IBM's STAIR Shows Structured Retrieval Still Has Room to Run

IBM's STAIR Shows Structured Retrieval Still Has Room to Run

IBM introduced STAIR, a retriever that uses a document's table of contents — rather than embeddings or a vector database — as the core retrieval unit. On the SearchTome benchmark (18 open textbooks across 6 domains), STAIR hit 82.6% Recall@1 versus 76.9% for a fine-tuned DSI baseline, and cut invalid non-leaf section predictions to 0.05% versus 3.25% for DSI [4][5][6].

Worth being precise here: the viral "65x fewer hallucinations" framing overstates things. That 65x figure applies to one narrow structural-error metric, not overall answer accuracy — the real end-to-end error improvement is closer to 1.3x (18.67% vs. 24.31%). Still meaningful, just not the headline number circulating on X.

What's notable for anyone building retrieval systems: STAIR works without a vector store at all, built on Mistral 7B, and focuses specifically on structured documents. As @alex_verem's thread and others note, it's a real step toward retrieval that validates what it pulls before generation — relevant well beyond textbooks.

Graphify Turns Any Folder Into a Local Knowledge Graph, Rockets Past 137k GitHub Stars

Graphify, an open-source tool, converts codebases, PDFs, SQL schemas, configs, and even images into local knowledge graphs using deterministic AST parsing (tree-sitter) with an optional semantic layer — no vector store required [7][8][9]. It ships as a /graphify skill for Claude Code, Cursor, and similar tools, and has exploded to 137k stars and 13k forks, gaining roughly 5,500 stars a week.

Users are reporting that persistent, structured graphs materially improve Claude's consistency on coding and design tasks — essentially giving the model a long-term memory of a project's actual structure rather than re-deriving it each session.

The broader signal: "second brain" and personal knowledge management workflows are moving from vector-search-and-pray toward graph-based structure, echoing the same theme as IBM's STAIR — structure beats brute-force embedding search for precision.

EU Council Opens the Door to AI Training on Personal Data via "Legitimate Interest"

On October 21, the EU Council agreed its negotiating position on the Digital Omnibus, adding a new GDPR provision (Article 88 bis) that would let companies process personal data to develop or operate AI systems under "legitimate interest" grounds, rather than requiring explicit consent — subject to case-by-case assessment and the standard right to object [10][11][12].

This is a Council mandate, not final law — it still needs to be negotiated with the European Parliament, and some of the Commission's original transparency and override safeguards have already been stripped from the Council's version. For Nordic and EU enterprises handling meeting data, customer records, or any personal data feeding AI systems, this is the provision to watch closely over the coming months.

X reaction splits predictably: critics call it "digital expropriation" of personal data, while others note it simply formalizes what many enterprise AI deployments are already doing informally. Either way, Parliament's response will determine how much this actually changes practice.

What This Means For Your Meetings

Today's news traces one throughline: meeting and knowledge tools are being judged less on "did it transcribe accurately" and more on "did it retrieve the right thing at the right time, without making it up." Otter's Work Mode and Live Assist push transcription into active, real-time assistance — but that only works if the retrieval underneath is trustworthy. IBM's STAIR and Graphify's explosive adoption both point to the same fix: structure — tables of contents, knowledge graphs, AST parsing — beats pure vector similarity for precision and auditability. That's precisely the architecture Proudfrog has built around: a structured knowledge graph of your meeting history, not just a searchable transcript dump.

The GDPR development matters more quietly but just as much. As "legitimate interest" potentially becomes a valid basis for AI training on personal data, the gap widens between tools that process your meetings as disposable training fodder and tools — like Proudfrog, built on Nordic data principles — that treat your meeting history as a private, structured asset you control and query, not a resource to be mined.

Put together: the market is converging on the idea that your meetings are a long-term knowledge base, not a transcript archive. Whoever retrieves from that base most accurately — and most respectfully of data provenance — wins the next phase of this category.

Key takeaway: The industry is racing toward structured, verifiable retrieval across meeting history — and whether that history respects your data rights is becoming just as important as how smart the AI sounds.

Sources

1. https://otter.ai/blog/otter-rethinks-the-meeting-new-ai-preps-you-coaches-you-and-follows-through
2. https://otter.ai/blog/otter-ai-introduces-live-assist-the-first-live-coaching-agent-for-every-call
3. https://otter.ai/work-mode
4. https://indrajitswain.substack.com/p/ibms-stair-does-not-hallucinate-65x
5. https://arxiv.org/abs/2609.03874
6. https://x.com/alex_verem/status/2110055545852043385
7. https://github.com/safishamsi/graphify/tree/main
8. https://github.com/Graphify-Labs
9. https://ossradar.net/repo/Graphify-Labs/graphify
10. https://anoni.net/news/en/2026/10/eu-digital-omnibus-ai-training/
11. https://legal247.no/en/insights/digital-omnibus-gdpr/
12. https://cisonews.ai/regulation/eu-council-agrees-position-digital-omnibus-gdpr-ai-training