Jev AI · Retrieval & knowledge

Jev AI with live web context

Better context in, better decisions out. Ask a yes/no question about anything current. Live web results go straight into Jev state as evidence. See Jev's answer before and after the web, and which sources moved it.

Ask Jev with live web context

Search query and yes/no rules

Search query defaults to your questionAnswer yes whenAnswer no when

Decision

Recorded example · 2026-09-23 · jev-1.13.0

Has SpaceX caught a Starship booster with the launch tower?

Yes · 99% confident

Before web

With web

99%

+95 ptsNo → Yes

No single source moved the answer on its own: the 6 sources agree. Pick an example or type your own question to run it live.

Evidence

Web search results appear here and go straight into Jev state. In the recorded example, the search returned:

In engineering feat by SpaceX, mechanical arms catch ...pbs.org · 2024-10-13
Watch SpaceX catch Starship Super Heavy booster with 'chopsticks' in this animation | Spacespace.com · 2024-07-02
SpaceX pulls off unprecedented feat, grabs descending rocket with mechanical arms - CBS Newscbsnews.com · 2024-10-13
SpaceX catches Starship rocket booster in dramatic ...nbcnews.com · 2024-10-13
SpaceX Catches a Falling Starship—A First in Spaceflight History | Scientific Americanscientificamerican.com · 2024-10-15
SpaceX Launches Starship Mega-Rocket and Catches Its Booster in Midair on First Trysmithsonianmag.com · 2024-10-15

How live web context works

Jev reads only the state you send. It does not browse. Live web context adds the missing step: fresh evidence from the web, with nothing rewriting it on the way.

01
Search the web
Your question, or your own search query, runs as a live web search. Each result becomes one evidence line: title, site, date and the most relevant passages.
02
Ask Jev twice
The evidence goes verbatim into state.search_results. Jev answers the same yes/no question with the evidence and with an empty list, so you can see what the web changed.
03
Weigh each source
Jev AI reruns the question once without each source. The drop in the winning probability is that source's weight. Exclude a source and rerun to check how much the answer depends on it.

Live web context API

Use the same flow from your backend with a Jev AI API key. Leave out attribution for the fastest response, or send your own sources to skip the search. See the API reference.

curl https://jev-ai.pro/api/v1/web-context \
  -H "Authorization: Bearer $JEV_AI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question":"Should I take an umbrella in Tokyo today?","query":"Tokyo weather forecast rain today","criteria":{"yes":"Rain is likely enough that carrying an umbrella is worth it","no":"Dry or negligible chance of rain; leave it at home"},"num_results":6,"attribution":true}'

Pricing

2 credits for a run with a fresh web search, including source weights.
1 credit to rerun with some sources excluded.
Token billing: actual input tokens across all Jev calls, plus 50,000 tokens per search. See plans.
Failed runs are refunded, including when the search or Jev times out.

Sources and method

Reviewed 2026-09-23.

Search provider: web results come from Exa (exa.ai), a web search API. Your question and search query are sent to Exa to run the search; see our privacy policy.
Decision: each run asks Jev one Choice question (TypeSafe documentation, “Choice”) with yes and no options, once with the search results and once with none. Nothing rewrites the results in between.
Source weights: computed by Jev AI by rerunning the question without each source, following TypeSafe’s composite-scoring pattern of independent checks combined in code. They are not a Jev response field.
Limits: answers are only as current and reliable as the retrieved pages. Check the sources before acting on an answer.

Live web context FAQ

Back to the Jev AI playground

Does Jev search the web by itself?

No. Jev is a fast decision model that only reads the state you send it. This page runs a web search for your question first, then passes the search results to Jev as evidence. No other model rewrites or summarizes them in between.

What do "Before web" and "With web" mean?

Each run asks Jev the same question twice: once with an empty evidence list and once with the search results. Before web shows what Jev assumes on its own. With web shows the answer grounded in current sources. The gap between them is what the evidence changed.

How are source weights calculated?

For each source, Jev AI reruns the question with that source removed and measures how much the probability of the winning answer drops. A large positive weight means the source pushed Jev toward its answer; a negative weight means it pulled the other way. Weights are computed by this site, not returned by Jev.

How much does a web context run cost?

A run with a fresh search costs 2 credits. Rerunning after you exclude sources costs 1 credit because no new search is needed. When credits run out, the run is billed by actual input tokens plus 50,000 tokens for each search. Failed runs are refunded.

Can I call live web context from code?

Yes. Send the same question to POST https://jev-ai.pro/api/v1/web-context with a Jev AI API key. You can also send your own sources to skip the search and use evidence from your own systems.

Which questions work best?

Yes/no questions whose answer depends on recent facts: releases, launches, policy decisions, weather or research findings. Clear yes and no descriptions help Jev apply the same rule every time.