XERJ · RC.101 · A NEW NEUROSYMBOLIC MODEL

XERJ Query-1

It turns your question into a small program over facts you have written, and the engine runs it on CPU. No embeddings, no GPU. Lookups name the facts they used.

1VS9–15
wrong of 160 questions from our dev set
Query-1 given each question's records vs RAG on Opus 5.5
$42VS$13,400+
per million questions, CPU time
vs RAG on Opus 5.5, list prices, our 160-question test
17 minVS62–97 h
BM25 index of 10 GB of English Wikipedia
vs embedding it on CPU (estimated)
xerj.org/xerj-query-1Off by default · enable with --facts-model-dirEngine Apache-2.0 · model crates proprietary