@atlaskit/editor-plugin-autocomplete 0.1.0 → 0.2.0

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

@@ -0,0 +1,783 @@
+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";
+import _defineProperty from "@babel/runtime/helpers/defineProperty";
+import _regeneratorRuntime from "@babel/runtime/regenerator";
+function _createForOfIteratorHelper(r, e) { var t = "undefined" != typeof Symbol && r[Symbol.iterator] || r["@@iterator"]; if (!t) { if (Array.isArray(r) || (t = _unsupportedIterableToArray(r)) || e && r && "number" == typeof r.length) { t && (r = t); var _n = 0, F = function F() {}; return { s: F, n: function n() { return _n >= r.length ? { done: !0 } : { done: !1, value: r[_n++] }; }, e: function e(r) { throw r; }, f: F }; } throw new TypeError("Invalid attempt to iterate non-iterable instance.\nIn order to be iterable, non-array objects must have a [Symbol.iterator]() method."); } var o, a = !0, u = !1; return { s: function s() { t = t.call(r); }, n: function n() { var r = t.next(); return a = r.done, r; }, e: function e(r) { u = !0, o = r; }, f: function f() { try { a || null == t.return || t.return(); } finally { if (u) throw o; } } }; }
+function _unsupportedIterableToArray(r, a) { if (r) { if ("string" == typeof r) return _arrayLikeToArray(r, a); var t = {}.toString.call(r).slice(8, -1); return "Object" === t && r.constructor && (t = r.constructor.name), "Map" === t || "Set" === t ? Array.from(r) : "Arguments" === t || /^(?:Ui|I)nt(?:8|16|32)(?:Clamped)?Array$/.test(t) ? _arrayLikeToArray(r, a) : void 0; } }
+function _arrayLikeToArray(r, a) { (null == a || a > r.length) && (a = r.length); for (var e = 0, n = Array(a); e < a; e++) n[e] = r[e]; return n; }
+/**
+ * Fast Lane Predictor: Local autocomplete using weighted trie + frequency + semantic scoring.
+ *
+ * Two prediction modes:
+ *   1. Word boundary → bigram-based next-word suggestion (grammar-filtered)
+ *   2. Mid-word (≥3 chars) → trie prefix search → scoring pipeline → top result
+ *
+ * Scoring is delegated to scoring-pipeline.ts which handles:
+ *   Stage 1 (semantic + frequency), grammar filter, Stage 2 (optional LM re-ranking).
+ *
+ * Context vector: average of word vectors from text before cursor (last N words).
+ * Falls back to cold mode (freq-only) when vectors not yet loaded.
+ *
+ * Session personalization (L1): words the user types are incrementally boosted
+ * via incrementSessionFreq(), called on word boundaries from the plugin.
+ */
+
+// import bigramsData from './data/bigrams.json';
+import l3VocabularyData from './data/l3_vocabulary.json';
+import vocabularyData from './data/vocabulary_10k.json';
+import wordIndexData from './data/word_index_10k.json';
+// import { rankCandidates, isGrammarAllowed } from './scoring-pipeline';
+import { rankCandidates } from './scoring-pipeline';
+import { getStoredContextVector, getStoredLmLogits } from './slow-lane-client';
+
+// ─── Constants ───────────────────────────────────────────────────────────────
+
+var PUNCTUATION_BOUNDARY_REGEX = /^[!"'-\),\.:;\?\[\]`\{\}]+|[!"'-\),\.:;\?\[\]`\{\}]+$/g;
+var MIN_PREFIX_LENGTH = 3;
+var MAX_CANDIDATES = 200;
+var CONTEXT_WORDS = 10;
+var MIN_SCORE_THRESHOLD = 0.2;
+var L3_BASELINE_FREQ = 0.001;
+
+// ─── Types ───────────────────────────────────────────────────────────────────
+var TrieNode = /*#__PURE__*/_createClass(function TrieNode() {
+  _classCallCheck(this, TrieNode);
+  _defineProperty(this, "children", new Map());
+  _defineProperty(this, "word", null);
+  _defineProperty(this, "tenantFreq", 0);
+  _defineProperty(this, "docFreq", 0);
+  _defineProperty(this, "authorFreq", 0);
+  _defineProperty(this, "sessionFreq", 0);
+});
+var WeightedWordTrie = /*#__PURE__*/function () {
+  function WeightedWordTrie() {
+    _classCallCheck(this, WeightedWordTrie);
+    _defineProperty(this, "root", new TrieNode());
