How Does StillMe Actually Learn? Technical Deep Dive

[anhmtk]

❓ How Does StillMe Actually Learn? Technical Deep Dive

Question

I see that StillMe has "Self-Evolving," but I don't understand how it learns. Could you explain the technical process?

Answer

1. RSS Fetching (Input Layer)

• Scheduler runs automatically every 4 hours
• Fetches content from RSS feeds (ArXiv, TechCrunch, HN, etc.)
• Parses and extracts text content
• Filters duplicate and low-quality content

2. Ethical Filtering (Safety Layer)

• Red-Team Agent scans content for safety
• EthicsGuard checks compliance with ethical rules
• Blocks harmful, toxic, or biased content
• Complete audit trail of every decision

3. Vector DB Storage (Memory Layer)

• Text content → Embeddings (sentence-transformers model)
• Stores embeddings into ChromaDB
• Indexes for semantic search
• Metadata tracking (source, timestamp, trust score)

4. RAG Retrieval (Recall Layer)

• User query → Generates embedding
• Semantic search within Vector DB
• Retrieves top-k relevant contexts
• Combines with conversation history

5. Learning Cycle (Evolution Layer)

• Session count → Stage progression (Infant → Adult)
• Trust scores → Quality metrics
• Community votes → Feedback loop
• Accuracy tracking → Continuous improvement

🔍 Technical Details

Embeddings

• Model: sentence-transformers/all-MiniLM-L6-v2
• Dimension: 384
• Use case: Semantic similarity search

Vector Database

• Technology: ChromaDB
• Collections: stillme_knowledge, stillme_conversations
• Search method: Cosine similarity

RAG Pipeline

Query → Embedding → Vector Search → Context Retrieval → LLM Prompt → Response

🤔 Questions?

Do you have any questions about the learning process? Ask anything! 💬