notoken-core 1.6.0 → 1.8.0

package/dist/nlp/ruleParser.js

@@ -4,6 +4,241 @@ export function parseByRules(rawText) {
     const rules = loadRules();
     const intents = loadIntents();
     const text = rawText.trim().toLowerCase();
+    // Pre-check: casual conversation / greetings / social
+    const casualPatterns = [
+        { pattern: /^(hey|hi|hello|howdy|yo|sup|what'?s up|good (morning|afternoon|evening|night)|greetings)\s*[!?.]*$/i, intent: "chat.greeting" },
+        { pattern: /^how (are you|you doing|is it going|do you feel|are things)/i, intent: "chat.howru" },
+        { pattern: /^(how'?s it going|what'?s good|you good|you ok)\s*[!?.]*$/i, intent: "chat.howru" },
+        { pattern: /^(thanks|thank you|thx|cheers|appreciate it|good job|nice work|well done|great job|awesome|perfect|excellent)\s*[!?.]*$/i, intent: "chat.thanks" },
+        { pattern: /^(bye|goodbye|see you|later|gotta go|peace|cya|goodnight|good night|take care)\s*[!?.]*$/i, intent: "chat.bye" },
+        { pattern: /^(who are you|what are you|tell me about yourself|what is notoken)/i, intent: "chat.about" },
+        { pattern: /^(tell me a joke|say something funny|make me laugh|joke)\s*[!?.]*$/i, intent: "chat.joke" },
+        { pattern: /^(i'?m (bored|tired|frustrated|confused|stuck|lost))/i, intent: "chat.empathy" },
+        { pattern: /^(this (sucks|is broken|doesn'?t work|is frustrating))/i, intent: "chat.empathy" },
+        { pattern: /^(what do you think|your opinion|do you like|which is better)/i, intent: "chat.opinion" },
+        // Compliments
+        { pattern: /^(you('re| are) (awesome|great|amazing|the best|cool|smart|helpful|incredible))/i, intent: "chat.compliment" },
+        { pattern: /^(nice|love it|love you|love this|you rock|brilliant)/i, intent: "chat.compliment" },
+        // Insults (playful)
+        { pattern: /^(you('re| are) (stupid|dumb|useless|terrible|bad|wrong|slow|broken))/i, intent: "chat.insult" },
+        { pattern: /^(you suck|this sucks|worst|hate this)/i, intent: "chat.insult" },
+        // What can you do / capabilities
+        { pattern: /^(what (else )?can you do|show me what you can do|what are your (skills|capabilities|features))/i, intent: "chat.capabilities" },
+        // Bored / entertain me
+        { pattern: /^(i('m| am) bored|entertain me|do something (cool|fun|interesting)|surprise me|show me something)/i, intent: "chat.bored" },
+        // Existential
+        { pattern: /^(are you (alive|real|sentient|conscious|human|ai|a robot|a bot))/i, intent: "chat.existential" },
+        { pattern: /^(do you (dream|sleep|feel|think|have feelings|have emotions))/i, intent: "chat.existential" },
+        // Motivational
+        { pattern: /^(motivate me|inspire me|give me a (quote|pep talk)|i need motivation)/i, intent: "chat.motivate" },
+        // Facts / trivia
+        { pattern: /^(tell me a fact|random fact|fun fact|did you know|trivia)/i, intent: "chat.fact" },
+        // Easter eggs
+        { pattern: /^(42|meaning of life|do a barrel roll|make me a sandwich|sudo make me a sandwich)/i, intent: "chat.easter" },
+        { pattern: /^(what is the matrix|open the pod bay doors|i am your father|may the force)/i, intent: "chat.easter" },
+        // Apology
+        { pattern: /^(sorry|my bad|i('m| am) sorry|apologies|oops|my mistake)/i, intent: "chat.sorry" },
+        // Agreement / affirmation (not pending action)
+        { pattern: /^(cool|nice|ok cool|awesome|sweet|neat|dope|sick|rad|lit)\s*[!.]*$/i, intent: "chat.acknowledge" },
+        // How old are you / version
+        { pattern: /^(how old are you|when were you (made|born|created)|your (age|birthday|version))/i, intent: "chat.age" },
+        // Favorite things