+    /** Highest tenantFreq seen — used to normalize freq scores at query time */
+    _defineProperty(this, "maxTenantFreq", 1);
+  }
+  return _createClass(WeightedWordTrie, [{
+    key: "insert",
+    value: function insert(word, tenantFreq, docFreq, authorFreq) {
+      var node = this.root;
+      var _iterator = _createForOfIteratorHelper(word.toLowerCase()),
+        _step;
+      try {
+        for (_iterator.s(); !(_step = _iterator.n()).done;) {
+          var char = _step.value;
+          var next = node.children.get(char);
+          if (!next) {
+            next = new TrieNode();
+            node.children.set(char, next);
+          }
+          node = next;
+        }
+      } catch (err) {
+        _iterator.e(err);
+      } finally {
+        _iterator.f();
+      }
+      node.word = word;
+      node.tenantFreq = tenantFreq;
+      node.docFreq = docFreq;
+      node.authorFreq = authorFreq;
+      if (tenantFreq > this.maxTenantFreq) {
+        this.maxTenantFreq = tenantFreq;
+      }
+    }
+
+    /**
+     * Return all words matching this prefix, up to maxResults.
+     * O(prefix_length + results) — traverses to the prefix node then collects subtree.
+     */
+  }, {
+    key: "getCandidates",
+    value: function getCandidates(prefix) {
+      var maxResults = arguments.length > 1 && arguments[1] !== undefined ? arguments[1] : MAX_CANDIDATES;
+      var node = this.root;
+      var _iterator2 = _createForOfIteratorHelper(prefix.toLowerCase()),
+        _step2;
+      try {
+        for (_iterator2.s(); !(_step2 = _iterator2.n()).done;) {
+          var char = _step2.value;
+          var next = node.children.get(char);
+          if (!next) {
+            return [];
+          }
+          node = next;
+        }
+      } catch (err) {
+        _iterator2.e(err);
+      } finally {
+        _iterator2.f();
+      }
+      var candidates = [];
+      var stack = [node];
+      while (stack.length > 0 && candidates.length < maxResults) {
+        var current = stack.pop();
+        if (!current) {
+          continue;
+        }
+
+        // Only add candidates that are longer than the prefix
+        if (current.word && current.word.length > prefix.length) {
+          candidates.push({
+            word: current.word,
+            node: current
+          });
+        }
+        var _iterator3 = _createForOfIteratorHelper(current.children.values()),
+          _step3;
+        try {
+          for (_iterator3.s(); !(_step3 = _iterator3.n()).done;) {
+            var child = _step3.value;
+            stack.push(child);
+          }
+        } catch (err) {
+          _iterator3.e(err);
+        } finally {
+          _iterator3.f();
+        }
+      }
+      return candidates;
+    }
+  }, {
+    key: "findNode",
+    value: function findNode(word) {
+      var node = this.root;
+      var _iterator4 = _createForOfIteratorHelper(word.toLowerCase()),
+        _step4;
+      try {
+        for (_iterator4.s(); !(_step4 = _iterator4.n()).done;) {
+          var char = _step4.value;
+          var next = node.children.get(char);
+          if (!next) {
+            return null;
+          }
+          node = next;
+        }
+      } catch (err) {
+        _iterator4.e(err);
+      } finally {
+        _iterator4.f();
+      }
+      return node.word !== null ? node : null;
+    }
+
+    /**
+     * Set the session frequency for a word.
+     * Returns true if the word exists in the trie.
+     */
+  }, {
+    key: "updateSessionFreq",
+    value: function updateSessionFreq(word, count) {
+      var node = this.findNode(word);
+      if (!node) {
+        return false;
+      }
+      node.sessionFreq = count;
+      return true;
+    }
+
+    /**
+     * Increment the session frequency for a word by 1.
+     * Returns true if the word exists in the trie.
