MinusX

Converts natural language questions into executable SQL queries by parsing the connected Metabase instance's database schema, table relationships, and metadata. The system maps user intent to appropriate tables and columns, handles JOIN logic automatically, and generates dialect-specific SQL that executes directly against the underlying database. This approach avoids hallucinated table names by grounding queries in the actual schema available in Metabase.

Maintains conversation context across multiple turns, allowing users to ask follow-up questions that reference previous queries and results. The system tracks query history, understands implicit references ('drill down into that', 'show me the top 5'), and regenerates SQL with accumulated context. This enables natural dialogue-based data exploration without requiring users to restate full context with each question.

Operates as a native Metabase plugin or embedded interface that intercepts natural language input and returns results directly within the Metabase dashboard/query builder UI. The integration likely uses Metabase's plugin architecture or API to execute queries and render results in the native format, avoiding context-switching to external tools. Results appear as native Metabase visualizations (tables, charts, etc.) rather than raw text.

When a generated SQL query fails (syntax error, missing table, permission denied), the system captures the database error message, explains the issue in natural language, and regenerates a corrected query. This creates a feedback loop where the AI learns from execution failures within the conversation. The system likely sends error messages back to the LLM as context for the next generation attempt.

Leverages Metabase's semantic layer (custom expressions, field descriptions, table relationships, saved questions) to understand business context beyond raw schema. The system reads Metabase metadata like field descriptions, custom metrics, and relationship definitions to map natural language business terms to actual columns. For example, 'revenue' might map to a custom expression in Metabase rather than a raw column, improving semantic accuracy.

After executing a query and retrieving results, the system generates natural language explanations of what the data shows, highlights notable patterns or anomalies, and provides business context. This transforms raw query results into actionable insights without requiring users to interpret numbers themselves. The explanation is generated by the LLM based on the result set and original question.