads are fetched from the CDN by loadGrammarDataAsync, so
+ * the grammar filter is a pass-through until that resolves.
  */
 export interface ScoringCandidate {
     authorFreq: number;
@@ -29,6 +32,17 @@ export interface GrammarFilterMeta {
 export declare const STAGE1_WEIGHT = 0.35;
 export declare const STAGE2_WEIGHT = 0.65;
 export declare const MIN_STAGE1_SCORE = 0.35;
+export declare const isGrammarDataLoaded: () => boolean;
+/**
+ * Fetch the POS tag and grammar transition payloads from the CDN.
+ *
+ * Safe to call repeatedly: the in-flight promise is shared, and a failure is not
+ * cached so a later call can retry.
+ */
+export declare const loadGrammarDataAsync: (options?: {
+    isLocalLLM?: boolean;
+    surface?: string;
+}) => Promise<void>;
 export interface PipelineDebug {
     final: number;
     grammarRejected: string[];

package/dist/types/pm-plugins/text-predictor.d.ts

@@ -81,7 +81,6 @@ export declare const incrementSessionFreq: (word: string) => void;
 export declare const ingestDocumentPage: (pageContent: string | undefined) => void;
 export declare const predict: (textBefore: string) => string | null;
 export declare const loadVectorsAsync: (options?: {
-    getBinaryUrl?: () => Promise<string>;
     isLocalLLM?: boolean;
     surface?: string;
 }) => Promise<void>;

package/package.json

@@ -1,15 +1,12 @@
 {
 	"name": "@atlaskit/editor-plugin-autocomplete",
-	"version": "6.0.0",
+	"version": "8.0.0",
 	"description": "Client-side text autocomplete plugin for @atlaskit/editor-core",
 	"author": "Atlassian Pty Ltd",
 	"license": "Apache-2.0",
 	"publishConfig": {
 		"registry": "https://registry.npmjs.org/"
 	},
-	"scripts": {
-		"ak-postbuild": "ls -d dist/* | xargs -n 1 copyfiles -u 1 -V src/pm-plugins/data/*.json"
-	},
 	"atlassian": {
 		"team": "Editor: AI",
 		"singleton": true
@@ -30,8 +27,8 @@
 		"wink-nlp": "^2.4.0"
 	},
 	"peerDependencies": {
-		"@atlaskit/editor-common": "^118.0.0",
-		"@atlaskit/editor-plugin-analytics": "^14.0.0",
+		"@atlaskit/editor-common": "^119.0.0",
+		"@atlaskit/editor-plugin-analytics": "^15.0.0",
 		"react": "^18.2.0 || ^19.2.0"
 	},
 	"techstack": {

package/src/analytics/ufo.ts

@@ -37,6 +37,7 @@ export const EXPERIENCE_NAME = {
 	SLOW_LANE_FETCH: 'slow-lane-fetch',
 	LOAD_VOCABULARY: 'load-vocabulary',
 	LOAD_VECTORS: 'load-vectors',
+	LOAD_GRAMMAR: 'load-grammar',
 } as const;
 
 export type AutocompleteExperienceName = (typeof EXPERIENCE_NAME)[keyof typeof EXPERIENCE_NAME];
@@ -69,6 +70,10 @@ const experiences: Record<AutocompleteExperienceName, ConcurrentExperience> = {
 		EXPERIENCE_NAME.LOAD_VECTORS,
 		operationConfig,
 	),
+	[EXPERIENCE_NAME.LOAD_GRAMMAR]: new ConcurrentExperience(
+		EXPERIENCE_NAME.LOAD_GRAMMAR,
+		operationConfig,
+	),
 };
 