+     */
+  }, {
+    key: "incrementSessionFreq",
+    value: function incrementSessionFreq(word) {
+      var node = this.findNode(word);
+      if (!node) {
+        return false;
+      }
+      node.sessionFreq += 1;
+      return true;
+    }
+  }]);
+}(); // L1/L2 Trie (Session + Atlassian Domain)
+var wordTrie = new WeightedWordTrie();
+
+// L3 Trie (General English Fallback)
+var l3Trie = new WeightedWordTrie();
+
+// --- Initialization Function ---
+/**
+ * Loads the General English vocabulary.
+ * expects a simple array of strings: ["about", "above", "actually", ...]
+ */
+export var initL3Vocabulary = function initL3Vocabulary(l3Words) {
+  var _iterator5 = _createForOfIteratorHelper(l3Words),
+    _step5;
+  try {
+    for (_iterator5.s(); !(_step5 = _iterator5.n()).done;) {
+      var word = _step5.value;
+      // Insert with a tiny baseline frequency so it mathematically
+      // loses to any domain word in Stage 1, but still scores above 0.
+      l3Trie.insert(word, L3_BASELINE_FREQ, 0, 0);
+    }
+  } catch (err) {
+    _iterator5.e(err);
+  } finally {
+    _iterator5.f();
+  }
+  if (debugMode) {
+    // eslint-disable-next-line no-console
+    console.log("[text-predictor] L3 General English loaded: ".concat(l3Words.length, " words"));
+  }
+};
+
+// const bigramMap: Map<string, Record<string, number>> = new Map(
+// 	Object.entries(bigramsData as Record<string, Record<string, number>>),
+// );
+
+var isInitialized = false;
+var vectorStore = null;
+var vectorsLoadStarted = false;
+var debugMode = true;
+var lastPredictionDebug = null;
+var hasLoggedSemanticActive = false;
+
+/** Get vector for a word from the store. */
+var getWordVector = function getWordVector(word) {
+  if (!vectorStore) {
+    return null;
+  }
+  var idx = vectorStore.wordIndex[word.toLowerCase()];
+  if (idx === undefined) {
+    return null;
+  }
+  var start = idx * vectorStore.dim;
+  return vectorStore.float32.subarray(start, start + vectorStore.dim);
+};
+
+/**
+ * Compute context vector by averaging vectors of last N words in text.
+ * Falls back to null (cold mode) if no words have vectors.
+ */
+var computeContextVectorLocal = function computeContextVectorLocal(textBefore) {
+  if (!vectorStore) {
+    return null;
+  }
+  var tokens = tokenize(textBefore);
+  var words = tokens.slice(-CONTEXT_WORDS);
+  var vectors = [];
+  var _iterator6 = _createForOfIteratorHelper(words),
+    _step6;
+  try {
+    for (_iterator6.s(); !(_step6 = _iterator6.n()).done;) {
+      var word = _step6.value;
+      var _v = getWordVector(word);
+      if (_v) {
+        vectors.push(_v);
+      }
+    }
+  } catch (err) {
+    _iterator6.e(err);
+  } finally {
+    _iterator6.f();
+  }
+  if (vectors.length === 0) {
+    return null;
+  }
+  var dim = vectorStore.dim;
+  var avg = new Float32Array(dim);
+  for (var _i = 0, _vectors = vectors; _i < _vectors.length; _i++) {
+    var v = _vectors[_i];
+    for (var i = 0; i < dim; i++) {
+      avg[i] += v[i];
+    }
+  }
+  for (var _i2 = 0; _i2 < dim; _i2++) {
+    avg[_i2] /= vectors.length;
+  }
+  return avg;
+};
+
+/**
+ * Get context vector for scoring. Prefers Slow Lane (BE) context when available
+ * and dimension matches; otherwise falls back to local averaging.
