Building a Workflow

Create a workflow

Open Workflows under Build and click New Workflow. Pick a starting point:

Option	What you get
Blank Template	An Entry point, one LLM Call and an End node, already connected.
Custom Evaluator	An Entry point with a `question` input, one LLM judge node and an End node with the fixed evaluator results. See Workflow as evaluator.
From Export	A workflow imported from a JSON file that Export Workflow produced on another project.

Give the workflow a name, an icon and a description, then click Create Workflow. The editor opens.

The Create new workflow dialog with Blank Template, Custom Evaluator and From Export

The editor

The Components panel on the left lists the node types you can drag onto the canvas. The canvas holds the nodes and the edges between them. Click a node to open its settings in a drawer on the right.

Component	What it does
Prompt	Calls a model with a prompt. The node is named LLM Call on the canvas.
Code	Runs a Python class you write.
HTTP Call	Sends an HTTP request to an external API and reads the response.
If/Else	Routes execution to a `true` or a `false` branch from a condition over its inputs. Nodes on the branch that is not taken are skipped.
Agent	Calls one of the agents of the project (HTTP, code or workflow agent).
Evaluator	Runs one of the evaluators of the project and outputs `passed`, `score` or `label`.
Custom Components	Workflows of the project that were published as components. They appear below the built-in components.

Every node lists its inputs and outputs with a type: str, float, bool, image, list[str], dict, json_schema or chat_messages. Drag from an output handle on the right side of a node to an input handle on the left side of another node to connect them. In a node drawer, the mapping field next to an input does the same thing: pick the node and the output the input reads from. The top bar shows the workflow name, undo and redo, the version history, Evaluate and Publish. The Results button at the bottom opens the evaluation runs of the workflow.

Entry point

The Entry point declares the inputs of the workflow. Every other node reads from these inputs or from the outputs of an earlier node.

Inputs: add one field per value the workflow receives. Each field has a type and an optional default value.
Attached Dataset: a new workflow starts with a draft dataset. Open edits its rows and columns, Replace attaches another dataset of the project, and the cross detaches it. Attaching a dataset adds its columns to the inputs. Save as dataset in the dataset editor stores the draft under Datasets.
Workflow Outputs: Go to end node opens the End node, which declares what the workflow returns.

An input that no dataset column provides is a parameter of the workflow. Runs from the API and from Agent Testing set it per run.

LLM node

The LLM node calls a model with a prompt and returns the fields you declare as outputs.

The LLM Call drawer with the model, the system prompt, the user message and the variable mapping

Model: the picker at the top selects the model. The list holds the model providers configured for the project.
Prompt: the view switch next to the section title shows either the system prompt alone (Prompt) or the full message list (Messages). Write {{name}} in a message to insert a variable. A variable used in a message but not declared shows an Undefined variables notice with a Create action.
Variables: one row per input of the node, each with a type and a mapping to the source it reads from, for example Entry.question.
Outputs: one row per field the model returns. With one output of type Text the node returns the model answer as is. With several outputs, or an output of another type, the node asks the model for structured output with those fields.

Apply keeps the changes on this node. Save stores the prompt in the prompt library as a new version, so other workflows and your code can use it. An orange dot on the node marks prompt changes that are applied but not saved. See Prompts in workflows.

Code node

The Code node runs a Python class. The __call__ method receives one keyword argument per input and returns a dictionary with one key per output.

The Code drawer with the Python source, one input mapped to the LLM answer and two outputs

class Code:
    def __call__(self, answer: str = None):
        length = len(answer or "")
        return {"within_limit": length <= 300, "length": length}

Inputs and Outputs work as on the LLM node. Adding an input also adds the parameter to the __call__ signature. Every parameter defaults to None, so an input that is not connected does not raise an error.
Click the source to open the code editor. Save keeps the editor open, Save & Close returns to the drawer.
Secrets: the editor lists the secrets of the project and inserts secrets.NAME at the cursor. The value reaches the code at run time only; the editor and the saved workflow hold the name. Manage the values under Settings, Secrets, described in Secrets.