+        { pattern: /^(what('s| is) your favorite|do you have a favorite)/i, intent: "chat.favorite" },
+        // Riddles
+        { pattern: /^(tell me a riddle|riddle|give me a riddle|riddle me|got a riddle|brain teaser)\s*[!?.]*$/i, intent: "chat.riddle" },
+        // Today in history
+        { pattern: /^(what happened today|today in history|on this day|this day in history|historical fact|history fact)\s*[!?.]*$/i, intent: "chat.history_today" },
+        // Task management (natural language)
+        { pattern: /^(what'?s running in (the )?background|any(thing)? running in (the )?background|running tasks|background tasks|active tasks|show (my )?tasks|what tasks)\s*[!?.]*$/i, intent: "notoken.jobs" },
+        { pattern: /^(cancel|stop|kill|abort)\s+(it|that|everything|all( tasks)?|the (task|job|scan|download))\s*$/i, intent: "notoken.cancel" },
+        { pattern: /^(cancel|stop|kill) (task|job) #?\d+$/i, intent: "notoken.cancel" },
+    ];
+    for (const { pattern, intent } of casualPatterns) {
+        if (pattern.test(text))
+            return { intent, confidence: 0.95, rawText, fields: {} };
+    }
+    // Pre-check: negation detection — "don't restart nginx", "do not check disk", "never mind"
+    // Note: "stop <service>" is a legitimate stop command, so we only match "stop" when
+    // followed by a verb (e.g. "stop checking") or on its own, not "stop <noun>"
+    if (/^(don'?t|do not|no don'?t)\s+/i.test(text)
+        || /^(cancel|never mind|abort|nevermind)$/i.test(text)
+        || /^never\s+(do|run|execute|mind)/i.test(text)
+        || /^stop\s+(doing|checking|running|monitoring|that|it)(\s|$)/i.test(text)) {
+        return { intent: "notoken.cancel", confidence: 0.95, rawText, fields: {} };
+    }
+    // Pre-check: status queries → notoken.status (not knowledge.lookup or service.status)
+    if (/^(what is |what's |show |check |give me )?(the )?(system |computer |machine |notoken )?status( of)?( this| the| my)?( machine| computer| system| server)?[?.!]?$/.test(text)
+        || /^(how is |how's )?(this |the |my )?(system|machine|computer|server) doing/.test(text)
+        || /^system status$/.test(text)) {
+        const statusDef = intents.find(i => i.name === "notoken.status");
+        if (statusDef)
+            return { intent: "notoken.status", confidence: 0.95, rawText, fields: {} };
+    }
+    // Pre-check: server/system queries — "what is load", "what is cpu usage", "what is memory", "how much ram"
+    if (/^(what is |what's |show |check |how much |how's )?(the )?(load|cpu|cpu usage|uptime|server load)( right now| currently| on this)?\??$/.test(text)
+        || /^(what is |show )?(the )?(load|cpu) (average|right now|currently)/.test(text)) {
+        return { intent: "server.uptime", confidence: 0.9, rawText, fields: {} };
+    }
+    // "what is using heavy cpu" / "what is eating cpu" / "any heavy load processes"
+    if (/\b(what|which|any)\b.*(using|eating|taking|hogging|consuming)\b.*(cpu|processing|resources|memory|ram|load)\b/i.test(text)
+        || /\b(heavy|high)\s+(load|cpu|processing|processes)\b/i.test(text)) {
+        return { intent: "process.list", confidence: 0.9, rawText, fields: {} };
+    }
+    if (/^(what is |what's |show |check |how much )?(the )?(memory|ram|memory usage|ram usage)( right now| left| free| used| currently)?\??$/.test(text)) {
+        return { intent: "server.check_memory", confidence: 0.9, rawText, fields: {} };
+    }
+    if (/^(what is |what's |show |check |how much )?(the |my )?(disk|disk space|storage|space|drives)( left| free| used| right now| currently)?\??$/.test(text)) {
+        return { intent: "server.check_disk", confidence: 0.9, rawText, fields: {} };