+ */
+var getContextVectorForScoring = function getContextVectorForScoring(textBefore) {
+  var slowLaneVector = getStoredContextVector();
+  if (slowLaneVector && vectorStore && slowLaneVector.length === vectorStore.dim) {
+    return slowLaneVector;
+  }
+  return computeContextVectorLocal(textBefore);
+};
+var tokenize = function tokenize(text) {
+  var tokens = [];
+  // eslint-disable-next-line @atlassian/perf-linting/no-expensive-split-replace
+  var _iterator7 = _createForOfIteratorHelper(text.toLowerCase().split(/[\t-\r \xA0\u1680\u2000-\u200A\u2028\u2029\u202F\u205F\u3000\uFEFF]+/)),
+    _step7;
+  try {
+    for (_iterator7.s(); !(_step7 = _iterator7.n()).done;) {
+      var raw = _step7.value;
+      // eslint-disable-next-line @atlassian/perf-linting/no-expensive-split-replace
+      var clean = raw.replace(PUNCTUATION_BOUNDARY_REGEX, '');
+      if (clean.length >= 2) {
+        tokens.push(clean);
+      }
+    }
+  } catch (err) {
+    _iterator7.e(err);
+  } finally {
+    _iterator7.f();
+  }
+  return tokens;
+};
+var extractPreviousWord = function extractPreviousWord(text) {
+  // 1. Split the text by newlines or punctuation (. ? !)
+  var sentences = text.split(/[\n!\.\?]+/);
+
+  // 2. Only look at the current sentence/line the user is typing in
+  var currentSentence = sentences[sentences.length - 1];
+
+  // 3. Extract the previous word as normal
+  var words = currentSentence.trimEnd().split(/[\t-\r \xA0\u1680\u2000-\u200A\u2028\u2029\u202F\u205F\u3000\uFEFF]+/);
+  return words.length >= 2 ? words[words.length - 2] : '';
+};
+
+// ─── Debug Helpers ───────────────────────────────────────────────────────────
+
+/** Enable or disable debug logging. Also checks localStorage key `autocomplete-debug`. */
+export var setDebugMode = function setDebugMode(enabled) {
+  debugMode = enabled;
+};
+
+/** Check localStorage for autocomplete-debug on first access. */
+var ensureDebugModeFromStorage = function ensureDebugModeFromStorage() {
+  if (typeof localStorage !== 'undefined' && localStorage.getItem('autocomplete-debug') === '1') {
+    debugMode = true;
+  }
+};
+
+/**
+ * Get predictor status for debugging.
+ * vectorsLoaded: true when semantic scoring is active
+ * wordCount: number of words in vector store (0 if not loaded)
+ */
+export var getPredictorStatus = function getPredictorStatus() {
+  ensureDebugModeFromStorage();
+  return {
+    vectorsLoaded: vectorStore !== null,
+    wordCount: vectorStore ? Object.keys(vectorStore.wordIndex).length : 0,
+    vectorsLoadStarted: vectorsLoadStarted,
+    isInitialized: isInitialized
+  };
+};
+
+/**
+ * Get details of the last prediction (for debugging).
+ * Returns null if no prediction has run yet or debug was off.
+ */
+export var getLastPredictionDebug = function getLastPredictionDebug() {
+  ensureDebugModeFromStorage();
+  return lastPredictionDebug;
+};
+export var initVocabulary = function initVocabulary(vocabulary) {
+  var _iterator8 = _createForOfIteratorHelper(vocabulary.terms),
+    _step8;
+  try {
+    for (_iterator8.s(); !(_step8 = _iterator8.n()).done;) {
+      var term = _step8.value;
+      wordTrie.insert(term.word, term.freq, term.docFreq, term.authorFreq);
+    }
+  } catch (err) {
+    _iterator8.e(err);
+  } finally {
+    _iterator8.f();
+  }
+  isInitialized = true;
+};
+
+/**
+ * Increment L1 session frequency for a single word.
+ * Called from the plugin on word boundaries for efficient incremental boosting.
+ */
+export var incrementSessionFreq = function incrementSessionFreq(word) {
+  wordTrie.incrementSessionFreq(word);
+};
+
+/**
+ * Prime session frequencies from a document page string.
+ *
+ * Iterates through every token in `pageContent` and increments its session
+ * frequency so that words already present on the page receive an L1 boost
+ * before the user starts typing.
+ *
+ * Pass `undefined` (or omit the argument) to skip priming — useful when the
+ * calling context does not yet have a page value available.
