MagLab / Case studies / Knowledge graph
AI read a whole library and built a knowledge graph from it
AI and knowledge graphs: an open EleutherAI community project
The answers sit in thousands of documents, and finding the right one takes days of rereading. We built a pipeline that reads everything and links what matters into a graph: you ask a question and get the answer and its source.
Problem
A whole field of science and no map.
- Thousands of papers, posts and transcripts on AI safety.
- The goal: find every risk, every measure against it and the links between them.
What we did
LLMs read, the graph remembers.
- An LLM goes through every paper and pulls out the main ideas: risks, methods and measures.
- The ideas become vectors, and similar ones merge.
- Everything is linked into one graph with search by meaning.
- The graph was built in an open EleutherAI community project on AI safety: our part was the parsing pipeline and loading into the graph.
What changed
- One question, and the graph shows every measure linked to a risk across the whole library.
- The whole library was processed in an evening instead of a month, one paper at a time.
- The code is open, and a research paper is in preparation.
Built with
- OpenAI
Also: Python, vector search and the knowledge graph, Docker.
Other cases
01Restaurant chain
The whole business on one screen
The money lived in five systems. Now the whole business is on one screen every morning.
02High-load platform
Ten microservices, a terabyte a day and one source of truth
Data was lost between ten services. Now there is one event stream and one database.
04Mosena
Mosena: we did it all ourselves, start to finish
There was an idea for an app. Now there is a working AI product, built start to finish.
Let's grow together
Tell us what eats your time and money. We reply within a working day and suggest a time for a call.
The first conversation is 30 minutes, free, and you talk to an engineer, not a salesperson. Email: hello@maglab.app
Who you will talk to
Gleb Maksimov, founder of MagLab. Eight years in data, ML and high-load systems: startups, research labs, corporations. LinkedIn