Parse-Flow: Open-Source Visual Document Intelligence Workflow Designer

Unstructured documents, which are prevalent in business operations, pose challenges for downstream systems that require structured, machine-readable data, leading to the necessity of document intelligence to transform these documents effectively. Parse-Flow is an open-source project designed to tackle this challenge by focusing on four document processing primitives—Parsing, Extraction, Classification, and Splitting—within a visual workflow designer. The system leverages a React frontend, a Bun server, a Python worker, Redis, and Postgres to create a seamless and efficient workflow, with the Bun server distributing tasks to the Python worker, which processes them and returns results via Redis and Postgres. The project emphasizes a narrow workflow vocabulary supported by the LlamaParse Platform, allowing for versatile compositions of document processing tasks while ensuring transitions are validated and observable. The backend operates on a LlamaAgent workflow, which interprets user-defined processes in real-time and maintains a robust state management system, ensuring each step of the workflow is transparent and auditable. By focusing on comprehensive document intelligence, Parse-Flow provides a durable solution to common pitfalls in document processing, such as misclassification or extraction errors, highlighting the importance of composable, validated, and observable workflows.