First version.
- Engine with four backends: Laya in-process (default,
predict_batch), sqljev gateway, laya-serve, TypeSafe Jev. - Row de-duplication, LRU answer cache, spend guard, streaming evaluation with bounded read-ahead.
- SQL Server (
jev.judge+ lookup functions;sqljev judgefor 2016–2022), Snowflake (SPCS service
functions and vectorized UDF), Databricks/Spark (pandas UDFs), BigQuery (remote functions), Redshift (Lambda),
DuckDB (Arrow UDFs), any SQLAlchemy database (CLI). - PostgreSQL and MySQL/MariaDB (answers table + lookup functions, no extension needed) and one
sqljev judge
command for SQL Server, PostgreSQL, MySQL and MariaDB; the gateway also speaks Jev's/v1/systemone, so pg-jev
runs on Laya. - Fine-tuning:
sqljev dataset,sqljev finetune(single GPU,--train-layerslow-memory mode),sqljev eval,
sqljev publish(Hugging Face Hub), and a Colab notebook (notebooks/sqljev_finetune_colab.ipynb). - Benchmark:
python -m sqljev.demo(100,000 rows, 10 domains, 13 questions with exact labels) andbench/run.py. - One-command releases:
scripts/release.sh 0.1.0. - Laya prompt: "Is it true that ?" (11/12 vs 9/12 for "Is this true of this database row: ...", on a
small labelled set of tickets and names, English checkpoint).
Install: pip install sqljev[laya,db] — database install scripts are attached below.