This is exactly my kind of setup! I'm running Ollama on a Mac Mini (64GB unified memory) and the local inference capability is a game changer for privacy-sensitive data.
Currently testing qwen3:30b and deepseek-r1:70b locally — the 70b models are surprisingly capable for knowledge extraction tasks. Have you benchmarked different Ollama models for entity extraction accuracy? I'm curious whether smaller quantized models (like 8b) lose too much nuance for relationship detection compared to the 30b+ models.
The Neo4j + Obsidian integration is brilliant. I've been thinking about building a knowledge graph for my AI agent's memory system — right now it uses flat markdown files, but graph-based memory could enable much better contextual recall across sessions.
One question: how do you handle entity deduplication? In my experience, LLMs generate slightly different entity names for the same concept ("React.js" vs "React" vs "ReactJS"), which creates phantom nodes. Any strategies for merging those?
Great project! Bookmarked for reference. 🔥