ContextGem Jev Extraction
Extract document aspects and closed concept values with Jev, while retaining probabilities and thresholds.
Added to Jevfast
How it uses Jev
ContextGem's DocumentDecisionModel selects the typesafe/jev-latest provider through LiteLLM. It builds closed predicates for document aspects and Boolean concepts, plus choices for single-label concepts. It evaluates aspect predicates per paragraph, and concept questions against the relevant document text. The library applies thresholds and retains answer probabilities on extracted items.
What Jev decides
Illustrative notice-form concept ChoiceExample answers · not a recorded Jev response · Source ↗
Question 1 · C1
YOUR APP
INSTRUCTION
Choose the required form of termination notice.
STATE
The repository example supplies a service agreement stating that either party may terminate for any reason with 30 days of written notice, governed by New York law. This chart shows the notice-form concept question against that document; paragraph-aspect predicates are evaluated separately.
JEV · CHOICE
- Written
- Any form
- Not specified
App workflow
- Describe the document
Create a Document from the text, then declare the aspects and concepts to extract with their descriptions, labels, and classification settings.
- Configure the decision model
Select typesafe/jev-latest and provide the TypeSafe API key through CONTEXTGEM_TYPESAFE_API_KEY.
- Build and send questions
DocumentDecisionModel.extract_all builds a predicate per paragraph for each aspect and questions for supported concepts, then submits the closed questions through the configured decision model.
- Apply answers
The extraction uses probabilities and the decision threshold to include qualifying paragraphs and concept values. The returned items expose their probabilities.
- Read extracted items
Retrieve aspect and concept items by name from the returned document.
Why it is interesting
It separates paragraph selection from closed concept extraction and retains probabilities for each extracted item. The same document interface can use a decision model for closed questions and a language model for open-ended extraction.