Question 1

How does a Jev-per-page crawl differ from an LLM scoring every page?

Accepted Answer

The expensive model is called twice per run instead of once per page: once to turn the question into four to six searches, once to write the report. Every per-page decision is a choice between known options, which is what Jev is for, so triage, page reads and link selection are all cheap typed calls. The README frames the split as the whole point, and estimates a depth-8 run at roughly a cent. Discarded pages are never expanded, so a deep run narrows into tendrils instead of exploding.

Question 2

What has to be running before this works?

Accepted Answer

Two keys and a search backend. Copy .env.example to .env and fill in TypeSafe for Jev and OpenRouter for the writing model. Web search goes through SearXNG with the JSON format enabled, and the shipped setup expects it bound to loopback on another machine, with npm run tunnel forwarding it over SSH using SEARXNG_SSH_HOST; without that a run fails on the first search with a message saying so. You can point SEARXNG_URL at any SearXNG instance with JSON enabled instead. The server itself listens on port 8790.

Question 3

Why is reasoning turned off for the writing model?

Accepted Answer

Because Qwen 3.6 is a reasoning model and left alone it spends the whole completion budget thinking and returns empty content. The README is explicit that this is not a bug in the code, it is what a finish_reason of stop with 247 of 261 tokens in the reasoning field looks like. Calls send reasoning disabled, which answers directly and costs about 20 times less, $0.0002 against $0.0043 on the planning call. Setting LLM_REASONING=on restores it.

Question 4

What stops the crawl from chasing garbage pages?

Accepted Answer

Two thresholds plus a yield table, all applied in code. TRIAGE_FLOOR at 0.15 means a search hit scoring below that on its snippet is never fetched at all, and THRESHOLD at 0.62 decides whether a read page is kept and expanded; MAX_PAGES at 120 is a hard stop. The yield table records found, kept, cut, unreachable and skipped per query, and when the queue runs dry but the run is short of its target, the planning model gets that table plus the hosts and titles that produced keepers, and is asked for new queries that borrow the vocabulary that worked.

Question 5

How do I run it on my own machine?

Accepted Answer

Run npm install, then npm run tunnel in one terminal to forward a SearXNG node to :8888, and npm start for http://localhost:8790. Copy .env.example to .env and fill in the TypeSafe key for Jev and an OpenRouter key for the writing model; every outbound call is made from the server process, not the browser.

Question 6

What do the main settings control?

Accepted Answer

THRESHOLD 0.62 keeps and expands a page at or above that relevance, TRIAGE_FLOOR 0.15 means a search hit scoring below it on its snippet is never fetched, MAX_PAGES 120 is a hard stop, MIN_SOURCES 10 is the default source target, MAX_WAVES 6 caps query batches, PER_PAGE_LINKS 3 limits children per page, and CONCURRENCY 18 sets pages in flight.

Question 7

What did a measured run cost and how did it go?

Accepted Answer

A depth-3 run with a 14-source target on 22 September 2026 crawled in 26 seconds, fetching 48 pages, keeping 17 and cutting 30, using 63 Jev calls at 438 ms average and 127k tokens for $0.005. The report added 81 seconds and 17 cited sources, for a $0.039 total.

Jev Deepresearch

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community