IBM's STAIR retriever cuts invalid references 65× and hits 82.6% accuracy by using document structure
solyarisoftware · x · 2026-11-02
IBM researchers unveiled STAIR (STructure Aware Information Retriever), which replaces fixed-length chunking with a Table-of-Contents-style global structure for retrieval.
- Idea: standard RAG chunking discards chapter headings and semantic hierarchy; STAIR guides an LLM to locate the right section via the corpus's structure, like using a book's contents page.
- Results: on SearchTome, a benchmark of 18 books across 6 domains, STAIR retrieves the correct section first 82.6% of the time vs. 76.9% for a fine-tuned DSI baseline, and beats BM25 (59.5%), DPR (68.7%), and out-of-the-box Mistral (13.8%).
- Headline number: invalid section references fell from 3.25% to 0.05% — a 65× reduction.
- Code and the SearchTome dataset are open-sourced on GitHub and Hugging Face.
More from AGI Musings
- Apple's new CEO John Ternus draws a line against AI automation: 'Creating products is deeply human' — anthara_ai · 2026-11-02
- Will Manidis' year-long essay argues abundant intelligence will reorder economic power beyond SF — abhiadesai · 2026-11-02
- AI bird sticker project swarmed by artists on Xiaohongshu over 'poisoning' claims — AlchainHust · 2026-11-02
- Why Anthropic's 6-to-1 agent-hour ratio isn't a staffing forecast — krishnan · 2026-11-02
- Managing agents is babysitting: every ping arrives equally loud — krishnan · 2026-11-02
- Anthropic report: agent hours already outnumber employee hours 6-to-1 — krishnan · 2026-11-02