Integration demo · Exa Search API
That is the only question a buyer actually asks, and a marketing benchmark never answers it. So here is the bench: twenty five real regulatory questions, each run through two Exa retrieval modes, scored on latency, cost, and groundedness that is verified in code rather than claimed by a model.
The verdict at the end is not a score. It is a map of where the API wins outright, and where it needs a guard.
The readout
The table below is a recorded run, so the page costs nothing to open. This box is live: it fires the same pipeline at Exa right now, both modes plus the guard, and scores the answer the same way. Repeat questions come from cache, and the endpoint is capped at five runs per address per five minutes and forty a day across everyone, because a public URL should not be able to spend without a ceiling.
How the numbers are made
The model answers using only the passages Exa returned, and every claim must carry a
quote. Code then looks for that quote in the page it cites. If the quote is not there
word for word, the claim is not grounded, however confident the sentence reads. The
score you see is verified claims / claims made.
Every question runs through neural and keyword inside the same
API, on the same day, with the same passage budget. This compares two retrieval
strategies against each other, never Exa against somebody else's product.
The leading document is re-checked against a second search with a published-date floor. A supersession only counts when the newer document says so in a quote that code finds in it. When in doubt the guard stays quiet, because a false alarm costs a regulatory reader more than a miss.