GraphRAG serialization format can double multi-hop accuracy

A benchmark of 10 graph serialization formats found that verbose formats like JSON waste ~70% of tokens on syntax, while tabular layouts achieve 80% multi-hop accuracy versus 40-70% for nested formats. The author built ISONGraph, a token-efficient format that reduces token usage by 70% and achieves 92% traversal accuracy. The findings suggest that serialization choice significantly impacts LLM reasoning performance on knowledge graphs.