Published
Exa posts a demo that puts web results into Jev's state
Ishan Goswami posted a demo at demos.exa.ai/jev-web-context. The post says Jev without web search is confidently wrong, and Jev with Exa results is more accurate. The page we loaded lists seven yes/no prompts and does not print a before-and-after. Diogo Almeida quoted the post and asked if it was real.
Ishan Goswami posted on September 22 that Jev without web search confidently returns wrong outputs, and that Jev with Exa results is much more accurate. The demo link is in the next post: demos.exa.ai/jev-web-context. A reply credits @yoimnotkesku. Exa’s account replied with a fire emoji and no figure.
The page we loaded is titled Live Web Context for Jev. Seven prompts are on the screen: whether you need an umbrella in Tokyo, whether a Starship booster was caught, whether GPT-6 is out, the iPhone 18 Pro camera, the Fed’s latest move, whether AI hacked real companies, and whether LLM judges love long answers. The instruction says to pick one, or to type a yes/no question. Jev is asked with live Exa results and without them. The search text is described as going straight into Jev’s state. The HTML we received does not include a probability for any of those prompts.
Diogo Almeida quoted the original post at 07:08 UTC and wrote, “is this real? or an exajevration? (sorry)”. The thread does not show a follow-up from him with a rerun. The post does not name a model id, a trial count, or a repository. We did not press the buttons.
jevsearch is the closest measurement already on this desk: 41 labelled queries over TypeSafe’s docs, Hit@1 83 percent after a Jev re-rank against 41 percent for the keyword pass. That set is a fixed corpus. Goswami’s claim is about questions whose answers move, and the demo page does not print the with-search and without-search pair.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @TheIshanGoswami, September 22, 2026, 00:38 UTC. The claim in the post is that Jev without web search confidently gives wrong outputs, and that Jev with web search is much more accurate. The next post, https://x.com/TheIshanGoswami/status/2102195821429309649, links https://demos.exa.ai/jev-web-context. A reply credits @yoimnotkesku. @ExaAILabs replied with a fire emoji and no numbers. The page we fetched is titled Live Web Context for Jev. Visible prompts: Umbrella in Tokyo, Starship booster catch, Is GPT-6 out, iPhone 18 Pro camera, Fed's latest move, AI hacked real companies, LLM judges love long answers. The page says to pick an example or type a yes/no question, that Jev answers with and without live web evidence from Exa, and that the search results go into Jev state. No pair of probabilities was in the HTML we received. Diogo Almeida quoted the original post at 07:08 UTC, https://x.com/CompleteSkeptic/status/2102294042185023816, and wrote "is this real? or an exajevration? (sorry)". We did not click the examples and we did not send a question. The post does not name the Jev model id, the number of trials, or a repository.
Compare
jevsearch re-ranks a keyword pool on TypeSafe's own docs and prints Hit@1, 83 percent against 41 percent, on 41 labelled queries. This demo claims that putting Exa results into state changes yes/no answers about the live web. The page we loaded does not print those pairs, so there is no table to set beside jevsearch. MotherDuck's AG News run is a closed label set, not a question about today's news.
Terms
- web context
- In the Exa demo, search snippets placed into the state Jev judges. The page says Jev is asked the same yes/no question with that state and without it. No scored pair was printed on the page we loaded.