ISONGraph

ISONGraph is an open-source AI knowledge-graph format for LLMs that serializes property graphs into a compact, tabular representation so retrieved subgraphs fit in a model's context window without wasting tokens on syntax. It targets developers building GraphRAG and agentic AI pipelines, where a graph is normally dumped as JSON or RDF and much of the prompt is consumed by brackets, quotes, and repeated keys. Benchmarked against 9 other formats, ISONGraph reports roughly 70% token savings (about 68.6% fewer tokens), 92% data-traversal accuracy, and 80% multi-hop reasoning accuracy versus 40-70% for alternative serializations, with 90% knowledge-graph QA accuracy. Beyond the format, the library provides an in-memory property-graph store with node and edge CRUD, O(1) lookups, multi-hop traversal, shortest-path search, schema validation, and an SQL-like ISONQL query language, plus optional embedding-based semantic search. It is MIT licensed and ships across Python, JavaScript/TypeScript, Rust, Go, C++, and C#, so the same format can be produced and consumed on either side of a service boundary. The Python package installs with pip install ison-graph and requires no external graph database. For teams whose GraphRAG cost and accuracy hinge on how a knowledge graph is written into the prompt, ISONGraph swaps the serialization layer rather than the retriever or the model.

Best for developers building GraphRAG or agentic AI pipelines who want to cut the tokens spent serializing a knowledge graph into LLM prompts while preserving traversal and multi-hop reasoning accuracy.