Examples in this demo are picked to stress Jev's prior knowledge.
Some are recent facts like GPT-6, the Fed's latest move, or today's weather.
Others need harder retrieval: documented experiences, niche technical evidence, opinions matching a specific description, or research papers.
Exa finds the evidence, the selected results go straight into Jev state, and Jev answers the same typed question with and without them.
The demo sends a query to /search and asks for six results with short, query-relevant highlights.
For the presets, the search query is written for the evidence we want to find. For a custom question, your question is used directly.
/search
Each selected result becomes a short string containing its title, source, date, and highlights.
Those strings go directly into state.search_results.
There is no model between Exa and Jev. The Request drawer shows the exact state Jev receives.
Turning a source on or off only reruns Jev. It does not search again.
The demo asks Jev the same typed question twice: once with Exa results in state.search_results, once with search_results: [].
The difference between those two answers shows what the web evidence changed.
Each preset defines its own multiple-choice question and answer criteria. Custom questions are asked as yes/no.
The main result shows the answer with web context. Before web shows Jev's answer without it.
/v1/systemone
Source weights are computed by this demo, not returned by Jev. Jev only returns the answer and its probabilities.
For each selected source, the demo runs Jev once more with that source removed from state.search_results.
The weight is the winning probability with all sources minus the winning probability without that one, in points. Positive means the source supports the answer, negative means it pulls against it.
The Weights tab in the Request drawer shows the raw array, one number per source in evidence order.
Every run is live: the search, every Jev call, and the latencies shown are the real ones.