nvolt - think in graphs
About a month ago I wrote a short post about nvolt, my little experiment in building a personal knowledge vault. Since then, the project has grown up a bit. It’s still a tool for storing and searching your notes and documents, but now it’s got a few new tricks up its sleeve.
Check it out here: nvolt.raptofu.com
# What is nvolt?
At its core, nvolt is a place to put your knowledge - notes, PDFs, snippets, whatever - and make it searchable. But I wanted more than just a search box. I wanted something that could help me understand my own data, not just retrieve it.
# What’s new
- Google authentication: You can now sign in with Google. No more password headaches.
- Visual knowledge graph: Your knowledge isn’t just a list anymore. It’s a graph you can see and explore. Entities, relationships, clusters—click around and see how your ideas connect.
- Graph-powered RAG: The real magic. When you ask a question, nvolt doesn’t just do a quick vector search. It also builds and consults a knowledge graph, giving you answers that are both fast and deep.
# Under the hood: GraphRAG and nano-graphrag
Retrieval-Augmented Generation (RAG) is the standard way to make LLMs answer questions about your data. Usually, it just grabs the most relevant chunks from a vector database and feeds them to the model. That works, but it’s shallow. Great for fact lookup, not so great for big-picture questions and connecting the dots between different entities of your data.
Microsoft’s GraphRAG paper introduced a smarter approach: build a knowledge graph from your documents, summarize communities in the graph, and use those summaries to answer global questions. I’ve based nvolt’s new engine on this idea, using the excellent nano-graphrag Python library.
# How it works: two-stage answers
When you ask nvolt about a document, here’s what happens:
- Quick answer: The system does a fast vector search and gives you a short, relevant answer. This is almost instant, but sometimes a bit surface-level.
- Deep answer: In the background, the AI starts thinking harder. It consults the knowledge graph—entities, relationships, clusters—and crafts a much more detailed, holistic answer. This can take a few seconds, but it’s worth the wait.
This mirrors the “local to global” approach from the GraphRAG paper: start with the obvious, then zoom out and connect the dots.
# Limitations (for now)
Building a knowledge graph is expensive - both in compute and in my time as the site maintainer. For now, I’ve limited graph construction to the first few thousand tokens of each document. That means the deep, graph-based answer won’t always be available, especially for very large files. I’m working on ways to make this more scalable.
# What’s next?
I’m excited about where this is going. There’s a lot more to do: smarter graph building, better visualizations, maybe even letting users tweak their own graphs. If you have feedback or ideas, let me know. For now, nvolt is my personal lab for exploring what it means to really understand your own data.
