Benchmarking graph serialization formats for GraphRAG: token efficiency and reasoning accuracy across 10 formats

Short summary

Benchmarking 10 graph serialization formats for GraphRAG reveals that verbose formats like JSON waste ~70% of tokens on syntax, and format choice alone swings multi-hop reasoning accuracy from ~40% to ~80%. Tabular and relational layouts consistently outperform nested markup because LLMs have seen those patterns extensively in training. The author introduces ISONGraph, an MIT-licensed property-graph format achieving 92% traversal accuracy and 80% multi-hop accuracy with ~70% fewer tokens.

• •Verbose graph serialization formats waste ~70% of tokens on syntax overhead
• •Format choice alone swings multi-hop accuracy from ~40% to ~80%
• •ISONGraph: MIT-licensed format with 92% traversal accuracy, implementations in 6 languages

Generated with AI, which can make mistakes.