Destinations

Where the data lands. The same flow can be pointed at any of these.

A SQL table

Created for you, or matched to a table you already have, in any relational connection.

You see the exact CREATE TABLE and INSERT statements before anything runs. Writing modes are append (add the rows) and replace (empty the table first).

A MongoDB collection

Nesting is kept rather than flattened. This is the inverse of what the SQL destination does with a nested record, and it is usually the right answer when the source was already document-shaped.

A file

CSV, JSON, JSONL, or a Markdown report you can hand to somebody. Useful when the point of the flow is to produce an artifact rather than to load a database.

A fine-tuning set

Validated JSONL, in the shape a fine-tuning job expects. Together with generation as a source, this makes "produce a dataset and train something smaller on it" one path in one app, rather than two tools and a script in between.

A vector index

Embedded and upserted into Pinecone, so you can retrieve over what you just brought in.

Production destinations are guarded

If the destination connection is flagged as production, the same guardrails that protect a query you type by hand apply here: destructive operations are held until you confirm them, and the run is recorded.