Nova: Self-Hosted AI Learns From Corrections and Fine-Tunes Itself

Nova is a self-hosted personal AI assistant designed to learn from user corrections and fine-tune itself for improved accuracy. Unlike other AI assistants, Nova runs entirely on the user's hardware, ensuring data privacy and eliminating cloud dependencies. It employs a unique self-improvement pipeline that includes correction detection, lesson storage, DPO training pair generation, reflexion, curiosity engine, success patterns, and automated fine-tuning.

Key features of Nova include a knowledge graph for temporal understanding, hybrid retrieval (Vector + BM25 + RRF), and prompt injection defense. It supports multiple messaging channels and proactive monitors. The system is built using async Python, httpx, and FastAPI, without relying on LangChain or LangGraph.

Nova's architecture involves a multi-step process: user query, context loading, intent classification, document retrieval, system prompt building, response generation, tool loop (if needed), token streaming, and post-response actions (correction detection, fact extraction, reflexion, curiosity). This comprehensive approach enables Nova to learn from its mistakes and continuously improve its performance.

While Nova offers significant advantages in terms of privacy and personalization, it also has certain limitations. The hardware requirements (NVIDIA GPU with 20GB+ VRAM) may restrict its accessibility to users with high-end systems. Additionally, the technical setup process involving Docker and command-line commands could be challenging for some users.

_Context: This intelligence report was compiled by the DailyAIWire Strategy Engine. Verified for Art. 50 Compliance._