Key Concepts
This page defines the building blocks of the platform. Read it once and the rest of the documentation will be easier to follow.
Workflow (also: flow)
A workflow is a connected graph of steps that turns an input into an output. Steps include AI model calls, knowledge retrieval, conditional branches, and external system calls. Workflows are authored visually in the Workflows section.
Every workflow is owned by a user, has a unique identifier, and is governed by two key fields:
- Type - categorises the workflow (e.g. assistant, ingestion, report).
- Configuration - the saved graph and node settings.
Version and publish
The committed shape of a workflow is captured as a version. Each save creates a new version; publishing marks one version as live, while drafts continue to evolve. This is how the platform supports safe iteration - a draft can be tested without affecting the published flow that callers depend on.
Run, message, and trace
When a workflow is executed it generates a run. The platform records:
- A run record with the input, output, status, and any error.
- An execution trace with the model call, prompt, latency, and token usage.
- (Optional) Asset outputs such as generated PDFs.
Statuses you will see: Pending, Running, Completed, Failed, Cancelled.
Knowledge collection and source
A knowledge collection is a logical group of documents that workflows can retrieve from. Typical patterns are one collection per product, per audience, or per language.
Each uploaded file becomes a knowledge source. Its processing lifecycle is:
- Pending - uploaded, awaiting processing.
- Processing - content is extracted, split into searchable chunks, and indexed.
- Completed - ready for retrieval.
If processing fails, the source is marked accordingly so the author can fix the issue and re-upload.
Model
A model is a language model that can be called inside a workflow. The Models page lists what is currently available in your environment; flows pick a model per node. The platform never silently falls back to an external provider - if a node references a model that is not installed, the run fails fast and tells you why.
Configuration vs. user configuration
Two concepts are easy to confuse:
- Platform configuration - global platform settings, such as dashboard layout or default workflow behaviour.
- User configuration - per-user settings and integration credentials.
Both are governed by the same role-based access control.
Compliance assessment
A compliance assessment is the result of walking through the EU AI Act risk wizard. It captures:
- Your answers to the risk-classification questions.
- The derived risk level (Minimal, Limited, High, Prohibited).
- A checklist of requirements applicable to that level, each marked complete or pending.
- A generated PDF you can download.
Assessments are individually owned and can be revisited or regenerated as a workflow evolves.
Generated document
Any artefact the platform produces (today, compliance PDFs and workflow descriptions) lands as a generated document, versioned per category and per source entity. The Documents Hub UI lists them with download and delete actions.
Roles
The platform ships with two roles:
- Manager - the typical workflow author role; can use the UI but not access advanced configuration screens.
- Developer - full access including configuration and advanced settings.
Roles are assigned through your organisation's identity provider.
Job and queue
Long-running work - workflow execution, document processing - runs through a background queue. Jobs retry automatically on transient failure. Workers process jobs independently of the user-facing request, so the UI stays responsive even when the underlying work takes minutes.
What to read next
- Functionality & Features for a deeper tour of each user-facing area.
- Compliance Check for the regulatory mapping.