developer tools
The Financial Number Kept Its Unit
Deterministic search and bounded reads beat dense retrieval on one 780-page government report—but not classical BM25.
Summary
Deterministic search and bounded reads beat dense retrieval on one 780-page government report—but not classical BM25.
READ answered 58.8 percent of 51 verified questions against 15.7 percent for dense retrieval and 35.3 percent for a tuned version. A top-k tool in the same loop reached 27.5 percent, while BM25 remained statistically indistinguishable from READ, narrowing the finding to embedding-based versus embedding-free retrieval.
Why it matters
Deterministic search and bounded reads beat dense retrieval on one 780-page government report—but not classical BM25.
Limits and context
No additional limitation was separately recorded.
Key claims
Deterministic search and bounded reads beat dense retrieval on one 780-page government report—but not classical BM25.
Evidence: source-2026-08-08-020
Sources
- arXiv preprint 2608.06305arXiv · primary research
Corrections
No corrections have been recorded for this story.