Nova - 수정을 통해 학습하고 스스로 미세 조정하는 자체 호스팅 개인 AI

The personal AI that actually learns from its mistakes. Correct Nova once. It remembers forever. Correct it enough times, it fine-tunes itself into a smarter model. All on your hardware. Your data never leaves. You: "What's the capital of Australia?" Nova: "Sydney" You: "That's wrong, it's Canberra" Nova: [saves lesson, generates DPO training pair, updates knowledge graph] --- 3 months later, different conversation --- You: "What's the capital of Australia?" Nova: "Canberra" ← learned permanently No other open-source project combines all of these capabilities. Nova is a sovereign personal AI that runs entirely on your hardware with zero cloud dependencies. It doesn't just answer questions — it gets permanently smarter through a self-improvement pipeline that no other open-source project has: | Nova | Khoj (32K stars) | Open WebUI (124K stars) | | |---|---|---|---| | Learns from corrections | Full pipeline | No | No | | Fine-tunes itself | DPO + A/B eval | No | No | | Knowledge graph | Temporal | Experimental | No | | Hybrid retrieval | Vector + BM25 + RRF | Vector only | Vector only | | Zero cloud dependency | Yes (bundled Ollama) | Partial | Partial | | Prompt injection defense | 4-category detection | No | No | | Messaging channels | 4 (all with allowlisting) | 3 | 0 | | Proactive monitors | 52 across 35+ domains | Automations | No | | MCP (client + server) | Both | No | Client only | Prerequisites: Docker + Docker Compose, NVIDIA GPU (20GB+ VRAM), NVIDIA Container Toolkit # Clone and start git clone https://github.com/HeliosNova/nova.git cd nova_ cp .env.example .env docker compose up -d # Pull models (one-time) docker exec nova-ollama ollama pull qwen3.5:27b # Main model docker exec nova-ollama ollama pull nomic-embed-text-v2-moe # Embeddings Open http://localhost:5173 — that's it. docker exec nova-ollama ollama pull qwen3.5:9b # Vision model docker exec nova-ollama ollama pull qwen3.5:4b # Fast model (greetings, simple queries) User query -> brain.think() -> load context (history + facts + lessons + skills + knowledge graph) -> classify intent (regex, no LLM call) -> retrieve documents (ChromaDB vectors + SQLite FTS5 + Reciprocal Rank Fusion) -> build system prompt (8 prioritized blocks with truncation budget) -> generate response (Ollama / OpenAI / Anthropic / Google) -> tool loop if needed (max 5 rounds, 20 built-in tools) -> stream tokens via SSE -> post-response: correction detection, fact extraction, reflexion, curiosity Meanwhile, 52 monitors run autonomously: -> web search across 35+ domains every 1-24h -> extract knowledge graph triples from every result -> send alerts via Discord/Telegram when something changes -> quiz itself on learned lessons, validate skills, research gaps No LangChain. No LangGraph. No agent frameworks. ~79 files of async Python + httpx + FastAPI. This is what makes Nova unique. Every conversation makes it smarter: - Correction Detection (2-stage) — regex pre-filter + LLM confirmation extracts what was wrong and what's correct - Lesson Storage — topic, wrong answer, correct answer, lesson text — retrieved on future similar queries - DPO Training Pairs — every correction generates {query, chosen, rejected} data for fine-tuning - Reflexion (experimental) — heuristic failure detection (bad tool choices, short answers, exhausted loops) stored as warnings for future reference - Curiosity Engine (experimental) — detects knowledge gaps ("I don't know", hedging, tool failures), queues background research via scheduled monitors - Success Patterns — high-quality responses (score >= 0.8) stored as positive reinforcement - Automated Fine-Tuning — when enough pairs accumulate, runs DPO training with A/B evaluation before deploying python scripts/finetune_auto.py --check # Check readiness python scripts/finetune_auto.py # Full pipeline: train -> eval -> deploy | Tool | What it does | |---|---| web_search | Privacy-respecting search via SearXNG | calculator | Math via SymPy — never does arithmetic in its head | http_fetch | Fetch URLs with SSRF protection (blocks private IPs, DNS rebinding) | knowledge_search | Hybrid retrieval: ChromaDB vectors + SQLite FTS5 + RRF fusion | code_exec | Sandboxed Python (AST-analyzed, tier-restricted imports) | memory_search | Search conversations and user facts | file_ops | Read/write files (path-restricted per access tier) | shell_exec | Shell commands (blocked patterns, tier-restricted, disabled by default) | browser | Playwright-based web browsing with cookie clearing | screenshot | Capture website screenshots | email_send | SMTP email with recipient allowlist | calendar | ICS calendar (create, list, search, delete) | webhook | HTTP webhooks (URL-restricted) | reminder | Schedule reminders via heartbeat system | monitor | Create/manage proactive heartbeat monitors | delegate | Delegate subtasks to parallel sub-agents | background_task | Submit/track long-running background work | integration | Connect to GitHub, Slack, Notion, etc. (10 templates) | de