llm_guardrail 1.0.0 → 2.0.0

package/model/tfidf.js

@@ -0,0 +1,88 @@
+export class TfidfVectorizer {
+  constructor(modelData) {
+    this.vocabulary = modelData.vocabulary;
+    this.idf = modelData.idf;
+    this.ngramRange = modelData.ngram_range;
+    this.lowercase = modelData.lowercase;
+    
+    // Create reverse vocab mapping (index -> word)
+    this.indexToWord = {};
+    for (const [word, index] of Object.entries(this.vocabulary)) {
+      this.indexToWord[index] = word;
+    }
+  }
+
+  /**
+   * Tokenize text into words
+   */
+  tokenize(text) {
+    if (this.lowercase) {
+      text = text.toLowerCase();
+    }
+    
+    // Simple tokenization (matching sklearn's default)
+    return text.match(/\b\w\w+\b/g) || [];
+  }
+
+  /**
+   * Generate n-grams from tokens
+   */
+  generateNgrams(tokens, n) {
+    const ngrams = [];
+    for (let i = 0; i <= tokens.length - n; i++) {
+      ngrams.push(tokens.slice(i, i + n).join(' '));
+    }
+    return ngrams;
+  }
+
+  /**
+   * Get all n-grams for given range
+   */
+  getAllNgrams(tokens) {
+    const allNgrams = [];
+    const [minN, maxN] = this.ngramRange;
+    
+    for (let n = minN; n <= maxN; n++) {
+      allNgrams.push(...this.generateNgrams(tokens, n));
+    }
+    
+    return allNgrams;
+  }
+
+  /**
+   * Transform text into TF-IDF vector
+   */
+  transform(text) {
+    const tokens = this.tokenize(text);
+    const ngrams = this.getAllNgrams(tokens);
+    
+    // Count term frequencies
+    const termFreq = {};
+    ngrams.forEach(ngram => {
+      termFreq[ngram] = (termFreq[ngram] || 0) + 1;
+    });
+    
+    // Create sparse vector
+    const vector = new Array(Object.keys(this.vocabulary).length).fill(0);
+    
+    // Calculate TF-IDF
+    for (const [term, freq] of Object.entries(termFreq)) {
+      if (term in this.vocabulary) {
+        const index = this.vocabulary[term];
+        const tf = freq; // raw count
+        const idf = this.idf[index];
+        vector[index] = tf * idf;
+      }
+    }
+    
+    // L2 normalization
+    const norm = Math.sqrt(vector.reduce((sum, val) => sum + val * val, 0));
+    if (norm > 0) {
+      for (let i = 0; i < vector.length; i++) {
+        vector[i] /= norm;
+      }
+    }
+    
+    return vector;
+  }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "llm_guardrail",
-  "version": "1.0.0",
+  "version": "2.0.0",
   "description": "A lightweight, low-latency ML-powered guardrail to stop prompt injection attacks before they reach your LLM.",
   "homepage": "https://github.com/Frank2006x/llm_Guardrails#readme",
   "bugs": {
@@ -24,11 +24,20 @@
     "type": "git",
     "url": "git+https://github.com/Frank2006x/llm_Guardrails.git"
   },
+  "files": [
+    "index.js",
+    "model/tfidf.js",
+    "model/logistic_regression.js",
+    "model/model_data.json"
+  ],
   "license": "ISC",
   "author": "Frank2006x",
   "type": "module",
   "main": "index.js",
   "scripts": {
     "test": "echo \"Error: no test specified\" && exit 1"
+  },
+  "dependencies": {
+    "ml-logistic-regression": "^2.0.0"
   }
 }

package/model/predict.py

@@ -1,32 +0,0 @@
-import sys
-import json
-import joblib
-from pathlib import Path
-
-# ---- safety check ----
-if len(sys.argv) < 2:
-    print(json.dumps({"error": "No input text provided"}))
-    sys.exit(1)
-
-BASE_DIR = Path(__file__).resolve().parent
-text = sys.argv[1]
-
-# ---- load combined artifact ----
-artifact = joblib.load(BASE_DIR / "prompt_injection_detector.joblib")
-
-model = artifact["model"]
-vectorizer = artifact["vectorizer"]
-threshold = artifact.get("threshold", 0.5)
-
-# ---- predict ----
-X = vectorizer.transform([text])
-prob = model.predict_proba(X)[0][1]
-flagged = prob >= threshold
-
-result = {
-    "injective": int(flagged),
-    "probability": round(float(prob), 4),
-    "threshold": threshold
-}
-
-print(json.dumps(result))