How a GPT-5.4 harness scored above GPT-6 Astra and Claude Code on ExtractBench

Automated excerpt

On LlamaIndex ExtractBench (370 documents, 4,869 pages), my open-source harness KYNEX scores 92.97 F1. RAG: chunking separates table headers from their rows, and embeddings flatten table structure. Table parsing: get rows with their headers preserved.

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.