STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentation
STAIR is a retrieval system that feeds an LLM a book's table of contents so it can store and fetch information from its parameters, rather than chunking text by length. The authors also release SearchTome, a benchmark built from 18 books across six domains.
- Standard retrievers split long documents into fixed-length chunks and discard global structure like the table of contents; STAIR keeps that structure.
- On SearchTome, STAIR reaches 82.6% Recall@1, versus 76.9% for DSI, 68.7% for DPR, and 59.5% for BM25; the authors report the DSI gap as statistically significant.
- The authors claim a hallucination rate under 0.05% for the generative retrieval setup and say it generalizes to cases with very few training samples.
STAIR is a retrieval system aimed at a known weak spot in retrieval-augmented generation. Standard retrievers cut a long document into fixed-length chunks, which throws away the document's global organization. The authors argue that a corpus's Table of Contents carries useful semantic structure, and that an LLM can be trained to exploit it. Their setup is a generative retrieval approach built on a Differentiable Search Index (DSI), where the model stores information in its parameters and produces the identifier of the relevant passage directly rather than searching an external index. The stated motivation is partly the "lost in the middle" problem, where LLMs underuse information buried in long contexts, so precise retrieval matters more than raw context length.
The headline numbers are on the authors' new benchmark, SearchTome, which is built from 18 books across six domains. STAIR reaches 82.6% Recall@1, against 76.9% for DSI, 68.7% for DPR, 59.5% for BM25, and 13.8% for an off-the-shelf Mistral model. The authors say the gap over DSI is statistically significant. They also report a hallucination rate below 0.05% for the generative retrieval setup and claim the ToC-based approach generalizes to cases with very few training examples. The ablations in the paper are meant to isolate the contribution of the ToC signal, so the comparison that matters most is STAIR against DSI with and without that structure.
Some things to keep in mind before leaning on these results. The benchmark is small (18 books), and it is the authors' own construction, so there is no independent check on how representative it is of real retrieval workloads, which tend to involve many more documents and far messier structure than a clean book ToC. Books have unusually reliable hierarchical structure, and that may be exactly what makes the method look good here. The comparison baselines also matter: a Recall@1 gap of about six points over DSI is meaningful, but the paper's value depends on whether the same gain holds for document collections without tidy ToCs, which the abstract doesn't address. The paper is a v2 revision from October 2026, so the details may have shifted from the first version. Worth reading the full PDF for the ablation tables and the exact definition of the hallucination metric before drawing firm conclusions.