Nobody budgets for the tokens an agent spends finding its way around.
Lists tables. Samples rows. Tries a join, fails, tries another. Then rediscovers the whole schema on the next question.
Token spend is an architecture problem.
@coginiti.co
@coginiti.coNobody budgets for the tokens an agent spends finding its way around.
Lists tables. Samples rows. Tries a join, fails, tries another. Then rediscovers the whole schema on the next question.
Token spend is an architecture problem.
@mmullins.coginiti.coTelling a real semantic layer from a rebranded metrics catalogue, in 3 questions:
1. Can anything other than your BI tool read it?
2. Is policy enforced here or only documented here?
3. Does it survive a question nobody anticipated?
@coginiti.coThe failure mode of an agent on raw schemas isn't that it breaks. It's that it succeeds. Counterintuitive right?
The agent returns some number. Correctly formatted, plausible, wrong, because it picked orders.amount and you define revenue net of returns.
Nobody catches it.
@mmullins.coginiti.co"Can't we just tell the agent what it's allowed to see?"
You can. It'll mostly kinda work, and "mostly" is doing a lot of work there.
Prompt policy is a request. Semantic layer policy is a boundary.
@coginiti.coAn analyst joining your company gets onboarding and a colleague to ask.
An agent gets a connection string.
No institutional memory, nobody to ask, machine speed. When it guesses wrong about revenue it does so confidently and at scale.
@mmullins.coginiti.coAn agent's confidence is not correlated with its correctness.
When an analyst doesn't know how you define revenue, they ask, and the question is visible.
When an agent doesn't know, it picks, and the picking is invisible.
@coginiti.coSemantic model. Ontology. Context engineering. Three terms used interchangeably by people selling three different things.
Richard Hall sorts them out at Big Data LDN. AI Context Theatre, Thu 24 Sept, 10:40. What each one does, and when it doesn't.
@coginiti.coFor 20 years the semantic layer was a BI convenience. Friendly names so nobody had to remember which table held revenue.
That was built for humans clicking dashboards. The user has changed.
Matthew Mullins, Big Data LDN, Wed 23 Sept 10:40, Lightning Talks Theatre.