IBM's STAIR uses a book's table of contents as the index: 82.6% recall, 0.05% hallucination

alex_verem · x · 2026-09-26

IBM Research's STAIR (STructure Aware Information Retriever) paper proposes skipping equal-sized chunking and vector embeddings entirely: the model exploits a document's global structure, like its table of contents, and learns to answer each question with the right section title — the outline itself becomes the index.

The team built SearchTome, a benchmark of 18 books across law, medicine, finance, education and the sciences with tens of thousands of questions. STAIR hits 82.6% Recall@1 vs 76.9% for the best comparable method, 68.7% for dense retrieval, 59.5% for keyword search, and 13.8% for off-the-shelf Mistral. Hallucination rate (pointing to nonexistent sections) is 0.05%, vs 3.25% for the closest method.

Caveats: requires a document with a ToC, fine-tunes a separate 7B model per book (200 passes each), hasn't been tested on company-sized collections, and doesn't verify whether answers built from the section are correct.

Related event: IBM's STAIR Uses Book TOCs as Index, Slashing Hallucinations 65x vs RAG(3 posts)→

Original post →

More from Research

Research channel →