Notion AI Database vs Sup AI

“The primitive here is vector search layered on top of an existing document graph — Notion is essentially running embeddings over workspace content and letting you query the index in natural language. The DX bet is zero-config: you don't set up a vector store, you don't manage chunking, you just ask a question. That's the right call for 90% of users, but it also means you have no visibility into why a result surfaces or why it doesn't, which will frustrate anyone trying to build reliable workflows on top of it. The auto-tagging is the more interesting primitive — inferring structured properties from unstructured content is legitimately hard and if it works reliably it saves real hours of metadata hygiene. I'd ship it for the search alone, but I want to see the accuracy numbers before I trust the auto-tagging on anything consequential.”

“No API, no self-hosting option, and the ensemble approach means your per-query cost is 3-5x a single model call. The benchmark numbers are compelling but I cannot integrate this into a product. Ship an API and I will reconsider.”

“Direct competitor is Obsidian with a vector search plugin, or just asking ChatGPT to summarize a doc you paste in — except those require you to leave Notion, which is the actual moat here. The scenario where this breaks is a workspace with 5,000 pages of inconsistent structure: semantic search will surface loosely related content confidently, and auto-tagging will hallucinate property values on pages with thin content, creating a database that looks complete but isn't. The 12-month threat is not OpenAI — it's Notion itself deciding this should be free to stop the Coda and Linear encroachment, which guts the AI add-on revenue line. What keeps me from skipping entirely is that the integration surface is real: this is search that knows your custom properties, your linked databases, your team's taxonomy. That's not a generic API call.”

“The benchmark result is legitimately impressive and the methodology is transparent. My concern is latency — querying multiple models and aggregating adds significant time. For research and high-stakes questions it is worth the wait. For everyday chat it is overkill.”

“The output of semantic search is ranked page excerpts with the relevant passage highlighted — it reads like a competent research assistant who's actually read your wiki, not a keyword matcher spitting back titles. The taste layer here is delegation: Notion doesn't impose a taxonomy, it infers one from your existing content, which means it amplifies whatever organizational instincts you already have rather than forcing you into a template. The editing surface on auto-tagging is where this needs work — you can correct a wrong tag after the fact, but there's no feedback loop that teaches the model your corrections, so you're fixing the same class of mistake repeatedly. The fingerprint problem is subtle but real: every workspace with this enabled will start converging on the same inferred tag vocabulary, which flattens the idiosyncratic structure that makes a good Notion setup actually useful.”

“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”

“The buyer is a Notion Business or Enterprise admin who's already paying for the AI add-on — this is an upsell to existing customers, not a new motion, which means the TAM is capped by Notion's existing install base and churn rate. The pricing architecture is the problem: $10 per member per month for the AI add-on means a 50-person team is paying $6,000 a year on top of their base plan for features that Coda ships in their base tier and that Confluence is actively cloning. The moat argument is 'our AI knows your Notion graph' but that moat erodes the moment a better-funded competitor trains on the same content type. What would make me reconsider: evidence that AI add-on attach rate is above 40% and that semantic search meaningfully reduces churn — if this is a retention feature disguised as a revenue feature, the unit economics could actually work.”

“Confidence-weighted ensembling is the quiet breakthrough everyone is sleeping on. Individual models plateau — but smart aggregation keeps pushing the frontier. Sup AI scoring 52% on Humanity's Last Exam when no single model breaks 40% proves the thesis.”