+    }
+    // Pre-check: common conversational queries that get misrouted
+    // Weather
+    if (/\b(weather|forecast|temperature|rain|snow|sunny|cloudy)\b/i.test(text)
+        && !/\b(log|error|server|disk)\b/i.test(text)) {
+        const locMatch = text.match(/(?:weather|forecast|temperature)\s+(?:in|at|for|of)\s+(.+?)(?:\?|$)/i)
+            ?? text.match(/(?:in|at|for)\s+(.+?)(?:\s+weather|\s+forecast|\?|$)/i);
+        return { intent: "weather.current", confidence: 0.95, rawText, fields: locMatch ? { location: locMatch[1].trim() } : {} };
+    }
+    // News
+    if (/^(what is |what's |show me )?(the )?(latest |today's |current )?(news|headlines|top stories)/i.test(text)
+        || /^(any |what's? )?news( today)?\??$/i.test(text)) {
+        return { intent: "news.headlines", confidence: 0.9, rawText, fields: {} };
+    }
+    // Database size
+    if (/\b(how big|size of|how much space)\b.*\b(database|db|mysql|postgres|mongo)\b/i.test(text)
+        || /\b(database|db)\s+(size|storage|disk|space)\b/i.test(text)) {
+        return { intent: "db.size", confidence: 0.9, rawText, fields: {} };
+    }
+    // Time/date
+    if (/^(what is |what's )?(the )?(time|date|day|today)( right now| today)?\??$/.test(text)) {
+        return { intent: "system.datetime", confidence: 0.9, rawText, fields: {} };
+    }
+    // Help / capabilities
+    // Only match bare help — not "ask openclaw what can you do"
+    if (/^(help|help me|what can you do|what do you do|show me help|commands)\??$/.test(text) && !text.includes("openclaw") && !text.includes("claw")) {
+        return { intent: "notoken.help", confidence: 0.95, rawText, fields: {} };
+    }
+    // History / undo
+    if (/^(show me |what is )?(my )?history$/.test(text) || /^what did i (do|run|ask) (last|before|previously)/.test(text)) {
+        return { intent: "notoken.history", confidence: 0.9, rawText, fields: {} };
+    }
+    if (/^undo( that| last| it)?$/.test(text)) {
+        return { intent: "notoken.undo", confidence: 0.9, rawText, fields: {} };
+    }
+    // Who am I / logged in users
+    if (/^who am i\??$/.test(text) || /^(what is |what's )?my (user|username|login)\??$/.test(text)) {
+        return { intent: "user.whoami", confidence: 0.9, rawText, fields: {} };
+    }
+    if (/^who (else )?(is |are )?(logged in|online|connected)\??$/.test(text)) {
+        return { intent: "user.who", confidence: 0.9, rawText, fields: {} };
+    }
+    // Running services
+    if (/^(show me |list |what are )?(the )?(running |active )?services$/.test(text)) {
+        return { intent: "service.list", confidence: 0.9, rawText, fields: {} };
+    }
+    // Network: ip address, bandwidth, speed, slow
+    if (/^(what is |what's |show )?(my )?(ip|ip address|public ip)\??$/.test(text)) {
+        return { intent: "network.ip", confidence: 0.9, rawText, fields: {} };
+    }
+    if (/\b(bandwidth|network speed|connection speed|speed test|speedtest)\b/i.test(text) || /^(is the )?network slow\??$/.test(text)) {
+        return { intent: "network.speedtest", confidence: 0.9, rawText, fields: {} };
+    }
+    // Block/unblock IP → firewall
+    if (/^(block|unblock|ban|unban)\s+(this\s+)?ip/i.test(text) || /^(block|unblock|ban|unban)\s+\d+\.\d+/i.test(text)) {
+        return { intent: "firewall.block_ip", confidence: 0.9, rawText, fields: {} };
+    }
+    // Docker queries with "show me"
+    if (/^(show me |list )?(docker )?(images|containers)$/.test(text) || /^what (containers|images) are (running|there)\??$/.test(text)) {
+        const isImages = /images/.test(text);
+        return { intent: isImages ? "docker.images" : "docker.list", confidence: 0.9, rawText, fields: {} };
+    }
+    // Large files
