Workflows¶
After data is registered in the Catalog, it is still in its raw form. PDFs, spreadsheets, audio, and video cannot yet be retrieved directly by AI. A workflow parses, splits, vectorizes, and otherwise processes that data into a form that AI can use. A workflow is a definition of how to process a set of data: a group of operators and the order in which data flows between them.
Distinguish these three terms¶
The term “workflow” is often used interchangeably with two related concepts. Keep them separate:
Term |
Meaning |
|---|---|
Workflow |
A definition that records processing steps and their order. It is stored in the Catalog; until it is started, it takes no action. |
Run |
A single execution after the definition is started. Each start creates another workflow job. |
Canvas |
One way to edit a workflow definition. It is not the workflow itself. |
When troubleshooting, first determine whether the definition is incorrect or whether a particular run failed.
Where workflows fit¶
Workflows occupy the stage between raw and usable data:
Raw data registered in the Catalog → Workflow processing → Knowledge bases / agents
Upstream, a workflow retrieves data to process from the Catalog. Downstream, it writes results back to the Catalog and records lineage, where knowledge bases and agents can use them. A workflow does not store data itself; it performs the processing.
Three ways to create the same definition¶
Processing steps are represented as a data definition rather than a program that only engineers can modify. You can therefore produce the same workflow definition in three ways:
Use the canvas to drag operators and connect them.
Write the definition file directly.
Describe the requirement in natural language, then use the built-in AI to generate and refine the definition.
All three approaches produce the same type of definition file and run the same way. Natural language and platform-provided operators make it possible to compose processing flows for different data types: SQL operators for structured data, parsing and splitting operators for unstructured data, and custom operators for special requirements.
A workflow that starts might still produce incorrect results¶
The scheduling engine executes a definition as written; it does not understand business intent. Incorrect connections or parameter values are still run. A workflow that can be saved and started does not necessarily produce correct results. Check each workflow job to validate its output. For steps to locate a failure, see Run and debug.
Further reading¶
Architecture and core concepts: where workflows fit in the platform
Data sources and outputs: Catalog and knowledge bases
Build, run, and troubleshoot: Workflow guide
Try it first: Quickstart