edgextract
Asking a chat model for JSON gives you a graph that parses, but it cannot tell you which links it guessed. edgextract takes a different route. You write the ontology in YAML, saying which entity types and links are legal, and a small local model answers one closed yes or no question per candidate pair, with a probability. Your cutoff decides what is kept, what goes to a person for review, and what is dropped. On 288 news sentences it produced 218 wrong links, against 504 from a JSON writing chat model. It is still early, with PyPI release 0.1.0 from 3 October 2026, and it runs locally with no cloud call.
What you get
- An ontology you define in YAML, so only the entity types and links you allow can appear in the graph
- One closed yes or no question per candidate pair, each answered with a probability by a small local model
- A cutoff you control, routing each link to keep, human review, or drop
- Local runs with no cloud call, plus a Rust crate and a browser WASM demo to try it without installing anything
The point is not a prettier graph. It is a graph where every link survived a check. Try it on a small corpus first, set the cutoff before you trust the output, and spot check the links it keeps.
Repo: https://github.com/raphaelmansuy/edgextract
Quick poll
What would you use edgextract for first?
- Extracting a knowledge graph from my own docs
- Testing it against a JSON writing model
- Trying the WASM demo in the browser
- Watching it for now
Vote in the comments.
If you found edgextract interesting, I write about tools like this every week in The AI-Native Engineering Leader, what actually works for engineering leaders building with AI agents, and what to skip.
Top comments (0)