+ */
+// NOTE: We ingest full page context here
+export var ingestDocumentPage = function ingestDocumentPage(pageContent) {
+  if (!pageContent) {
+    return;
+  }
+  var words = tokenize(pageContent);
+  var validBoostedWords = new Set();
+  var _iterator9 = _createForOfIteratorHelper(words),
+    _step9;
+  try {
+    for (_iterator9.s(); !(_step9 = _iterator9.n()).done;) {
+      var word = _step9.value;
+      var didBoost = wordTrie.incrementSessionFreq(word);
+      if (didBoost) {
+        validBoostedWords.add(word);
+      }
+    }
+  } catch (err) {
+    _iterator9.e(err);
+  } finally {
+    _iterator9.f();
+  }
+  if (debugMode && validBoostedWords.size > 0) {
+    // eslint-disable-next-line no-console
+    console.groupCollapsed("%c[L1 Session] %cPrimed ".concat(validBoostedWords.size, " valid dictionary words from page"), 'color: #00b8d9; font-weight: bold;', 'color: inherit; font-style: italic;');
+    // eslint-disable-next-line no-console
+    console.dir(Array.from(validBoostedWords).sort());
+    // eslint-disable-next-line no-console
+    console.groupEnd();
+  }
+};
+export var predict = function predict(textBefore) {
+  ensureDebugModeFromStorage();
+  if (!isInitialized) {
+    loadDefaultVocabulary();
+  }
+  var t0 = performance.now();
+
+  // ── Step 1: Bigram-based next-word suggestion at word boundary ───────────
+  // if (textBefore.length > 0 && /\s$/u.test(textBefore)) {
+  // 	const words = textBefore.toLowerCase().trimEnd().split(/\s+/u);
+  // 	const prevWord = words[words.length - 1];
+  // 	const nextWords = bigramMap.get(prevWord);
+  // 	if (nextWords) {
+  // 		const sorted = Object.entries(nextWords).sort((a, b) => b[1] - a[1]);
+  // 		let bestWord = '';
+  // 		for (const [word] of sorted) {
+  // 			if (!isGrammarAllowed(prevWord, word)) {
+  // 				continue;
+  // 			}
+  // 			bestWord = word;
+  // 			break;
+  // 		}
+  // 		if (bestWord) {
+  // 			if (debugMode) {
+  // 				const latencyMs = performance.now() - t0;
+  // 				console.log(
+  // 					'%c[autocomplete] BIGRAM',
+  // 					'color:cyan',
+  // 					'| "' + prevWord + '" -> "' + bestWord + '" | ' + latencyMs.toFixed(1) + 'ms',
+  // 				);
+  // 			}
+  // 			return bestWord;
+  // 		}
+  // 	}
+  // 	if (debugMode) {
+  // 		console.log(
+  // 			'%c[autocomplete] BIGRAM-MISS',
+  // 			'color:gray',
+  // 			'| no bigram for "' + words[words.length - 1] + '", skipping prefix completion',
+  // 		);
+  // 	}
+  // 	return null;
+  // }
+
+  // ── Step 2: Prefix completion (≥3 chars typed) ──────────────────────────
+  if (textBefore.length > 0 && /[\t-\r \xA0\u1680\u2000-\u200A\u2028\u2029\u202F\u205F\u3000\uFEFF]$/.test(textBefore)) {
+    return null;
+  }
+  var trimmed = textBefore.trimEnd();
+  var lastSpaceIdx = trimmed.lastIndexOf(' ');
+  var currentWord = lastSpaceIdx === -1 ? trimmed : trimmed.slice(lastSpaceIdx + 1);
+  if (currentWord.length < MIN_PREFIX_LENGTH) {
+    return null;
+  }
+
+  // 1. Primary Query: Ask the L2 Domain Trie
+  var candidates = wordTrie.getCandidates(currentWord, MAX_CANDIDATES);
+
+  // 2. Fallback Query: Gap-fill with the L3 General English Trie
+  if (candidates.length < MAX_CANDIDATES) {
+    // Ask L3 for MAX_CANDIDATES to guarantee we have enough buffer
+    // to survive the deduplication process.
