AI assistant: compact YAML context, safer INSERTs and a New chat button

The query console's AI assistant now receives your database as YAML instead of JSON. The schema it is given maps each table to its columns and their types, and the rows of your last query arrive as a YAML list of column-and-value pairs rather than a text table. Models read YAML more reliably, and the same schema takes far fewer tokens. That matters most on large legacy databases, where the schema is most of every prompt. The MCP tools that AI clients call also return their data as YAML: list tables, describe table, procedures, schema graph, read query and the MongoDB tools.

The assistant now knows which columns are keys. Primary-key, unique and auto-increment columns are marked in the schema it receives, and it follows new rules when it writes INSERT statements. It leaves auto-generated columns for the database to fill. For a numeric key, it writes one statement per row and takes the next value from the data already in the table, instead of guessing 1, 2, 3. Asking for five sample INSERTs no longer produces rows that clash with the ones you already have.

Less context, better focus: only your 10 most recent messages are sent with each question. A long conversation no longer makes every prompt bigger, and old answers stop steering new ones.

A New chat button in the assistant's header clears the conversation so you can start over. The prompt suggestions and sample queries come back, as they are when you first open the chat. The chat also works better from the keyboard: Tab now moves around inside the open chat instead of leaving it, and Esc closes it.