The 94% Threshold

Dependency parsing achieves 94% of LLM-based performance for knowledge graph construction (61.87% vs. 65.83% semantic alignment scores).

This comes from SAP’s enterprise GraphRAG research on code migration tasks. The lightweight, classical NLP approach processed documents at a fraction of the cost while retaining nearly all the extraction quality.

The finding suggests a useful heuristic: when the gap between expensive and cheap approaches is single digits, the cheap approach often wins on total value.

Related: 06-atom—construction-bottleneck-problem, 07-molecule—good-enough-classical-nlp, 07-molecule—vectors-vs-graphs