The Python code editor with the Secrets button and Save and Close

Evaluator node

The Evaluator node runs an evaluator of the project on the values you map to it and outputs passed, score or label, depending on the evaluator type. Dragging Evaluator onto the canvas opens the evaluator picker. Pick an existing evaluator, or click New Evaluator and choose a type from the categories (Expected Answer, LLM as Judge, RAG Quality, Quality Aspects, Safety, Custom). The evaluator is saved to the project and the node references it.

The evaluator drawer with the name, the model, the judge prompt and the variable mappings

The drawer shows the settings of the evaluator type, for example the model and the prompt of an LLM judge.
Variables: map input, output, expected_output or contexts to the entry inputs or to the outputs of earlier nodes. Only the fields the evaluator uses are required.
Apply keeps the changes on this node only. Save updates the saved evaluator for every place that uses it.

Also check: Built-in evaluators for the list of evaluator types and their fields.

End node

The End node declares the results of the workflow. Add one result per value the workflow returns, with its type, and connect an output of an earlier node to each one.

The End drawer with four results: answer, within_limit, length and passed

In a workflow published as an evaluator, the results are fixed to details, passed, score and label. See Workflow as evaluator.

Run one input

The play button on a node opens two options:

Run with manual input runs this node alone. A form asks for a value per input of the node.
Run workflow until here runs this node and everything it depends on. The Run until here dialog shows one field per Entry point input, prefilled with the first dataset row or with the values of your last run. Select dataset value picks a row of the attached dataset instead.

The Run until here dialog with the question field prefilled from the dataset

Use it on the End node to run the whole workflow. Each node shows a check mark when it finishes or an error mark when it fails.

Read the result of a node

Click a node after a run to see what it received and what it produced. The expanded view shows the inputs on the left, the settings in the middle and the outputs on the right, with the duration of the run. Full Trace opens the trace of the run in the Trace Explorer, with every model call, its cost and its messages.

The End node after a run, with the answer, length, passed and within_limit outputs

Evaluate on a dataset

Evaluate runs the workflow on every row of the attached dataset and records the outputs of the evaluator nodes. The dialog asks for a description of the version it saves, and for the rows to evaluate: Full dataset, Test entries, Train entries or Specific entry.

The Evaluate Workflow dialog with the version description and the Full dataset option

The Results panel at the bottom lists the runs of the workflow. Select a run to see the pass rate of each evaluator, the total cost and the runtime. Open full results opens the run under Experiments, with the row-by-row results. Run via API shows the code that starts the same evaluation from a script or a CI job, in Python, TypeScript or shell.

The Results panel with one run, its pass rate, total cost and runtime

Also check: Experiments for the results page and comparisons between runs.

Versions

The editor autosaves every change. The history button in the top bar opens Workflow Versions: type what changed and click Save new version to store a numbered version, or restore an earlier one from Previous Versions. Evaluate and Publish save a version when the workflow changed since the last one.

Publish

Publish opens a menu with three actions:

Action	What it does
Publish Workflow	Saves a version and marks it as the published one. This is the version the API, the evaluator lists and the Components panel of other workflows use. A workflow created as a custom evaluator becomes an evaluator of the project when you publish it.
View API Reference	Shows how to run the published version through the Workflows API. Available after the first publish.
Export Workflow	Downloads the workflow as JSON, with the attached dataset. Import it on another project with From Export.

The Publish menu with Publish Workflow, View API Reference and Export Workflow

Copy a workflow to another project

On the Workflows page, the menu of a workflow card offers Replicate to another project. Pick the target project and choose whether the attached dataset is copied with it.

Get Started

Agent Testing

Observability

Evaluations

Prompt Management

Platform

Examples & Cookbooks

Help

Building a Workflow

Create a workflow

The editor

Entry point

LLM node

Code node

Evaluator node

End node

Run one input

Read the result of a node

Evaluate on a dataset

Versions

Publish

Copy a workflow to another project

​Create a workflow

​The editor

​Entry point

​LLM node

​Code node

​Evaluator node

​End node

​Run one input

​Read the result of a node

​Evaluate on a dataset

​Versions

​Publish

​Copy a workflow to another project

Create a workflow

The editor

Entry point

LLM node

Code node

Evaluator node

End node

Run one input

Read the result of a node

Evaluate on a dataset

Versions

Publish

Copy a workflow to another project