Isn't model-native web search already a form of a model and agent on a retrieval task?
Certainly it is! It's less about general retrieval, though, and more about what is being retrieved. Web search retrieves from the open internet while enterprise agents retrieve from data that isn't on the web at all.
When is it cheaper or better to use a specialist model instead of free online tools like ChatGPT or Gemini?
The clearest case is proprietary data. If your organization has, for example, a collection of legal documents, customer account data, financial reports, etc., you won't be able to get this from the web & you likely would not be permitted to use free-tier models for analysis anyway, on account of their data retention policies. So proprietary documents, in general, are one such use case.
For another, you may be interested in the guidance that exists in a specific set of tables and documents. The bee example above pulled a table from the Ontario Beekeepers' Association, but maybe you work for an organization where you are required to follow guidance issued by a different organization. In that case you need retrieval grounded in that set of documents, not whatever the open web ranks highest.
Do specialist models actually perform better on complex enterprise data?
Lastly, especially for more complex data, you are likely to see better performance from a specialist model that is very good at writing SQL (Genie) in an environment where it has access to your schemas, metadata, and metrics. A general-purpose model doesn't have access to this context. Working with a generalist model and correcting numerous false assumptions, misinterpretations, etc., can be both costly and inaccurate. That's part of what the KARL article demonstrates.
