STAIR uses tables of contents to guide AI retrieval, a user says

A user’s summary of an IBM paper describes STAIR as a way to preserve document hierarchy that gets lost when long texts are split into chunks by length. It uses a table of contents as an addressing system for a generative retriever—a model that stores and retrieves information from its own parameters.

The summary reports 82.6% Recall@1 on SearchTome, compared with 76.9% for a fine-tuned Differentiable Search Index, calling the gap statistically significant. It also reports hallucination below 0.05% and says tests showed generalization where very few training samples exist.