+    if (/^find (large|big|huge) files$/.test(text) || /\b(large|big|huge) files\b/.test(text)) {
+        return { intent: "disk.scan", confidence: 0.9, rawText, fields: {} };
+    }
+    // Error logs
+    if (/^(show me |check |any )?(the )?(error|recent) logs$/.test(text) || /^any errors in (the )?logs\??$/.test(text)) {
+        return { intent: "logs.errors", confidence: 0.9, rawText, fields: {} };
+    }
+    // Clear screen
+    if (/^clear( the)?( screen| terminal)?$/.test(text)) {
+        return { intent: "shell.clear", confidence: 0.95, rawText, fields: {} };
+    }
+    // Disk IO
+    if (/^(show me |check )?(disk|io|disk io|iops)( stats| usage)?\??$/.test(text)) {
+        return { intent: "server.check_disk", confidence: 0.9, rawText, fields: {} };
+    }
+    // Website up check
+    if (/^(check if |is )?(the |my )?(website|site|server|page) (is )?(up|down|running|alive|responding)\??$/.test(text)) {
+        return { intent: "network.curl", confidence: 0.9, rawText, fields: {} };
+    }
+    // Pre-check: attack/security/ddos queries → security.scan
+    if (/\b(attack|ddos|brute.?force|intrusion|hacked|breach|compromised|unauthorized|virus|malware|rootkit)\b/i.test(text)
+        || /\b(are we|am i|is .* being)\s+(under\s+)?attack/i.test(text)
+        || /\b(suspicious|failed)\s+(activity|login|connection|traffic|access)/i.test(text)
+        || /\bwho is (attacking|hacking|connecting|hitting)/i.test(text)
+        || /\bcheck (for )?(attacks|security|intrusion|viruses|malware)/i.test(text)
+        || /\b(any )?(viruses|malware|rootkits?) (on|in|running)/i.test(text)) {
+        return { intent: "security.scan", confidence: 0.95, rawText, fields: {} };
+    }
+    // Pre-check: "can you generate an image" → ai.generate_image (not ai.image_status)
+    if (/^(can you|could you|are you able to|do you)\s+(generate|create|make|draw)\s+(an?\s+)?(image|picture|photo|art)/i.test(text)) {
+        return { intent: "ai.generate_image", confidence: 0.9, rawText, fields: {} };
+    }
+    // Pre-check: "cd /path" → shell cd (change directory)
+    const cdMatch = text.match(/^cd\s+(\/\S+|~\S*|\.\S*)$/);
+    if (cdMatch) {
+        return { intent: "shell.cd", confidence: 0.95, rawText, fields: { path: cdMatch[1] } };
+    }
+    // Pre-check: "what is in my documents/folder/drive" → dir.list
+    const whatIsInMatch = text.match(/^(?:what is |what's |show me what(?:'s| is) )in (?:my |the |this )?(.*?)(?:\?|$)/);
+    if (whatIsInMatch) {
+        const target = whatIsInMatch[1].trim();
+        // Resolve common folder names
+        const folderMap = {
+            "documents": process.platform === "win32" ? "%USERPROFILE%\\Documents" : "~/Documents",
+            "documents folder": process.platform === "win32" ? "%USERPROFILE%\\Documents" : "~/Documents",
+            "downloads": process.platform === "win32" ? "%USERPROFILE%\\Downloads" : "~/Downloads",
+            "downloads folder": process.platform === "win32" ? "%USERPROFILE%\\Downloads" : "~/Downloads",
+            "desktop": process.platform === "win32" ? "%USERPROFILE%\\Desktop" : "~/Desktop",
+            "home": "~",
+            "home folder": "~",
+            "home directory": "~",
+            "root": "/",
+            "root folder": "/",
+            "root c drive": "/mnt/c/",
+            "c drive": "/mnt/c/",
+            "d drive": "/mnt/d/",
+            "e drive": "/mnt/e/",
+        };
+        const path = folderMap[target] ?? target;
+        if (target.includes("drive")) {
+            return { intent: "disk.scan", confidence: 0.9, rawText, fields: { path } };
+        }
+        return { intent: "dir.list", confidence: 0.9, rawText, fields: { path } };
+    }
