JEV + ModernBERT Search

Create a Python 3.12 virtual environment, activate it, upgrade pip, then run pip install -e ".[dev]". The package exposes a jev-search command-line entry point. Python 3.12 or newer is required because jevclient needs it.

Hybrid search mode requires one. Set the JEV_API_KEY environment variable with a TypeSafe key. Without a key you can still run retrieval using the --embedding-only flag on the search commands.

The question-pair sample comes from Kaggle's Quora Question Pairs competition, so you authenticate with Kaggle and accept the competition rules. The PIM demo needs a Neo4j 5.x database, configured through NEO4J_URI, NEO4J_USERNAME, and NEO4J_PASSWORD.

In hybrid mode the query and the retrieved candidates are sent, plus Neo4j ontology context for PIM searches. The documentation advises reviewing the provider's terms and usage costs before sending catalog or search data.

ModernBERT embeds questions or product text and retrieves the closest candidates by cosine similarity. A separate Jev API service then evaluates candidate relevance with typed yes/no questions and reranks the candidates.

Yes. Both the question search and the PIM search commands accept --embedding-only, which returns the ModernBERT retrieval baseline and, for PIM, still looks up candidates in the Neo4j-backed catalog.

The prepare command deterministically samples 10,000 Quora rows by default, and PIM import and search default to a 10,000-product limit; raise either explicitly for larger catalogs. Datasets and embedding indexes are excluded from Git.