Flow editor

Flow steps

Enterprise Single-tenant

Combine flow steps to create your document processing pipeline. Most steps link to a module, a folder containing the artifacts which support the step’s actions, such as schema files, validations files, custom function files (also called user-defined functions, or UDFs), and refiner programs.

Importing a module into a flow copies the module code into that flow, so any future edits you make to the original module aren’t reflected. Importing is useful for reuse, but doesn’t support syncing.

Map steps

Use map steps to define processing steps in your pipeline.

StepPurposeLinked module
Agent classifierClassifies documents into predefined classes using LLMs. The platform handles the LLM call and grounds the result in the document content. This step is recommended for most classification needs.Agent classifier, linked to a classification schema JSON file
Agent extractExtracts structured data from documents using LLMs. The platform handles the LLM call and ensures extracted values are grounded in the source document. This step is recommended for most extraction needs.Agent extract, linked to an extraction schema JSON file.
Apply checkpointEvaluates extracted and refined data against validation rules. Files failing validation are routed to human review. Multiple checkpoints can be used in a single flow for staged review.Checkpoint, linked to a validations configuration.
Apply classifierRuns rule-based split classification using a custom Python classifier. Use for code-driven, deterministic class assignments, rule-based routing, or cases where external context is needed before classification.Classifier, linked to a classifier file.
Apply refinerPost-processes extracted data using refiner logic. Used to clean, transform, standardize, and enrich output before it reaches validation or downstream systems.Refiner
Map recordsSpecify how multipage documents are parsed into separate records.
Map UDFApplies a UDF to each record individually. Used for per-document transformations, lookups, or custom processing logic.UDF
Process caseRun a case program against a packet of input files for packet processing, also called case management, and populate cross-class fields using a refiner program.refiner
Process filesDigitizes input documents using OCR. This is typically the first step in any flow, converting raw files (PDF, TIFF, images) into machine-readable text that downstream steps can process.Reader

The following steps generally aren’t required when developing new flows and advanced apps. They remain available in the UI to support existing apps.

  • Apply redactor to refined fields
  • Doc gen
  • Run extraction model
  • Unified extractor

Filter step

The filter step lets you filter results based on defined parameters.

Reduce steps

Use reduce steps to combine streams in your pipeline, reducing the output.

StepPurposeLinked module
CombineCombine branches into a single flow output.
Reduce UDFApplies a UDF across all records in a batch. Used for aggregations, cross-document comparisons, or batch-level output generation. See Custom functions in flow for details.UDF