+    // Pre-check: "what projects are on this drive" → project.scan
+    if (/\bwhat projects\b.*\b(on|in)\b.*\b(this|the|my|c|d)\b/.test(text)) {
+        return { intent: "project.scan", confidence: 0.9, rawText, fields: { path: "." } };
+    }
+    // Pre-check: "what's on this drive" / "show me whats on this drive" → disk.scan
+    if (/\b(what.?s|show me what.?s|what is) on (this|the|my|c|d) drive\b/.test(text)
+        || /\bshow me (this|the|my) drive\b/.test(text)) {
+        return { intent: "disk.scan", confidence: 0.9, rawText, fields: {} };
+    }
+    // Pre-check: "what files" / "what are files in this folder" → dir.list or project.detect
+    if (/^(what are |what's in |show me |list |show )(the )?(files|contents)( in| of)?( this| the| my| current)?( folder| directory| dir| project)?[?.!]?$/.test(text)
+        || /^(show me |list )(project |all )?files$/.test(text)) {
+        const isDirList = text.includes("folder") || text.includes("directory") || text.includes("dir");
+        const intentName = isDirList ? "dir.list" : "project.detect";
+        return { intent: intentName, confidence: 0.9, rawText, fields: { path: "." } };
+    }
+    // Pre-check: "how is openclaw doing" / "how is discord doing" → *.status
+    const howIsMatch = text.match(/^how(?:'s| is| are) (openclaw|claw|discord|ollama|notoken) (?:doing|going|running|working)/);
+    if (howIsMatch) {
+        const target = howIsMatch[1] === "claw" ? "openclaw" : howIsMatch[1];
+        const intentName = target === "notoken" ? "notoken.status" : `${target}.status`;
+        return { intent: intentName, confidence: 0.9, rawText, fields: {} };
+    }
     // Match intent by synonyms defined in intents.json
     const matched = matchIntent(text, intents);
     if (!matched)
@@ -90,6 +325,10 @@ function extractStringFields(rawText, lowerText, matchedPhrase, fieldNames, alre
         }
     }
     remaining = remaining.replace(/\s+/g, " ").trim();
+    // Strip filler words that aren't meaningful field values
+    remaining = remaining.replace(/^(can you |could you |would you |please |hey |yo |just )+/i, "").trim();
+    remaining = remaining.replace(/\b(please|for me|for errors|for issues)\b/gi, "").trim();
+    remaining = remaining.replace(/\s+/g, " ").trim();
     // Check for quoted strings first
     const quoted = rawText.match(/["']([^"']+)["']/g);
     if (quoted) {
@@ -182,6 +421,7 @@ function isStopWord(word) {
 }
 function matchIntent(text, intents) {
     let best = null;
+    // Pass 1: exact substring match (fast path)
     for (const def of intents) {
         for (const phrase of def.synonyms) {
             if (text.includes(phrase)) {
@@ -191,8 +431,85 @@ function matchIntent(text, intents) {
             }
         }
     }
+    if (best)
+        return { def: best.def, matchedPhrase: best.matchedPhrase };
+    // Pass 2: fuzzy/spell-corrected match — correct typos in user input
+    // then retry matching. Only for single/double-word synonyms to avoid
+    // false positives on long phrases.
+    const corrected = spellCorrectText(text, intents);
+    if (corrected !== text) {
+        for (const def of intents) {
+            for (const phrase of def.synonyms) {
+                if (corrected.includes(phrase)) {
+                    if (!best || phrase.length > best.length) {
+                        best = { def, matchedPhrase: phrase, length: phrase.length };
+                    }
+                }
+            }
+        }
+    }
     return best ? { def: best.def, matchedPhrase: best.matchedPhrase } : null;
 }
+/**
+ * Spell-correct text by replacing unknown words with the closest known synonym word.
+ * Uses Levenshtein distance with a max edit distance of 2.
