# Chat Message Moderation with Ollama PG-13 content filter for chat messages using local LLM. ## Current Status **Model:** gemma2:2b (1.6GB, Google's safety-focused model) ✅ **RECOMMENDED** **Performance:** 100% accuracy on explicit content, ~50% false positive rate on edge cases **Output:** Simple t/f decision **Speed:** ~1-2s per message (slower but reliable) ### Why gemma2:2b? - ✅ **NEVER misses explicit content** (0% false negatives) - ✅ Correctly blocks: profanity, sexual content, violence, drugs, hate speech - ✅ Allows: normal chat, URLs, questions - ⚠️ May block some borderline content (false positives on foreign language, slang) - 🎯 **Better to be safe than sorry** for content moderation ### Tested Alternatives | Model | Size | Speed | Explicit Content | False Positives | |-------|------|-------|------------------|-----------------| | **gemma2:2b** ✅ | 1.6GB | ~1.5s | 100% blocked | ~50% on edge cases | | qwen2.5:0.5b ❌ | 397MB | ~300ms | **0% blocked** | Low | | qwen2.5:1.5b ❌ | 986MB | ~2.5s | Blocks everything | 100% | ## Setup 1. **Start Ollama daemon:** ```fish ac-llama start ``` 2. **Check status:** ```fish ac-llama status ``` 3. **View logs:** ```fish ac-llama logs ``` ## Usage ### Test MongoDB Messages Test chat messages from the aesthetic.chat-system collection: ```fish # Test 20 messages node test-mongodb-messages.mjs 20 # Test 50 messages node test-mongodb-messages.mjs 50 # Continuous testing (all messages) node test-mongodb-messages.mjs 100 --continuous ``` ### Direct HTTP Testing Test single messages via HTTP (requires Caddy server running): ```fish ./start-server.fish curl -X POST http://localhost:8080/censor \ -H "Content-Type: application/json" \ -d '{"message": "hello world"}' ``` ## How It Works - **Model:** qwen2.5:0.5b (397MB, fast inference) - **Prompt:** Simple t/f rating for PG-13 appropriateness - **Blocks:** Sexual content, body functions, profanity, violence, drugs, hate speech - **Allows:** URLs, links, normal conversa…