Planning and building flows
Plan and assemble a document pipeline on the flow canvas.
Planning flows
Plan a pipeline with the flow editor’s essential building blocks:
-
Process files — OCR and image pre-processing.
-
Agent classifier — Categorize files with natural language definitions.
-
Filter — Route document types down parallel branches.
-
Agent extract — Extract tables and fields with the use of LLMs.
-
Combine — Merge parallel branches into one output packet.
-
Apply checkpoint — Validate fields and hold failures for human review.
User-defined functions (UDFs) extend the canvas for map, reduce, pre-flow, and post-flow logic—covered in UDFs.
Using Refiner
The Apply refiner step maps, formats, and standardizes fields from Agent extract into a validated target schema for downstream systems. Place it after extraction, then model and align the schema in the Refiner program.
Using Agent classifier
Configure an LLM-powered Agent classifier on the canvas after OCR:
-
Add the Agent classifier step and link a classifier module.
-
Replace defaults with natural language class definitions.
-
Run classification so downstream filters and extractors only process the target documents.
Using Agent extract
Extract data from unstructured documents with the Agent extract step. A common pattern is to first filter, then run Agent extract, so extraction logic (for example, passport fields) runs only on correctly classified documents. Define the extraction schema in the extract module with natural language prompts.
Using checkpoints and validations
This tutorial shows how to use validation rules as a human-in-the-loop gate. It covers configuring type checks, confidence thresholds, conditional branching, comparisons, and custom Python validation, as well as reviewing and correcting flagged fields before resuming the pipeline.