v0.14.3 — Answer quality at its best measured level (75.5% on a 400-question benchmark)

Answer quality at its best measured level (75.5% on a 400-question benchmark)

Improved

We swapped the reader to GLM-5.3 Flash (then identified as Ox Alpha) and turned reasoning effort to max: 69.3% to 75.5% average score on the 400-question memory benchmark (323/400 correct, 80.8% binary). Result setup: reader GLM-5.3 Flash (Ox Alpha); judge GLM-5.2; effort=max; run 2026-08-23; ~44% slower than default effort; this benchmark setup is not the chat default. Along the way we moved dense HNSW indexes from float32 to halfvec (3.6× less disk, 4.7× faster, recall intact), fixed an O(n²) validation that kept index builds at hours, and built, then disabled, a ColBERT rerank that helped synthesis but hurt abstention and extraction.

Read the full report →