Hardcoding schema into NL2SQL prompts fails when column names are opaque, misleading, or near-duplicate — a common reality in enterprise data. This project connects external data catalogs (DataHub, OpenMetadata) to Couchbase so that column descriptions, not just names, inform SQL++ generation. Column descriptions are embedded as 384-dimension vectors and stored in Couchbase; at query time, vector search retrieves semantically relevant columns, which are then passed to a Claude model on AWS Bedrock to generate correct SQL++. A pluggable provider pattern means switching catalogs is a config change. Two concrete examples show how schema-only approaches silently return wrong results — correct SQL that filters on invented string values or picks the wrong near-duplicate column — while catalog-enhanced queries get the right answer because the encoding or tolerance clause is in the description.
