Cohere Compass 2 vs Notion AI 3.0

“The primitive here is a retrieval pipeline that can ingest mixed-format documents — PDFs with embedded charts, scanned tables, image-heavy slides — and return semantically relevant chunks without requiring a preprocessing ETL step per modality. That's a real problem: anyone who's tried to build RAG over a 10,000-document enterprise library knows the pain is 80% in the ingestion layer. The DX bet is that Cohere handles the multimodal parsing so you don't glue together a PDF parser, a table extractor, and a vision model yourself. The on-prem deployment option is actually the headline feature for the buyer, not the multimodal part — that's what gets it past legal review. My skip concern is documentation: the blog post is long on capability claims and short on API surface, schema design, and what 'image understanding' means at query time versus index time. Show me the query API, then we'll talk.”

“The primitive is: web search → LLM synthesis → structured Notion database write, and that is three API calls dressed up as a platform feature. If you already have a Notion workspace and an API token, you can replicate the core loop with a small script hitting Perplexity's API, a basic extraction prompt, and Notion's database API — in an afternoon. The DX bet Notion made is betting users won't want to maintain that script and will pay for the integration instead, which is a legitimate bet, but it's not craft — it's convenience. The moment of truth breaks when a developer needs to customize the research schema, add preprocessing steps, or integrate findings into an existing automation pipeline: Notion's closed orchestration layer blocks all of that. The specific technical decision that causes the skip is the lack of any webhook, API surface, or composability for the Research Mode itself — you get a black box, not a primitive.”

“The direct competitors are Azure AI Search with multimodal indexing, AWS Kendra, and increasingly any RAG stack bolted onto GPT-4o's native PDF vision. Compass 2's real differentiator is not the multimodal capability — every major cloud provider is shipping that — it's the on-premises deployment for enterprises with data residency requirements, combined with a retrieval model trained specifically for enterprise document retrieval rather than general web content. The scenario where this breaks is at the 'chart understanding' claim: interpreting a bar chart semantically in a way that survives a specific quantitative query ('find all documents where Q3 revenue exceeded Q2') is a much harder problem than the blog post implies, and I've seen this class of tool hallucinate chart data confidently. What kills this in 12 months isn't a competitor — it's that the chart and table comprehension doesn't hold up under production query loads and the feature gets quietly deprioritized. I'm shipping it narrowly: for text-heavy PDFs with some visual elements in air-gapped environments, this is probably the best available option right now.”

“The direct competitor here is Perplexity Pages plus a Notion export, and honestly that pipeline exists and works — but the friction of leaving Notion, running research, and re-importing structured data is exactly the gap this fills. The scenario where this breaks is multi-step research requiring domain-specific depth: ask it to synthesize primary legal filings or niche technical papers and the web-browsing layer will hallucinate citations or surface SEO slop. What kills this in 12 months isn't a competitor — it's OpenAI or Anthropic shipping deep-research natively into API responses, making Notion's orchestration layer redundant. For now it earns a weak ship because the workflow integration is genuinely tighter than the alternatives, not because the research quality is exceptional.”

“The buyer is a VP of IT or Chief Data Officer at a regulated enterprise — financial services, pharma, government — and the budget comes from the data infrastructure or compliance line, not a software tools budget. That's a real check-writer with a real problem: they have document libraries they legally cannot send to OpenAI's API, and they need search that works across formats. The on-prem deployment option is the actual moat here, not the multimodal capability — Cohere has been building that distribution channel for two years and it creates genuine switching costs once it's integrated into an enterprise's document management stack. The risk is that the pricing model is 'contact sales' all the way down, which means a long sales cycle and high CAC that has to be recovered on large contracts. What survives the model-gets-cheaper scenario is the enterprise integration layer and compliance certifications, not the retrieval model itself — Cohere needs to be pricing for that, not for compute.”

“The thesis Compass 2 is betting on: enterprise knowledge is fundamentally multimodal — it lives in slide decks, scanned contracts, financial tables, and annotated diagrams — and the first retrieval system that treats those formats as first-class citizens rather than edge cases will own the enterprise search layer. That's a plausible and falsifiable bet, but the dependency is that 'understanding' a chart means something semantically useful at query time, not just 'we embedded the image.' The second-order effect that matters here isn't faster document search — it's that if this works, structured data that currently lives locked in PDFs becomes queryable without a data engineering team to extract it, which shifts power from BI teams who own structured pipelines toward anyone with a document library. Cohere is riding the trend of on-premises LLM deployment for regulated industries — that trend is real and accelerating, and they're on-time to it, not early. The future state where this is infrastructure is 'every regulated enterprise has a Compass instance the same way they have an Active Directory instance.' I'd believe that in five years if the chart comprehension claim is real.”

“The thesis here is falsifiable: in three years, the primary interface for knowledge work is a persistent workspace that accumulates structured context over time, and retrieval-augmented generation over that context outperforms ad-hoc chat. Notion is betting that owning the context store — the databases, the linked pages, the historical docs — gives them a durable advantage as the research agent layer commoditizes. What has to go right: the AI Databases need to become genuinely queryable organizational memory, not just populated tables. What has to not happen: Microsoft Copilot cannot get good enough at structured knowledge organization to make Loop the default; and OpenAI's deep research cannot ship a native export-to-structured-data flow. The second-order effect that matters most is that if this works, it shifts research workflows from search-then-synthesize to synthesize-into-memory, and the team that owns the memory layer owns the workflow — Notion is riding the trend toward ambient knowledge bases and they are on time, not early.”

“The job-to-be-done is clear and singular: turn a research question into a structured, cited Notion database without leaving the app. That's a real job with a real switching cost reduction, and Notion is one of the few players with the workspace context to make the output land somewhere useful rather than a blank chat thread. The onboarding question is whether triggering Research Mode and getting a populated database takes under two minutes from a cold start — if it requires setting up database schemas and configuring AI properties first, that's a configuration screen masquerading as value delivery. The product opinion here is strong though: structured output with citations is a genuine point of view, not a flexibility punt, and that's the specific decision that earns the ship.”