 export const startExp = (

package/src/pm-plugins/artifacts-manifest.ts

@@ -0,0 +1,151 @@
+/**
+ * Artifacts Manifest Client: resolves the autocomplete model artifacts hosted on
+ * the CDN.
+ *
+ * `GET /gateway/api/v1/autocomplete/artifacts` returns one presigned CDN URL per
+ * artifact for the tenant's active model version:
+ *
+ *   { activeVersion, artifacts: [{ name, url, contentType, sizeBytes, checksum }] }
+ *
+ * Every loader in the plugin resolves its payload through here, so no model data
+ * is bundled with the package.
+ */
+
+import { isAutocompleteDebugEnabled } from './debug-mode';
+
+// ─── Types ───────────────────────────────────────────────────────────────────
+
+/**
+ * Manifest artifact names. These are the manifest's lookup keys and are
+ * hyphenated, unlike the underscored filenames the artifacts are generated from.
+ */
+export const ARTIFACT_NAME = {
+	FIRST_TOKEN_TO_WORDS: 'first-token-to-words.json',
+	GHOST_POS_TAGS: 'ghost-pos-tags.json',
+	GRAMMAR_TRANSITIONS: 'grammar-transitions-10k.json',
+	L3_VOCABULARY: 'l3-vocabulary.json',
+	POS_TAGS: 'combined-l2-l3-pos-tags.json',
+	VOCABULARY: 'vocabulary-10k.json',
+	WORD_INDEX: 'word-index-10k.json',
+	WORD_VECTORS: 'word-vectors-10k.bin',
+} as const;
+
+export type ArtifactName = (typeof ARTIFACT_NAME)[keyof typeof ARTIFACT_NAME];
+
+/** A single entry in the artifacts manifest. */
+export interface AutocompleteArtifact {
+	checksum: string;
+	contentType: string;
+	name: string;
+	sizeBytes: number;
+	url: string;
+}
+
+/** Response body of the artifacts endpoint. */
+export interface AutocompleteArtifactsResponse {
+	activeVersion: string;
+	artifacts: AutocompleteArtifact[];
+}
+
+// ─── Constants ───────────────────────────────────────────────────────────────
+
+const MANIFEST_ENDPOINT = '/gateway/api/v1/autocomplete/artifacts';
+
+/**
+ * Presigned artifact URLs expire ~10 minutes after the manifest is issued, so
+ * the manifest is re-requested well inside that window instead of being cached
+ * for the session and handing out dead URLs to a late loader.
+ */
+const MANIFEST_TTL_MS = 5 * 60 * 1000;
+
+// ─── Manifest ────────────────────────────────────────────────────────────────
+
+let manifestPromise: Promise<Map<string, AutocompleteArtifact>> | undefined;
+let manifestRequestedAt = 0;
+
+/** Test-only: drops the memoised manifest so the next lookup re-requests it. */
+export const resetArtifactsManifestCache = (): void => {
+	manifestPromise = undefined;
+	manifestRequestedAt = 0;
+};
+
+const requestManifest = async (): Promise<Map<string, AutocompleteArtifact>> => {
+	const res = await fetch(MANIFEST_ENDPOINT, {
+		method: 'GET',
+		headers: {
+			'x-experience-id': 'confluence-smart-typeahead-artifacts',
+			'x-product': 'confluence',
+		},
+	});
+
+	if (!res.ok) {
+		throw new Error(`[autocomplete-artifacts] Manifest request failed: ${res.status}`);
+	}
+
+	const body = (await res.json()) as AutocompleteArtifactsResponse;
+
+	if (!Array.isArray(body?.artifacts) || body.artifacts.length === 0) {
+		throw new Error('[autocomplete-artifacts] Manifest response contained no artifacts');
+	}
+
+	if (isAutocompleteDebugEnabled()) {
+		// eslint-disable-next-line no-console
+		console.log('[autocomplete-artifacts] Manifest loaded:', {
+			activeVersion: body.activeVersion,
+			names: body.artifacts.map((artifact) => artifact.name),
+		});
+	}
+
+	return new Map(body.artifacts.map((artifact) => [artifact.name, artifact]));
+};
+
+const getManifest = (): Promise<Map<string, AutocompleteArtifact>> => {
+	if (manifestPromise && Date.now() - manifestRequestedAt < MANIFEST_TTL_MS) {
+		return manifestPromise;
+	}
+
+	manifestRequestedAt = Date.now();
+	manifestPromise = requestManifest().catch((e) => {
+		// Never cache a rejection — a transient failure would otherwise stop every
+		// later loader in the session from retrying.
+		resetArtifactsManifestCache();
+		throw e;
+	});
+
+	return manifestPromise;
+};
+
+// ─── Public API ──────────────────────────────────────────────────────────────
+
+/**
+ * Resolve the CDN URL for an artifact.
+ *
+ * :raises Error: when the manifest cannot be fetched or omits the artifact.
+ */
+export const getArtifactUrl = async (name: ArtifactName): Promise<string> => {
+	const manifest = await getManifest();
+	const url = manifest.get(name)?.url;
+
+	if (!url) {
+		throw new Error(`[autocomplete-artifacts] Manifest has no URL for "${name}"`);
+	}
+
+	return url;
+};
+
+/**
+ * Download and parse a JSON artifact. The presigned URL carries its own auth
+ * token, so the request is sent without Atlassian headers or credentials.
+ *
+ * :raises Error: when the manifest lookup or the download fails.
+ */
+export const fetchArtifactJson = async <T>(name: ArtifactName): Promise<T> => {
+	const url = await getArtifactUrl(name);
+	const res = await fetch(url);
+
+	if (!res.ok) {
+		throw new Error(`[autocomplete-artifacts] Failed to download "${name}": ${res.status}`);
+	}
+
+	return (await res.json()) as T;
+};