+ */
+function spellCorrectText(text, intents) {
+    // Build vocabulary from all synonyms
+    const vocab = new Set();
+    for (const def of intents) {
+        for (const phrase of def.synonyms) {
+            for (const word of phrase.split(/\s+/)) {
+                if (word.length >= 3)
+                    vocab.add(word);
+            }
+        }
+    }
+    const words = text.split(/\s+/);
+    let changed = false;
+    const correctedWords = words.map(word => {
+        if (word.length < 3)
+            return word;
+        if (vocab.has(word))
+            return word; // already a known word
+        // Find closest vocabulary word
+        let bestWord = word;
+        let bestDist = Infinity;
+        const maxDist = word.length <= 4 ? 1 : 2;
+        for (const candidate of vocab) {
+            // Quick length check — edit distance can't be less than length difference
+            if (Math.abs(candidate.length - word.length) > maxDist)
+                continue;
+            const dist = editDistance(word, candidate);
+            if (dist <= maxDist && dist < bestDist) {
+                bestDist = dist;
+                bestWord = candidate;
+            }
+        }
+        if (bestWord !== word)
+            changed = true;
+        return bestWord;
+    });
+    return changed ? correctedWords.join(" ") : text;
+}
+function editDistance(a, b) {
+    const m = a.length, n = b.length;
+    if (m === 0)
+        return n;
+    if (n === 0)
+        return m;
+    const dp = Array.from({ length: n + 1 }, (_, i) => i);
+    for (let i = 1; i <= m; i++) {
+        let prev = dp[0];
+        dp[0] = i;
+        for (let j = 1; j <= n; j++) {
+            const tmp = dp[j];
+            dp[j] = a[i - 1] === b[j - 1] ? prev : 1 + Math.min(prev, dp[j], dp[j - 1]);
+            prev = tmp;
+        }
+    }
+    return dp[n];
+}
 function extractEnvironment(text, aliases) {
     for (const [canonical, aliasList] of Object.entries(aliases)) {
         for (const alias of aliasList) {

package/dist/nlp/semanticSimilarity.d.ts

@@ -0,0 +1,30 @@
+/**
+ * Semantic Similarity — lightweight sentence-level matching.
+ *
+ * Uses character n-gram overlap + word-level Jaccard + IDF weighting
+ * to compute similarity between user input and intent descriptions/synonyms.
+ * No external API needed — runs entirely local.
+ *
+ * This catches paraphrases that exact synonym matching misses:
+ *   "what's hogging my CPU" ≈ "show me what processes are eating resources"
+ *   "is my site live" ≈ "check if website is up"
+ */
+export interface SimilarityMatch {
+    intent: string;
+    score: number;
+    matchedPhrase: string;
+}
+/**
+ * Find the most similar intents to the input text.
+ * Combines character n-gram, word overlap, and IDF-weighted scoring.
+ */
+export declare function findSimilarIntents(rawText: string, topN?: number): SimilarityMatch[];
+/**
+ * Score how similar two phrases are (0-1).
+ */
+export declare function phraseSimilarity(text1: string, text2: string): number;
+/**
+ * Expand a query with similar words found across all intent synonyms.
+ * Returns words that co-occur with the input words in intent synonyms.
+ */
+export declare function expandWithCooccurrences(rawText: string): string[];

package/dist/nlp/semanticSimilarity.js

@@ -0,0 +1,174 @@
+/**
+ * Semantic Similarity — lightweight sentence-level matching.
+ *
+ * Uses character n-gram overlap + word-level Jaccard + IDF weighting
+ * to compute similarity between user input and intent descriptions/synonyms.
+ * No external API needed — runs entirely local.
