semantic layer

How an AI agent reads your database

A live agent writes SQL against a cryptic schema, and real Postgres judges the answer. Start with no layer and hit ask the model — watch it guess wrong. Then flip to context, then to a constrained tool, and watch the same question go from wrong to provably right.

how the layer reaches the model

Definitions are handed over as text — the model still writes SQL, but now it knows the codes and the formula.

the session — what you asked & got

What is the total revenue from active customers in Germany?

ground truth (engine): …

booting the engine…

the wire — what the model actually sees

TABLE t_0 (
  c1     text,
  c2     text,
  geo_cd text,
  amt_x  numeric,
  st_flg text,
  dt_0   date,
  q      integer
)

semantic layer · editable

edit me, then ask again

whose revenue? — why the layer is the contract

Same question, same rows. Two teams each have their own idea of “revenue” — and the engine gives each of them a different number.

Finance teamblessed

revenue = SUM(amt_x × q)

…

Growth team

revenue = SUM(amt_x)

…

Let your agent query this data anywhere in the app: