Documents are a source, not a table
A PDF has no rows to profile and no working copy to make. Rather than pretending otherwise, DataLens treats a document collection as its own kind of dataset — stored whole, indexed for retrieval, and available to AI steps.
How it works
Stored whole
The bytes are kept intact and never shredded into a table that loses the structure the document carried.
A dataset all the same
A collection appears in the catalogue beside your tabular datasets, with the same ownership and access control.
Available to AI steps
Flow nodes can retrieve from a collection, which is what makes grounded generation possible against your own material.
Governed like the rest
Access grants, ownership and lineage apply to a document collection exactly as they do to a table.
What happens after it lands
Once rows are in the catalogue, the connector that produced them stops mattering. The dataset is profiled, can be cleaned, modelled, put in a flow, governed and published exactly like every other dataset — which means switching how data arrives never means rebuilding what happens to it afterwards.
Questions about this connector
Why are documents not parsed into rows?
Because most of the value of a document is in structure that a table destroys. Collections are stored whole and retrieved from; where a document genuinely contains a table, you can extract it into a tabular dataset explicitly.
What formats are supported?
PDFs and common document formats. They are stored as uploaded rather than converted on the way in.
Can AI steps read from a collection?
Yes — that is the point of the source type. Retrieval from a collection is what grounds a generation step in your own material rather than the model’s training data.
Other ways in
CSV & Excel upload
Drag a file in and start working. Profiled on arrival, no schema declared up front.
Database
Read from your existing relational database and pull tables in as datasets.
REST API
Point at an endpoint, map the response shape, and land it as rows.
FTP & SFTP
Collect the drops that still arrive as files on a server.
CDC sync
Change data capture, so the dataset follows the source instead of ageing.
CRM
Bring customer records across without exporting them to a spreadsheet first.
Data lake
Read from — and write back to — object storage you already own.
AI model providers
Your own model credentials, held as a connection like any other source.
See it on your own data
DataLens is in private beta. Bring a file, a database or an API and work through the whole lifecycle in one sitting.
Request beta access