Quivr V2
Quivr V2 combines durable stream ingestion with natural-language article alerts judged by an optional Jev classifier.
Added to Jevfast
How it uses Jev
Quivr's first-party alerts subscription can send a new article's title, source, mapped metadata, and configured text parts to Jev. It asks one Noul question for each distinct plain-language alert description. Quivr treats Jev's yes probability at or above the configured threshold (0.5 by default) as a match; a classifier failure remains unavailable rather than becoming a negative match.
What Jev decides
Illustrative described-alert requestExample answers · not a recorded Jev response · Source ↗
Question 1 · q0
YOUR APP
INSTRUCTION
Judge whether the article is about the alert description, allowing different wording or language.
STATE
One article is represented by title, source and mapped metadata, plus configured text parts. Each distinct alert description becomes a separate question. This is a source-derived request example, not an observed answer.
JEV · NOUL
YesNo
App workflow
- Receive a searchable article
The alert rule evaluates an article and its available Parts and metadata.
- Build the classifier state
The plug-in selects configured text roles, builds article title, source and mapped metadata, and bounds the text sent in state.
- Ask Jev about descriptions
The classifier sends the article once with one Noul per distinct description. Large requests are divided into size-limited calls.
- Apply the threshold
The plug-in compares each returned yes probability with the subscription threshold and emits a match at or above it.
Why it is interesting
A subscription plug-in turns natural-language alert descriptions into repeatable, per-article decisions. The same article state can be shared across several alert questions, and operators can select Jev or a local vector path for supported alert kinds.