A Transformer node reshapes data from a single upstream source. You describe what you want done; the AI assistant generates the operation.
You choose where the transformed data should live:
Generated and stored inside the same database the source comes from.
Pulled from the source and stored in DagFlux's local workspace. This is automatically the only option when the source is already local.
When the storage location is local, the operation mode is forced to "new table" — you can't perform an in-place update on a remote database while storing the result locally.
You also choose how the result relates to the source:
Create a brand-new table or collection holding the transformed data. The source is left untouched.
Modify the source itself by updating, deleting, or altering rows or documents.
The chat only asks about storage location when the operation mode is "new table" — for in-place updates the question is skipped because the result has to live where the source lives.
The assistant uses a different prompt depending on the source's database type. Each prompt knows the dialect and idioms of that database and produces operations that follow them. The assistant first asks any clarifying questions it needs (which mode, which target name, which storage location), then generates the operation.
The chat applies database-specific identifier quoting automatically (schema.table for PostgreSQL, square brackets for SQL Server, dataset.table for BigQuery, etc.) — you don't have to think about the dialect.
The shape of the generated operation depends on the source:
CREATE TABLE, an
INSERT INTO ... SELECT, and a preview query.
SELECT to fetch and transform rows,
plus a local CREATE TABLE and parameterised insert that the local workspace will run.
ALTER TABLE, an UPDATE /
DELETE, or both. Filtering intent like "keep only X" becomes a delete of the
records that don't match.
CREATE TABLE,
INSERT, UPDATE, DELETE, or ALTER TABLE statements.
If the request is a better fit for a skill (Python, JavaScript, vector operations) than for vanilla SQL or NoSQL, the AI assistant offers to route the transformation through that skill instead. The skill generates and runs its own logic and writes the result to the chosen destination.
After the assistant proposes an operation, you can refine it by sending a follow-up message ("also exclude records older than 2020", "use a left join instead", "use the customer name column not the email"). The assistant regenerates without losing the previous context, so the conversation keeps building toward the final shape you want.
Generated operations may include placeholders like {{current_date}} and
{{current_timestamp}} that DagFlux substitutes with the actual runtime values at
execution time. This makes it easy to write transformations like "set archived_date to today"
once and have them work for every future run.
The chat shows the generated operation (SQL, document operation, or JS code) along with a brief explanation in plain language and a preview query so you can confirm the result before saving the node.
