ISONGraph
Property graphs that fit your LLM's context window
9 followers
Property graphs that fit your LLM's context window
9 followers
GraphRAG has a hidden tax: serializing your knowledge graph as JSON or RDF wastes ~70% of tokens on syntax. ISONGraph is a property-graph format designed for LLM comprehension - benchmarked against 9 other formats: 92% traversal accuracy, 80% multi-hop reasoning (vs 40-70%), ~70% fewer tokens. Node/edge CRUD, multi-hop traversal, Cypher-like queries, schema validation, optional semantic search. MIT licensed. Python, JS/TS, Rust, Go, C++, C#.
Hey PH! I'm Mahesh - AI architect by day, open-source builder at AroorA AI Labs by night. While building GraphRAG pipelines I noticed nobody measures what the serialization format costs - everyone benchmarks retrievers, then dumps the subgraph as JSON. So I benchmarked 10 formats. The format alone swung multi-hop accuracy from 40% to 80%. ISONGraph is what I built from those findings. It's MIT, in 6 languages, and the benchmark methodology is public - poke holes in it. Happy to answer anything!
there's a full browser playground at graph.ison.dev/playground.html (and the base-format one at ison.dev) no install needed, load a sample graph or paste your own, see the token savings and structure directly.
the token savings look genuinely impressive, especially for multi-hop queries. one thing that would help adoption though is adding a hosted playground or a simple web demo where you can paste a small graph and try queries right in the browser without setting up a local environment. would make it way easier to evaluate before committing to a new stack.
@napoleon_liu
good news -- we actually have this, just didn't surface it well in the launch itself. graph.ison.dev/playground.html is a full browser playground, no install needed: load a sample graph (social network, knowledge graph, org chart, movie database) or paste your own, and you get live conversion, visualization, and token counts right there.
Appreciate the nudge, clearly this needed to be linked from the launch post itself, not buried on the site.
Tried it on a small knowledge graph and the token savings were immediately obvious, plus the traversal queries just worked without weird syntax overhead. Really nice to see a format that treats LLM context as a first-class constraint.
@quentin_finch really glad the traversal syntax felt natural -- that's usually the hardest part to get right, since most compact graph formats save tokens but make you fight the query layer to get them back. Which language binding did you try it in? And if you push it further and hit a rough edge, please open an issue - real usage like this is exactly what should be shaping this, not just the synthetic benchmarks.
Quick update for anyone evaluating ISONGraph from this thread: there's a full browser playground at graph.ison.dev/playground.html (and the base-format one at ison.dev) no install needed, load a sample graph or paste your own, see the token savings and structure directly. Should've linked this in the launch itself.