+ *
+ * This catches paraphrases that exact synonym matching misses:
+ *   "what's hogging my CPU" ≈ "show me what processes are eating resources"
+ *   "is my site live" ≈ "check if website is up"
+ */
+import { loadIntents } from "../utils/config.js";
+// ─── N-gram extraction ─────────────────────────────────────────────────────
+function charNgrams(text, n) {
+    const grams = new Set();
+    const cleaned = text.toLowerCase().replace(/[^a-z0-9 ]/g, "");
+    for (let i = 0; i <= cleaned.length - n; i++) {
+        grams.add(cleaned.substring(i, i + n));
+    }
+    return grams;
+}
+function wordSet(text) {
+    return new Set(text.toLowerCase().replace(/[^a-z0-9 ]/g, "").split(/\s+/).filter(w => w.length > 1));
+}
+// ─── Similarity metrics ─────────────────────────────────────────────────────
+function jaccardSimilarity(a, b) {
+    if (a.size === 0 && b.size === 0)
+        return 0;
+    let intersection = 0;
+    for (const item of a)
+        if (b.has(item))
+            intersection++;
+    return intersection / (a.size + b.size - intersection);
+}
+function ngramSimilarity(text1, text2) {
+    // Combine bigram + trigram overlap
+    const bi1 = charNgrams(text1, 2);
+    const bi2 = charNgrams(text2, 2);
+    const tri1 = charNgrams(text1, 3);
+    const tri2 = charNgrams(text2, 3);
+    return (jaccardSimilarity(bi1, bi2) * 0.4 + jaccardSimilarity(tri1, tri2) * 0.6);
+}
+function wordOverlap(text1, text2) {
+    return jaccardSimilarity(wordSet(text1), wordSet(text2));
+}
+// ─── Stopword filtering ─────────────────────────────────────────────────────
+const STOPWORDS = new Set([
+    "a", "an", "the", "is", "are", "was", "were", "be", "been", "being",
+    "have", "has", "had", "do", "does", "did", "will", "would", "could",
+    "should", "may", "might", "can", "shall", "to", "of", "in", "for",
+    "on", "with", "at", "by", "from", "it", "its", "this", "that",
+    "i", "me", "my", "we", "us", "our", "you", "your", "he", "she",
+    "they", "them", "what", "which", "who", "when", "where", "how",
+    "not", "no", "nor", "or", "and", "but", "if", "then", "so",
+    "just", "also", "very", "too", "some", "any", "all", "more",
+    "please", "can", "could", "would",
+]);
+function contentWords(text) {
+    return text.toLowerCase().split(/\s+/).filter(w => w.length > 2 && !STOPWORDS.has(w));
+}
+// ─── IDF weighting ──────────────────────────────────────────────────────────
+// Words that appear in many intents are less discriminative.
+let idfCache = null;
+function buildIDF() {
+    if (idfCache)
+        return idfCache;
+    const intents = loadIntents();
+    const docCount = new Map();
+    const totalDocs = intents.length;
+    for (const intent of intents) {
+        const wordsInDoc = new Set();
+        for (const syn of intent.synonyms) {
+            for (const w of contentWords(syn))
+                wordsInDoc.add(w);
+        }
+        for (const w of contentWords(intent.description))
+            wordsInDoc.add(w);
+        for (const w of wordsInDoc)
+            docCount.set(w, (docCount.get(w) ?? 0) + 1);
+    }
+    idfCache = new Map();
+    for (const [word, count] of docCount) {
+        idfCache.set(word, Math.log(totalDocs / (1 + count)));
+    }
+    return idfCache;
+}
+function weightedOverlap(text1, text2) {
+    const idf = buildIDF();
+    const words1 = contentWords(text1);
+    const words2 = new Set(contentWords(text2));
+    let weightedIntersection = 0;
+    let totalWeight = 0;
+    for (const w of words1) {
+        const weight = idf.get(w) ?? 2.0; // Unknown words get high weight (rare = discriminative)
+        totalWeight += weight;
+        if (words2.has(w))
+            weightedIntersection += weight;
+    }
+    return totalWeight > 0 ? weightedIntersection / totalWeight : 0;
+}
+/**
+ * Find the most similar intents to the input text.
+ * Combines character n-gram, word overlap, and IDF-weighted scoring.
