schemagate

schemagate is an Apache-2.0 Python library for the schema-selection step of text-to-SQL and NL2SQL systems.

Most stacks let the model see the whole schema, or a retrieved subset of it, and apply access control later when the query runs, via row-level security or a permissions layer. That protects the data but not the answer: the model writes valid SQL against a table the caller cannot read, the filter strips every row, and the user is told "no records found" — which is false and indistinguishable from the table being empty.

schemagate decides what the model is shown per caller, before any SQL exists. A restricted table is absent from the prompt rather than ranked lower or filtered afterwards. Table-level and column-level restrictions are both supported, so a table can be visible with a column withheld.

It is not a platform or a chat UI. It replaces the schema-selection step; you keep whatever writes your SQL.

• Reflects any SQLAlchemy database (PostgreSQL, Oracle, SQLite, MySQL, SQL Server)
• Fielded BM25 plus vector hybrid ranking; no API key required
• Writes nothing to your database: no embeddings table, no schema changes
• Ships an MCP server, a LangChain retriever, a CLI and a browser Studio
• Optional local sentence-transformer embedder; runs fully offline

Live-tested on PostgreSQL 16, Oracle 26ai and SQLite. SQL Server and MySQL reflect but have not yet been run against a live instance.

pip install schemagate