+    var l3Candidates = l3Trie.getCandidates(currentWord, MAX_CANDIDATES);
+    var existingWords = new Set(candidates.map(function (c) {
+      return c.word;
+    }));
+    var _iterator0 = _createForOfIteratorHelper(l3Candidates),
+      _step0;
+    try {
+      for (_iterator0.s(); !(_step0 = _iterator0.n()).done;) {
+        var l3c = _step0.value;
+        if (candidates.length >= MAX_CANDIDATES) break; // Stop exactly at the limit
+
+        if (!existingWords.has(l3c.word)) {
+          candidates.push(l3c);
+        }
+      }
+    } catch (err) {
+      _iterator0.e(err);
+    } finally {
+      _iterator0.f();
+    }
+  }
+
+  // If both Tries are completely empty for this prefix
+  if (candidates.length === 0) {
+    return null;
+  }
+  var previousWord = extractPreviousWord(trimmed);
+  var contextVector = getContextVectorForScoring(trimmed);
+  var lmLogits = getStoredLmLogits();
+
+  // Raw LM Output Logger
+  if (debugMode && lmLogits && Object.keys(lmLogits).length > 0) {
+    var rawLmTop = Object.entries(lmLogits).sort(function (a, b) {
+      return b[1] - a[1];
+    }).slice(0, 5).map(function (_ref) {
+      var _ref2 = _slicedToArray(_ref, 2),
+        word = _ref2[0],
+        score = _ref2[1];
+      return {
+        Word: word,
+        Prob: Number(score.toFixed(5))
+      };
+    });
+    // eslint-disable-next-line no-console
+    console.log('%c[Raw LM Prediction]🧠', 'color: #e83e8c; font-weight: bold;', rawLmTop);
+  }
+  var mode = contextVector ? 'warm' : 'cold';
+  if (debugMode && contextVector && !hasLoggedSemanticActive) {
+    hasLoggedSemanticActive = true;
+  }
+
+  // Build ScoringCandidate array from TrieNodes
+  var scoringCandidates = candidates.map(function (_ref3) {
+    var word = _ref3.word,
+      node = _ref3.node;
+    return {
+      word: word,
+      tenantFreq: node.tenantFreq,
+      docFreq: node.docFreq,
+      authorFreq: node.authorFreq,
+      sessionFreq: node.sessionFreq
+    };
+  });
+  var _rankCandidates = rankCandidates(scoringCandidates, contextVector, function (w) {
+      return getWordVector(w);
+    }, lmLogits, wordTrie.maxTenantFreq, previousWord),
+    ranked = _rankCandidates.candidates,
+    grammarMeta = _rankCandidates.grammarMeta;
+  var best = ranked[0];
+  var suggestion = best && best.finalScore >= MIN_SCORE_THRESHOLD ? best.word.slice(currentWord.length) : null;
+  if (debugMode) {
+    var latencyMs = performance.now() - t0;
+    var tokens = tokenize(trimmed);
+    var contextWords = tokens.slice(-CONTEXT_WORDS);
+    var belowThreshold = best && best.finalScore < MIN_SCORE_THRESHOLD;
+    var suggestionLabel = belowThreshold ? '🚫 (below threshold)' : suggestion && suggestion.length > 0 ? "\u2728 \"".concat(suggestion, "\"") : '🚫 (no match)';
+    lastPredictionDebug = {
+      textBefore: trimmed,
+      currentWord: currentWord,
+      mode: mode,
+      contextWords: contextWords,
+      topCandidates: ranked.slice(0, 5).map(function (r) {
+        return {
+          word: r.word,
+          finalScore: r.finalScore,
+          semanticScore: r.semanticScore,
+          freqScore: r.freqScore,
+          lmScore: r.lmScore
+        };
+      }),
+      suggestion: suggestion && suggestion.length > 0 ? suggestion : null
+    };
+
+    // ── Mode label: COLD / WARM(local) / WARM(BE) ───────────────────────
+    var slowLaneVec = getStoredContextVector();
+    var isUsingSlowLaneVector = slowLaneVec !== null && vectorStore !== null && slowLaneVec.length === vectorStore.dim;
+    var modeLabel = !contextVector ? 'COLD' : isUsingSlowLaneVector ? 'WARM(BE)' : 'WARM(local)';
+    var modeColor = modeLabel === 'WARM(BE)' ? 'color: #ff9800; font-weight: bold;' : modeLabel === 'WARM(local)' ? 'color: #4caf50; font-weight: bold;' : 'color: #9e9e9e; font-weight: bold;';
