Skip to main content
Workflows consist of ordered stages. In Workflows, create an OCR Pipeline for document processing or Agent-Only (API) for agent interactions. Add stages with Add Stage, configure their type and settings, and save each changed stage using its own Save button. New agent-only workflows include an API Input start stage. For a complete walkthrough see your first OCR workflow and agent workflows.

​
Common controls

Stages have a name, type, and position. Review the stage order: downstream stages depend on data produced earlier. Changes to a stage remain unsaved until its Save button is used. Most stages also offer an optional Batch sync point: all documents in the batch must finish that stage before any advance to the next stage. The input stages do not offer this control.

​
Start / Ingestion

The document pipeline entry point. Choose the Input Source: API or Redis queue. For queue input, provide a queue name. In an OCR workflow, administrators can enable Folder Import to ingest files from the workflow’s shared input directory. Folder import is optional; verify permissions and monitoring before use.

​
Agent Input (API)

The fixed start stage of an Agent-Only workflow. The invocation API accepts a question and optional file. It supplies input_question, input_file_content, and input_filename when applicable. File content depends on how the input is parsed; image-capable workflows need a compatible model. See agent workflow invocation.

​
Classification

Uses a selected LLM to categorize incoming documents. Define classes with a Class Name, optional Document Type and exact version, and optional comma-separated Keywords / Expressions. Commit the keyword list by leaving the field. Define clear, distinct classes and test with samples from each category; the classification can supply the document type to a later OCR stage.

​
OCR / Extraction

Select a document type and version to define extraction fields. If a classification stage precedes OCR, you can instead select Use document type from classification result. Enable Extract Full Text when subsequent stages need full document text, for example vector indexing. Ensure the OCR provider and model support the documents being processed. Review document type configuration.

​
AI Agent

Configure three tabs:
  • Model & Prompt: select a model, enter a system prompt, and adjust temperature (default 0.2), maximum tokens (default 4096), and optional context size.
  • Tools: add only the services and actions the agent needs. Configure each tool and save the stage; see the tools catalogue.
  • Variables: select inputs such as ocr_result, input_question, or outputs from earlier agent stages, and define the JSON keys expected as outputs.
In Batch Execution, Per document (default) runs the agent independently for each document. Once for batch runs one leader and shares its result with the other documents. An empty input selection passes the whole context, and output parsing depends on the model returning valid JSON; it is not a guaranteed schema enforcement mechanism. Test inputs, tools, output keys, and permissions before production.

​
Human Verification

Stops the document pipeline for an operator to review and correct extracted results before approval and subsequent routing. This stage has no separate type-specific configuration panel. Assign authorized operators access and plan for a verification queue.

​
Export / Webhook

Choose the destination: For folder export, the displayed container path is /shared/{workflow_id}/output/{document_id}.json. Check access, network connectivity, sensitive content, and downstream behavior before enabling an export.

​
Vector DB Upload

Configure Qdrant URL, optional API key, Collection Name, Embedding Model, Chunk Size, Chunk Overlap, and the OCR fields to upload. Default chunk size and overlap are 1000 and 200. Select fields inferred from earlier OCR/document type schemas or enter fields manually when no schema is available. If full text is required, enable it in the OCR stage. Use compatible embeddings and collection configuration for indexing and querying; verify the uploaded result before relying on semantic search.

​
Generic

A general-purpose stage without a dedicated type-specific configuration panel. Use a more specific stage when you need OCR, agent processing, verification, export, or indexing; do not assume a generic stage performs those actions automatically.

​
Check a workflow before use

Save every modified stage; confirm the selected type versions, model, input/output variables, credentials, data access, and export destination. Assign users in User Config, run an authorized non-sensitive sample, inspect processing and verification, and check the final result. See Dashboard for monitoring.