GAARD - Governed AI Access to Relational Data

GAARD is a self-hosted AI SQL Gateway for governed natural-language access to relational data.

GAARD allows applications and users to ask questions about relational databases using natural language while keeping SQL generation, validation, execution, prompts, connectors, and auditability under control.

For more informacion see https://github.com/pkroliszewski/gaard

This package

gaard-api provides the GAARD FastAPI backend and bundled admin application.

After installation, start it with:

gaard-core install-example-database
gaard-core start

gaard-core install-example-database creates the bundled Medical POC SQLite database at examples/medical-poc/demo.db in the current working directory, registers it as the active default datasource in metadata.db, and matches the default datasource URL. Use gaard-core install-example-database --output /path/to/demo.db to place it elsewhere; the command prints the SQLite datasource URL saved in metadata.

gaard-core start accepts --host, --port, and --reload. By default the API is available at http://localhost:8000 and the admin application at http://localhost:8000/admin.

gaard-api start is an alias. gaard admin remains available for compatibility.

Conversation context

With context_mode: "auto", each question stores a new_topic or follow_up decision. The LLM receives all turns since the last new_topic, including that question. For a follow-up, a separate LLM call compresses this context together with the current question into one standalone sentence. Both calls use temperature zero. The resulting sentence is saved before SQL or Analysis runs; a new topic uses the current question as its context.

GET /api/v1/conversations/{conversation_id}/turns/{turn_id}/context returns the saved execution-time snapshot as JSON with context, context_decision, conversation_id, and turn_id. The IDs are available in the query response's metadata.conversation object. This authenticated endpoint only allows the conversation owner and never regenerates the context from later messages.

context_mode: "new" starts a new conversation; "off" remains stateless. Mock mode does not perform semantic context classification.