+
+    // 1. Collapsible group header
+    // eslint-disable-next-line no-console
+    console.groupCollapsed("%c[Autocomplete] %c".concat(modeLabel, " %c| \"").concat(currentWord, "\" \u2794 ").concat(suggestionLabel, " | \u23F1 ").concat(latencyMs.toFixed(1), "ms"), 'color: #00b8d9; font-weight: bold;', modeColor, 'color: inherit; font-weight: normal;');
+
+    // 2. Context window (what local vector averaging sees)
+    // eslint-disable-next-line no-console
+    console.log('%cContext Window:', 'color: #888; font-style: italic;', contextWords.length ? contextWords.join(' ') : '(none)');
+
+    // 3. Slow-lane status
+    var vectorStatus = isUsingSlowLaneVector ? "\u2705 BE semantic vector (dim=".concat(slowLaneVec === null || slowLaneVec === void 0 ? void 0 : slowLaneVec.length, ")") : vectorStore ? '⚠️ Local vector average (slow-lane not yet returned)' : '❌ No vectors (cold)';
+    var logitsStatus = lmLogits && Object.keys(lmLogits).length > 0 ? "\u2705 LM logits active (".concat(Object.keys(lmLogits).length, " tokens)") : '⏳ No LM logits (slow-lane pending or failed)';
+    // eslint-disable-next-line no-console
+    console.log('%cSlow Lane:', 'color: #888; font-style: italic;', vectorStatus, '|', logitsStatus);
+
+    // 4. Scoring formula active this prediction
+    var formulaLabel = lmLogits && Object.keys(lmLogits).length > 0 ? 'Stage1(×0.6) + LM(×0.4)' : 'Stage1 only (no LM logits)';
+    // eslint-disable-next-line no-console
+    console.log('%cFormula:', 'color: #888; font-style: italic;', formulaLabel);
+
+    // 5. Grammar filter result (collected inside rankCandidates, logged here)
+    if (grammarMeta) {
+      // eslint-disable-next-line no-console
+      console.log("%c[Grammar] \"".concat(grammarMeta.prevWord, "\" [").concat(grammarMeta.prevTags.join('|'), "] \u2192 ").concat(grammarMeta.before, " candidates \u2192 ").concat(grammarMeta.after, " after filter"), 'color: #4caf50; font-weight: bold;');
+      if (grammarMeta.dropped.length > 0) {
+        // eslint-disable-next-line no-console
+        console.log("%c\uD83D\uDEAB Dropped: ".concat(grammarMeta.dropped.join(', ')), 'color: #f44336; font-style: italic;');
+      }
+    }
+
+    // 6. Candidate table
+    if (ranked.length > 0) {
+      var lmCoverage = ranked.slice(0, 10).filter(function (r) {
+        return r.lmScore > 0.05;
+      }).length;
+      // eslint-disable-next-line no-console
+      console.log("%cLM coverage: ".concat(lmCoverage, "/").concat(Math.min(ranked.length, 10), " candidates had real logit scores"), 'color: #888; font-style: italic;');
+      var tableData = ranked.slice(0, 10).map(function (r) {
+        var rawLogit = 'Not in Payload';
+        if (lmLogits) {
+          var val = lmLogits[r.word.toLowerCase()];
+          if (val !== undefined) {
+            rawLogit = Number(val.toFixed(5));
+          }
+        }
+        var original = scoringCandidates.find(function (sc) {
+          return sc.word === r.word;
+        });
+        var source = 'Unknown';
+        if (original) {
+          if (original.docFreq === 0 && original.tenantFreq === L3_BASELINE_FREQ) {
+            source = '🌍 L3 (Generic)';
+          } else if (original.sessionFreq > 0 && original.tenantFreq === 0) {
+            source = '👤 L1 (Session Only)';
+          } else {
+            source = '🏢 L2 (Domain)';
+          }
+        }
+        return {
+          Candidate: r.word,
+          Source: source,
+          'Final Score': Number(r.finalScore.toFixed(4)),
+          Semantics: Number(r.semanticScore.toFixed(4)),
+          Freq: Number(r.freqScore.toFixed(4)),
+          'LM Score': Number(r.lmScore.toFixed(4)),
+          'Raw Logit': rawLogit,
+          'Session Freq': (original === null || original === void 0 ? void 0 : original.sessionFreq) || 0
