Skip to main content
This node is in preview. Its behavior may change in future releases.
The Filter node narrows a document to the pages you care about. It reads the page text produced by an upstream parse step, keeps or drops pages by keyword, page range or a plain-language description, and passes the surviving pages on as a single document.
Documents read as text (office and CSV) arrive as a single logical page, so a page range can only ever select page 1 or nothing, and a keyword rule matches the whole document at once. To narrow one of those, filter on its content rather than its pages. See supported formats.
Filter never fans a run out. One document goes in, one (smaller) document comes out. To cut one document into several sub-documents, use split.

​
When to use Filter

  • You want to drop boilerplate before extraction: cover sheets, terms and conditions, blank trailing pages.
  • You want to keep only the pages that matter, so the model sees less noise and extraction is more accurate.
  • You know the pages by number or range (1-3, 2,5, 4-end) or by a keyword that appears on them.
  • You can describe the pages in plain words (“cover sheets and standard terms”) and let a built-in check judge each page.
  • Use split instead when one upload holds several documents and each needs its own run.
  • Use classify instead when the decision is which workflow a whole document belongs to. Filter narrows one document to its useful pages.

​
Configuration

​
Modes

To classify documents by content type, use the dedicated Classify action node.

​
Keep

Passes only the pages that match the filter. Every other page is removed from the document for the rest of the run.

​
Drop

The inverse of Keep. Removes the pages that match the filter and passes everything else through. With a Document filter, a page the check is unsure about is always kept, in both modes: losing a page is unrecoverable downstream, keeping an extra one is not. If the check is unavailable the step fails so the run can be retried, and a failed step is not charged. The step fails when no pages survive the filter, so a filter that matches nothing is reported rather than silently producing an empty document.

​
Output

A single page group containing the surviving pages. Downstream nodes see only those pages: extract reads just the pages you kept, and page numbers stay the ones from the original document.
Filtering does not reduce credits. Per-page cost is charged against the document’s full page count, and that count narrows only inside a child run, which Filter never creates. Filter buys accuracy, not cost.

​
Inputs and outputs

Allowed inputs: parse, review, split, and another filter node. Trigger nodes are deliberately not allowed. Filtering needs page text, which only a parse step produces: a scanned document arrives at the trigger with no text at all. Output: the filtered document, passed to the next node.

​
Credits

The 0.5-credit tier applies when the node has to read text to decide, by keyword or by description. Page-number filters stay free. See credits for the full per-node pricing breakdown.

Split action

Cut one upload into several sub-documents

Parse action

Produce the page text Filter matches on

Extract action

Extract data from the pages you kept

Conditional routing guide

Branch a run with output ports