Metadata extraction on weakly-typed DBs for agents

What Phinest built

A metadata agent generates table and column descriptions, business meaning, data distribution statistics and relationships. It works from the schema, or examines the data itself when schema types provide too little information.

Profiling collects null counts, distinct counts and common values. The workflow checks candidate keys and joins against the data and records assumptions for human review. Expert corrections feed back into the affected metadata.

What it enables

Downstream agents receive context about what each row represents, which tables relate and where interpretation needs care. That context helps them select data and write queries, while reviewers can inspect the assumptions behind it.