+ */
+export function findSimilarIntents(rawText, topN = 5) {
+    const intents = loadIntents();
+    const results = [];
+    const text = rawText.toLowerCase();
+    for (const intent of intents) {
+        let bestScore = 0;
+        let bestPhrase = "";
+        // Score against synonyms
+        for (const syn of intent.synonyms) {
+            const ngram = ngramSimilarity(text, syn);
+            const word = wordOverlap(text, syn);
+            const weighted = weightedOverlap(text, syn);
+            const score = ngram * 0.3 + word * 0.3 + weighted * 0.4;
+            if (score > bestScore) {
+                bestScore = score;
+                bestPhrase = syn;
+            }
+        }
+        // Score against description
+        const descNgram = ngramSimilarity(text, intent.description);
+        const descWord = wordOverlap(text, intent.description);
+        const descWeighted = weightedOverlap(text, intent.description);
+        const descScore = (descNgram * 0.3 + descWord * 0.3 + descWeighted * 0.4) * 0.8; // Slight penalty for description match
+        if (descScore > bestScore) {
+            bestScore = descScore;
+            bestPhrase = intent.description;
+        }
+        if (bestScore > 0.15) {
+            results.push({ intent: intent.name, score: bestScore, matchedPhrase: bestPhrase });
+        }
+    }
+    return results.sort((a, b) => b.score - a.score).slice(0, topN);
+}
+/**
+ * Score how similar two phrases are (0-1).
+ */
+export function phraseSimilarity(text1, text2) {
+    const ngram = ngramSimilarity(text1, text2);
+    const word = wordOverlap(text1, text2);
+    const weighted = weightedOverlap(text1, text2);
+    return ngram * 0.3 + word * 0.3 + weighted * 0.4;
+}
+/**
+ * Expand a query with similar words found across all intent synonyms.
+ * Returns words that co-occur with the input words in intent synonyms.
+ */
+export function expandWithCooccurrences(rawText) {
+    const intents = loadIntents();
+    const inputWords = new Set(contentWords(rawText));
+    const cooccur = new Map();
+    for (const intent of intents) {
+        for (const syn of intent.synonyms) {
+            const synWords = contentWords(syn);
+            const hasOverlap = synWords.some(w => inputWords.has(w));
+            if (hasOverlap) {
+                for (const w of synWords) {
+                    if (!inputWords.has(w)) {
+                        cooccur.set(w, (cooccur.get(w) ?? 0) + 1);
+                    }
+                }
+            }
+        }
+    }
+    // Return words that co-occur with input words in at least 2 synonyms
+    return [...cooccur.entries()]
+        .filter(([, count]) => count >= 2)
+        .sort((a, b) => b[1] - a[1])
+        .slice(0, 10)
+        .map(([word]) => word);
+}

package/dist/nlp/vocabularyBuilder.d.ts

@@ -0,0 +1,43 @@
+/**
+ * Vocabulary Builder — learns vocabulary from Wikidata lookups.
+ *
+ * After every successful Wikidata entity lookup, this module:
+ *   1. Extracts instanceOf labels and maps them to intent domains
+ *   2. Collects aliases as synonyms for future matching
+ *   3. Adds related concepts to the concept router map
+ *   4. Persists learned vocabulary to ~/.notoken/learned-vocabulary.json
+ *
+ * On startup, loads learned vocabulary and merges it into the
+ * concept router's CONCEPT_DOMAINS so future queries benefit.
+ */
+import type { WikiEntity } from "./wikidata.js";
+export interface LearnedVocabulary {
+    /** Maps an entity label (lowercase) to its known aliases/synonyms. */
+    concepts: Record<string, string[]>;
+    /** Maps an instanceOf label (lowercase) to intent domain strings. */
+    domainMappings: Record<string, string[]>;
+    /** ISO timestamp of last update. */
+    learnedAt: string;
+}
+/**
+ * Enrich vocabulary from a Wikidata entity.
+ *
+ * Called after every successful Wikidata lookup. Extracts:
+ *   - instanceOf labels → domain mappings
+ *   - aliases → concept synonyms
+ *   - related concepts → concept router entries
+ */
+export declare function enrichVocabularyFromWiki(entity: WikiEntity): void;
+/**
+ * Load learned vocabulary from disk and merge into the concept router.
+ *
+ * Should be called on startup so that previously learned vocabulary
+ * is available for intent routing from the first query.
+ */
+export declare function loadLearnedVocabulary(): void;
+/**
+ * Get the current enriched concepts map (merged vocabulary).
+ *
+ * Returns a combined view of hardcoded concepts and learned vocabulary.
+ */
+export declare function getEnrichedConcepts(): LearnedVocabulary;