package/src/pm-plugins/autocomplete-plugin.ts

@@ -22,6 +22,7 @@ import {
 	type LocalSlowLaneLoadError,
 	type LocalSlowLaneLoadSuccess,
 } from './local-slow-lane-client';
+import { loadGrammarDataAsync } from './scoring-pipeline';
 import { createSlowLaneClient, setDefaultSlowLaneClient, isWordBoundary } from './slow-lane-client';
 import {
 	predict,
@@ -224,22 +225,11 @@ export interface AutocompleteContext {
 }
 
 export interface AutocompletePluginOptions {
-	/**
-	 * Product/editor surface where autocomplete runs (e.g. "comment", "chat").
-	 * Added to analytics and UFO metadata for cross-surface reporting.
-	 */
-	surface?: string;
 	/**
 	 * Async function called once on first editor focus to retrieve context for
 	 * word-frequency boosting. Called lazily so the preset can remain synchronous.
 	 */
 	getContext?: () => Promise<AutocompleteContext | undefined>;
-	/**
-	 * Async function that resolves to a URL for the word vectors binary file.
-	 * When provided, this takes precedence over the bundled asset URL.
-	 * Use this to serve vectors from a CDN or media service in production.
-	 */
-	getVectorsBinaryUrl?: () => Promise<string>;
 	/**
 	 * User locale used to determine whether autocomplete should run.
 	 * Defaults to browser locale when omitted.
@@ -251,6 +241,11 @@ export interface AutocompletePluginOptions {
 	 * Only meaningful alongside `getContext`; without it each notification is a no-op.
 	 */
 	subscribeToContextUpdates?: (onContextUpdated: () => void) => () => void;
+	/**
+	 * Product/editor surface where autocomplete runs (e.g. "comment", "chat").
+	 * Added to analytics and UFO metadata for cross-surface reporting.
+	 */
+	surface?: string;
 	/**
 	 * When true, uses on-device inference via WebGPU (MLC WebLLM) instead of
 	 * the network-based slow-lane backend. Defaults to false (network client).
@@ -831,7 +826,6 @@ export const createAutocompletePlugin = (
 						});
 					});
 					loadVectorsAsync({
-						getBinaryUrl: options?.getVectorsBinaryUrl,
 						isLocalLLM: options?.useLocalModel ?? false,
 						surface,
 					}).catch((error) => {
@@ -839,6 +833,14 @@ export const createAutocompletePlugin = (
 							location: 'editor-plugin-autocomplete/loadVectorsAsync',
 						});
 					});
+					loadGrammarDataAsync({
+						isLocalLLM: options?.useLocalModel ?? false,
+						surface,
+					}).catch((error) => {
+						logException(error as Error, {
+							location: 'editor-plugin-autocomplete/loadGrammarDataAsync',
+						});
+					});
 					if (!hasIngestedPage) {
 						hasIngestedPage = true;
 						refreshContext({