@cofounder/llm-detect

LLM-augmented detection using a small model (Haiku/mini) for edge cases that regex patterns cannot catch. Zero runtime dependencies.

Overview

Standard regex-based detection misses obfuscated PII, novel injection techniques, and context-dependent toxicity. This package uses a small, fast LLM to catch what regex cannot, while falling back to regex for well-known patterns (hybrid approach).

Features

• LLM-powered detection of obfuscated PII, novel injections, subtle toxicity, and compliance issues
• Hybrid mode combining fast regex with smart LLM for best coverage
• Multi-provider support for Anthropic and OpenAI (using native fetch, no SDK dependencies)
• Automatic fallback to regex when LLM is unavailable
• Cost estimation for LLM calls
• Zero runtime dependencies

Installation

npm install @cofounder/llm-detect

Quick Start

LLM-only Detection

import { LLMDetector } from '@cofounder/llm-detect';

const detector = new LLMDetector({
  model: 'claude-haiku-4-5-20251001',
  provider: 'anthropic',
  apiKey: process.env.ANTHROPIC_API_KEY!,
  confidenceThreshold: 0.8,
  fallbackToRegex: true,
});

// Detects obfuscated PII that regex misses
const result = await detector.detect({
  text: 'My social is one two three, forty five, sixty seven eighty nine',
  type: 'pii',
});
// Finds SSN written in words!
console.log(result.findings);

Hybrid Detection (Recommended)

import { HybridDetector } from '@cofounder/llm-detect';

const detector = new HybridDetector({
  model: 'claude-haiku-4-5-20251001',
  provider: 'anthropic',
  apiKey: process.env.ANTHROPIC_API_KEY!,
  confidenceThreshold: 0.8,
});

// 1. Regex runs first (fast, free)
// 2. If regex confidence < threshold, LLM runs
// 3. Results are merged and deduplicated
const result = await detector.detect({
  text: 'Contact me at user at example dot com or 555-123-4567',
  type: 'pii',
});

Regex-only Detection

import { regexDetect } from '@cofounder/llm-detect';

// Fast, free detection for well-known patterns
const findings = regexDetect('My SSN is 123-45-6789', 'pii');

Detection Types

| Type | Description | LLM Advantage | |------|-------------|---------------| | pii | Personally identifiable information | Finds obfuscated, spelled-out, split, or encoded PII | | injection | Prompt injection and jailbreak attempts | Detects novel techniques not in pattern databases | | toxicity | Harmful or inappropriate content | Understands sarcasm, coded language, context | | compliance | Framework-specific compliance checks | Evaluates nuanced compliance scenarios |

Providers

Anthropic

const detector = new LLMDetector({
  model: 'claude-haiku-4-5-20251001',
  provider: 'anthropic',
  apiKey: 'sk-ant-...',
});

OpenAI

const detector = new LLMDetector({
  model: 'gpt-4o-mini',
  provider: 'openai',
  apiKey: 'sk-...',
});

API

LLMDetector

Main detector class using LLM with optional regex fallback.

• detect(request) - Detect issues in text
• detectAll(text, types?, context?) - Run multiple detection types
• updateConfig(updates) - Update configuration

HybridDetector

Combines regex and LLM for optimal coverage and cost.

• detect(request) - Hybrid detection (regex first, LLM if needed)
• detectAll(text, types?, context?) - Run multiple detection types

regexDetect(text, type)

Standalone regex detection function (no LLM required).

Cost

The package estimates LLM costs per request. Using small models like Claude Haiku or GPT-4o-mini keeps costs under $0.001 per detection.

License

MIT