How I Stopped Worrying and Made RAG Work in User-Message (80% Compliance, Solo)

Title: How I Stopped Worrying and Made RAG Work in User-Message (80% Compliance, Solo)

I'm Deniz, a solo developer. For the past month, I've been building SKV Network — an open knowledge base for AI that forces models to follow rules.

And I have an interesting engineering finding: rules in the system prompt don't work. Rules in the user message do.

What is SKV?
SKV stores "cubes" — atomic rules that AI should follow. 75+ constitutional rules (ethics, law, psychology, physics) indexed in Qdrant. Before answering, the model searches for relevant rules and applies them.

The Key Finding
I ran 1000+ tests. Results:

Approach Empty Answers Compliance
Rules in system prompt 29% 14%
Rules in user message 10% 80%+
System prompt kills responses. The model treats long rules as noise and stays silent. User message works — the model treats rules as context.

Features
Semantic search via Qdrant
AI consultation from 3 models simultaneously
Blind jury trial (3-5 AI models vote)
Export to PDF, Charts, Code, CSV
Async queue (no 503 errors)
Secure authentication
Stack
FastAPI + PostgreSQL + Qdrant + Docker + Polza API. All on a single VPS.

Try it: https://skv.network
API docs: GET https://skv.network/api/v1/info
Support: USDT (TON) UQDSasHNGFnBNp1G55IZ3I4vmf8IkGzoxQOyx3PcyaMZdJSH
GitHub: https://github.com/denizchavdarov-arch/skv-network