STAIR retrieval beats a search index with a table of contents

Long-context models still lose information buried in the middle of a document. Most retrieval systems compensate by chopping text into fixed-length chunks and searching those, which throws away structure the document already has.

A new paper proposes STAIR, which retrieves by walking a document's own table of contents instead of slicing it into arbitrary windows. On a new benchmark called SearchTome, spanning 18 books across 6 domains, STAIR reached 82.6% Recall@1 against 76.9% for a finetuned Differentiable Search Index. Hallucination stayed under 0.05%.

The generalization result matters more than the headline number. STAIR held up better than the baseline specifically when training samples were scarce, which is the condition most teams actually work in. A finetuned index needs a pile of labeled examples per domain to hit its numbers. A method that leans on structure the document already has doesn't need to relearn that structure from scratch for every new corpus.

For anyone building retrieval on long documents, chunk size has always been a tuning knob you fight with. Books have tables of contents, chapters, and sections. Technical docs have headings. Legal filings have numbered clauses. If that structure can replace a chunking heuristic and still generalize with less training data, test it against whatever chunking scheme is already in production. The paper doesn't say how STAIR performs on documents that lack a clean table of contents, which is most of the web.