ReviewFast
Free · open source · runs in your browser

Find the studies that matter first, and know when you have found enough.

ReviewFast helps with title and abstract screening in systematic reviews. A decision model, a model that answers a yes-or-no question about each record with a probability instead of generating text, ranks your search results against your eligibility criteria. You screen from the top, and a statistical stopping criterion tells you when at least 95% of the relevant records have been found, with 95% confidence.

Try the demo (no key needed) Start screening your records Use it in Python or with an AI assistant

How it works

1

Rank

Jev, a decision model, gives every record a probability of meeting your criteria. Records go ten per request with the evaluation's prompt and request format; ReviewFast shuffles their order, which the evaluation did not test. Scoring costs about US$0.02 per 1,000 records.

2

Screen

You decide include, maybe or exclude, one record at a time, from the most to the least likely. The probability stays hidden by default so that it does not anchor your judgement.

3

Stop with confidence

After each decision, the stopping criterion of Callaghan and Müller-Hansen tests whether recall could still be below 95%. When it can be rejected at the 5% level, you can stop and report the rule. A random-sample check of the records you skipped is built in.

What the evaluation showed

The workflow comes from an evaluation of a decision model for screening (manuscript in preparation). Ranking with Jev followed by the statistical stopping criterion was a post hoc analysis in that study.

51 / 51
reviews reached 95% recall (23 held-out SYNERGY reviews and 28 Cochrane reviews)
87.6%
of records read on average in the SYNERGY reviews (86.1% in the Cochrane reviews)
0.966
mean AUC of Jev ranking against 0.831 for the best free ranker
$0.02
per 1,000 records scored, at the list price during the study

Savings concentrate in larger reviews; in small reviews the criterion needs almost every record before it can stop. The full workflow was evaluated with one decision model, Jev. A fixed probability threshold was not reliable as a stand-alone stopping rule, so ReviewFast does not offer one.

Good to know

Before you start

  • You need a Vercel AI Gateway API key with paid credits. Jev is not available on the free tier.
  • Use the eligibility criteria from your protocol, written before screening.
  • Export complete, unsorted search results (RIS, CSV, PubMed format or PubMed XML).
  • Records without an abstract, or apparently not in English, go to a manual queue that you screen in full.

Your data

  • Everything runs in your browser. jolab.ai serves the page and receives none of your records, decisions or keys.
  • Criteria, titles and abstracts go from your browser to TypeSafe AI through the Vercel AI Gateway (United States). The provider does not offer zero data retention, so send only published or cleared records.
  • Your key stays in memory for the session. Projects are saved in this browser; export them to keep a copy.