CSV and Excel
Upload a spreadsheet and query it directly, with no database connection and no SQL approval step.
Files are queried through DuckDB rather than your database, which means two things: you need no connection to get an answer, and there is no SQL approval step, because nothing runs against your infrastructure.
Upload
Attach a file from the composer, or drag it onto the conversation. Multi-sheet workbooks expand into one table per sheet.
Uploads are content-hashed, so re-uploading a file you already have reuses the existing copy instead of storing it twice.
The file library
Past uploads stay available. From the library you can re-attach a file to a new conversation, peek at its inferred schema before asking anything, or share it with your organization so teammates can query the same numbers.
Hybrid: a file plus a database
Attaching a file to a chat that already has a database connection puts both sources in one thread, so a question can span them, joining a spreadsheet of targets against live rows, for example. Cross-source joins are executed in Python rather than pushed into either engine.
Limits
Uploads are capped per file and per conversation. A very wide workbook is usually better loaded into a real database and connected properly: file mode is for the analysis you'd otherwise do in a spreadsheet.