+        };
+      });
+
+      // eslint-disable-next-line no-console
+      console.table(tableData);
+    } else {
+      // eslint-disable-next-line no-console
+      console.log('No candidates found.');
+    }
+
+    // eslint-disable-next-line no-console
+    console.groupEnd();
+  }
+  return suggestion && suggestion.length > 0 ? suggestion : null;
+};
+
+// ─── Data Loading ────────────────────────────────────────────────────────────
+
+var DEFAULT_VECTORS_URL = '/data/word-vectors_10k.bin';
+export var loadVectorsAsync = /*#__PURE__*/function () {
+  var _ref4 = _asyncToGenerator( /*#__PURE__*/_regeneratorRuntime.mark(function _callee(options) {
+    var _options$vectorsUrl;
+    var url, res, buffer, float32, wordIndex, nWords, dim;
+    return _regeneratorRuntime.wrap(function _callee$(_context) {
+      while (1) switch (_context.prev = _context.next) {
+        case 0:
+          if (!(vectorStore || vectorsLoadStarted)) {
+            _context.next = 2;
+            break;
+          }
+          return _context.abrupt("return");
+        case 2:
+          vectorsLoadStarted = true;
+          url = (_options$vectorsUrl = options === null || options === void 0 ? void 0 : options.vectorsUrl) !== null && _options$vectorsUrl !== void 0 ? _options$vectorsUrl : DEFAULT_VECTORS_URL;
+          _context.prev = 4;
+          _context.next = 7;
+          return fetch(url);
+        case 7:
+          res = _context.sent;
+          if (res.ok) {
+            _context.next = 12;
+            break;
+          }
+          vectorsLoadStarted = false;
+          // eslint-disable-next-line no-console
+          console.warn("[text-predictor] Failed to load vectors: ".concat(res.status));
+          return _context.abrupt("return");
+        case 12:
+          _context.next = 14;
+          return res.arrayBuffer();
+        case 14:
+          buffer = _context.sent;
+          float32 = new Float32Array(buffer);
+          wordIndex = wordIndexData;
+          nWords = Object.keys(wordIndex).length;
+          dim = float32.length / nWords;
+          vectorStore = {
+            float32: float32,
+            wordIndex: wordIndex,
+            dim: dim
+          };
+          ensureDebugModeFromStorage();
+          if (debugMode) {
+            // eslint-disable-next-line no-console
+            console.log('[text-predictor] Vectors loaded:', {
+              wordCount: nWords,
+              dim: dim,
+              sizeBytes: float32.byteLength
+            });
+          }
+          _context.next = 28;
+          break;
+        case 24:
+          _context.prev = 24;
+          _context.t0 = _context["catch"](4);
+          vectorsLoadStarted = false;
+          // eslint-disable-next-line no-console
+          console.warn('[text-predictor] Failed to load vectors:', _context.t0);
+        case 28:
+        case "end":
+          return _context.stop();
+      }
+    }, _callee, null, [[4, 24]]);
+  }));
+  return function loadVectorsAsync(_x) {
+    return _ref4.apply(this, arguments);
+  };
+}();
+export var initVectors = function initVectors(store) {
+  vectorStore = store;
+};
+export var loadDefaultVocabulary = function loadDefaultVocabulary() {
+  // 1. Load the Atlassian Domain (L2)
+  var data = vocabularyData;
+  var terms = Object.entries(data.words).map(function (_ref5) {
+    var _ref6 = _slicedToArray(_ref5, 2),
+      word = _ref6[0],
+      stats = _ref6[1];
+    return {
+      word: word,
+      freq: stats.freq,
+      docFreq: stats.doc_freq,
+      authorFreq: stats.author_freq
+    };
+  });
+  initVocabulary({
+    terms: terms
+  });
+
+  // 2. Load General English (L3)
+  var l3Words = l3VocabularyData;
+  initL3Vocabulary(l3Words);
+};