Nothing Left Behind: Finding Evidence from a Large Randomized Evaluation Corpus

An embedding-based two-stage retrieval system found 92% of the evidence a systematic review would include across a corpus of 7,354 randomized trials, compared with 45% for standard label-based search, according to an arXiv paper (2610.36910v1). The approach retrieves candidates by meaning using text embeddings, then uses a language model to keep only relevant items, at a cost per query that rises with search depth. The authors conclude the results support a two-stage architecture combining broad candidate retrieval with query-specific evidence filtering.

arXiv:2610.36910v1 Announce Type: new Abstract: Many AI systems need to find every relevant items in a large collection and then organize what they found. The standard way to do this is to tag each item with labels from a curated vocabulary and search on those labels. Those labels are useful for organizing results, but searching on them misses items that do not fit a label and returns too much when labels are broad. We compare this approach with an embedding-based approach: find candidates by meaning using text embeddings, then have a language model keep only the relevant ones. In a corpus of 7,354 randomized trials, label search found 45% of the evidence a systematic review would include, while the new approach found 92%, at a cost per query that rises with search depth. These results support a two-stage architecture that combines broad candidate retrieval with query-specific evidence filtering.