OPEN LINK ↗
// 30m agoMODEL RELEASE
XERJ Query-1 Compiles Questions Into Checkable Programs
XERJ Query-1 turns natural-language questions into executable search programs over stored facts, returning evidence traces and refusing with UNKNOWN or CONFLICT when data is insufficient. XERJ reports one wrong answer versus 9–15 for tested RAG pipelines, at $0.042 per 1,000 questions on four CPU cores. [Announcement](https://www.xerj.org/blog/xerj-query-ai-model)
// ANALYSIS
XERJ attacks RAG’s costly answer-generation loop with a narrower, more auditable contract. The approach is compelling for structured fact lookup, but its current benchmark and schema limitations make it a promising complement to RAG—not a universal replacement.
- –Executable programs and record-level traces make errors inspectable.
- –Local CPU inference avoids embeddings, GPUs, network calls, and per-question LLM fees.
- –Facts must be explicitly written into a namespace; arbitrary document extraction is not yet supported.
- –The model over-declined 383 of 1,600 evaluation items and currently lacks sums and threshold operations.
- –Large fact namespaces slow down and increase UNKNOWN responses, limiting near-term scale. [Product site](https://xerj.org/)
// TAGS
xerj-query-1ragsearchagentinferencestructured-outputlocal-first
DISCOVERED
30m ago
2026-11-06
PUBLISHED
57m ago
2026-11-06
RELEVANCE
9/ 10
AUTHOR
d0znpp