For AI agents: a documentation index is available at the root level at /llms.txt and /llms-full.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
BlogLog InRequest Demo
HomeProductDevelopersSelf-HostingChangelog
HomeProductDevelopersSelf-HostingChangelog
  • Getting Started
    • Overview
  • Agent Builder
    • Using the Agent Builder
  • Prompts
    • Prompt Engineering
    • Collaboration
    • Custom Models
    • Multimodality
    • Prompt Caching
  • Workflows
    • Introduction
    • Experimenting
    • Integrating
    • Function Calling
  • Evaluation & Test Suites
    • Quantitative Evaluation
    • Evaluating RAG Pipelines
    • Online Evaluations
  • Metrics
    • Out of the Box Metrics
    • Custom Metrics
    • Reusing Metrics in Test Suites
  • Deployments
    • Deployment Lifecycle Management
    • Observability in Production
    • Environments
    • Release Tags
    • Release Reviews
  • Monitoring
    • Monitoring Production Trends
    • Track Workflow Execution Costs
    • Datadog Integration
    • Webhook Integration
    • Execution URLs
  • Documents
    • Uploading Documents
    • Integrating w/ Search API
    • Metadata Filtering
  • Security
    • Data Privacy and Storage
    • HMAC Authentication
    • Role-Based Access Control (RBAC)
    • Static IPs
  • Organizations
    • Manage Organization Access
    • Data Retention Policies
LogoLogo
BlogLog InRequest Demo
On this page
  • Quick Links
  • Concepts
  • Input Variables
  • Scenarios
  • Nodes
  • Edges
  • Final Output
Workflows

Build multi-step AI apps with Vellum’s Worfklows

Was this page helpful?
Previous

Streamline AI App Development with Vellum's Workflows

Next
Built with

Vellum Workflows help you quickly prototype, deploy, version, and monitor complex chains of LLM calls and the business logic that tie them together.

It provides a low-code interface for defining these chains so that you get rapid feedback on how they work across a variety of test cases that you define. Once you’re happy with the Workflow, you can “deploy” it and hit an API to invoke that Workflow from your application.

Once deployed, future changes to the Workflow definition are versioned and invocations made from your application are logged. For a given invocation, you can view the inputs, outputs, and latency of each step along the way.

Quick Links

Workflows Nodes
Learn about the different types of Nodes you can use to build Workflows.
Use Case Directory
Browse real-world use cases and reference architectures, from simple chatbots to complex multi-agent systems.
Example Workflows
See example Workflows that you can use as a starting point for your own applications.
Common Architectures
See frequently used architectures that you can use within your Workflows.

Concepts

Workflows make heavy use of the following concepts:

  1. Input Variables
  2. Scenarios
  3. Nodes
  4. Edges
  5. Final Outputs

Let’s take a look at each

Input Variables

The behavior of most Workflows depend on 1 or more dynamic Input. For example, you could define a single Input named query that your Workflow depends on.

Workflow Input Variables

Scenarios

A Scenario is a set of values for your Input Variables. In the above example, we have Scenario 1 which assigns a value of What is fine tuning? to the query Input Variable.

You can define as many Scenarios as you want and swap between them to test that your Workflows behaves the way you expect for each.

Nodes

Nodes are the steps in your Workflow where some action will take place. Some Nodes generate Outputs, whereas some Nodes are used purely to direct the flow of execution.

For example, the Prompt Node is used to pass Input Variables into a Prompt and execute an LLM. It generates an output that can then be used as an input to other downstream Nodes.

Workflow Nodes

Edges

Edges connect Nodes and define the order in which they are executed. The are represented as the lines in between Nodes.

Workflow Edges

Note that a Node has access to the output data from all upstream Nodes, not just the Node(s) that it’s directly connected to via an Edge.

Final Output

There’s a special Node called a “Final Output Node.” They’re used to indicate which Node output you actually care about and want to surface as the overall output for the Workflow.

In the below example, I have a Final Output Node named final-output that subscribes to a string output that comes from the OpenAI Help Center Prompt Node.

Workflow Final Output

Final Output Nodes are particularly important when you Deploy a Workflow and invoke it via API. By default, only the data that Final Output Nodes subscribe to will be returned by the API.

Note that you can have as many Final Output Nodes in a Workflow and can assign each a name to differentiate the data associated with each in API calls.