Node Types
Subtask, Review, Logic Gate, Start, and End nodes — behavior, configuration, and connection rules for every node type.
Each node type in the Pipeline Builder serves a distinct purpose. This page covers the behavior, configuration, and connection rules for every node type.
Start node
The entry point of every pipeline. Tasks begin their journey here.
- Configuration: None — the start node is automatic and not editable.
- Rules: Every pipeline must have exactly one start node. The start node can only connect to subtask nodes — it cannot connect directly to review nodes, logic gates, or the end node.
End node
The completion point of a pipeline. When all tracked nodes in a task have completed, the task is marked Finished.
- Configuration: None.
- Rules: Every pipeline must have exactly one end node. Any node type (subtask, review, or logic gate) can connect to the end node.
Subtask node
A work step that can collect human input, serve predefined data, or generate LLM responses. Subtask nodes have four tabs: Fields (form builder), Evaluators, Preview, and Settings.
Fields tab (form builder)
Design the contributor form with 12 field types: Short Text, Long Text, Code, Boolean, Numeric, Datetime, Select, Rating, Ranking, Pairwise, Criteria, and File Upload. See the Form Builder page for full details.
Additional options on the Fields tab:
- Grid Layout — toggle at the top to enable a grid layout for the form. When on, fields can have custom widths.
- Add Group — create field groups that batch multiple fields together for shared LLM processing. Fields in a group share a model, temperature, context sources, and tool settings. Groups are a container concept, not a field type.
- Upstream Fields — control visibility and editability of fields from upstream nodes. Useful for showing prior work as read-only context.
Evaluators tab
The Evaluators tab shows two kinds of evaluative fields:
- Hidden evaluators — added directly on this tab. These run automatically when the node is submitted and are not visible to contributors. Each evaluator uses a criterion (LLM-as-judge or programmatic) and targets specific fields on the node.
- Inline evaluative fields — Criteria fields and LLM-response fields with evaluation mode enabled that were added on the Fields tab. These are visible to contributors on the form but also appear here for a consolidated view of all evaluation activity on the node.
Preview tab
Shows a read-only preview of what the form looks like to a contributor, including upstream fields and grid layout. Template variables in prompts are resolved to current values for visual verification. Useful for checking the contributor experience before publishing.
Settings tab
| Setting | Description |
|---|---|
| Assignee Roles | Which contributor roles can claim tasks at this node. Required if the node has any human-input fields. If the node only has LLM/predefined/static fields, the role is optional (the node auto-submits). |
| LLM Metadata | Toggle to control which generation details are visible to contributors (model, tokens, cost, prompt, thinking). All five can be toggled individually. Off by default. |
| LLM Generation Settings | Node-level defaults for LLM fields (see below). |
| Contributor Permissions | Rules for who can work on this node based on upstream participation (exclude contributors who worked on specific nodes, require same contributor as an upstream node, sequential exclusion, parallel exclusion). |
LLM generation settings
When a subtask node has LLM-generated fields, the Settings tab shows compact controls for:
| Setting | Description | Default |
|---|---|---|
| Model | Default model for all LLM fields on this node. Choose a platform/custom model, or select "Use model from dataset" to load the model from a CSV column. | — |
| Temperature | Default temperature (0–2). Individual fields can override this. | 0.7 |
| Regeneration | Toggle Allow regeneration to let contributors re-trigger LLM fields. Set Max retries to cap how many times they can regenerate (1–10). | Off; 3 retries |
| Tool History | Toggle to show tool call history to contributors on this node. | Off |
| Thinking | Toggle extended thinking (only shown when the selected model supports it). | Off |
| Tools | Configure node-level tool calling — select tool endpoints, set max rounds (1–20, default 5). | — |
LLM field-level config
Individual fields set to LLM Response mode have their own overrides. For a comprehensive guide to prompt sources, context mechanisms, and placeholder mapping across all LLM field types, see Prompt Configuration.
- Model and temperature overrides per field (default: inherit from node).
- A prompt source — one of three options:
- Use prompt from dataset — the prompt is loaded from a CSV column at seeding time.
- Use custom prompt — opens a prompt editor with support for template variables
(placeholders). Type
{{to autocomplete available variables from upstream nodes and fields. - Use prompt from library — select a prompt from the Prompt Library. Placeholders defined in the library prompt can be mapped to pipeline fields.
- An optional system prompt (configured in a separate tab alongside the user prompt in the Prompt Configuration Modal).
- Extended thinking with reasoning effort (
low,medium,high). Defaults tohighwhen thinking is enabled. - Tool bindings for MCP/HTTP tool-calling.
- Context sources —
allupstream fields (default) or a specific list of node/field references. Toggling off "all" lets you pick individual fields.
Review node
A quality control node where reviewers evaluate upstream work. A review node can receive input from one or more upstream subtask or review nodes — reviewers see the upstream submissions alongside review-specific fields. Review nodes have the same four tabs as subtask nodes: Fields, Evaluators, Preview, and Settings.
Fields tab
- A mandatory Pass / Fail outcome field is always shown at the bottom — it cannot be removed or reconfigured.
- Upstream linkage is auto-determined from the graph topology based on incoming edges. It is not user-selectable — connect the review node to subtask or review nodes to establish which upstream work is reviewed.
- You can add additional review fields (feedback, notes, ratings) using the same form builder as subtask nodes.
- Upstream Fields lets you recall and display fields from any upstream node in the graph — not just the immediately connected node. This works the same as subtask nodes.
- Grid Layout works the same as subtask nodes.
Settings tab
The review Settings tab includes LLM Metadata, LLM Generation Settings, and Contributor Permissions (same as subtask), plus:
| Setting | Description |
|---|---|
| Reviewer Roles | Which contributor roles can claim review tasks at this node. Always required. |
| Reassignment Rules | Controls for rejection routing, pass routing, re-review limits, and escalation (see below). |
Reassignment rules
| Setting | Options |
|---|---|
| On Rejection | Same Person (default) — return to original contributor. Pool Including Original — return to pool, original can reclaim. Pool Excluding Original — return to pool, original excluded. |
| On Passing | Same three options — controls who gets assigned to the next downstream node. |
| Max Re-reviews | How many times a task can be sent back for rework. -1 = unlimited (default). |
| When Limit Exceeded | Quarantine task (default) — node enters Quarantined status, needs admin intervention. Escalate — node enters Escalated status, assigned to the escalation role. |
| Escalation Role | Which role receives escalated tasks (shown only when action is Escalate). |
Review outcomes
- Pass — the task moves forward to the next node via the approve edge.
- Fail — the task is sent back via the reject edge for rework based on the reassignment rules.
Logic gate node
An automated decision node that evaluates conditions and routes tasks without human intervention.
- Inputs: Accepts multiple incoming connections. The node waits for all connected upstream nodes to complete before evaluating (AND-join behavior).
- Outputs: 2 fixed branches: Then (condition met) and Else (condition not met). Both branches must be connected for the pipeline to be valid.
- Configuration: Rule builder with conditions and operators. The node name is editable.
Rule builder
The logic gate rule builder supports nested condition groups:
IF (all/any of these conditions are true):
├── field_name operator value
├── AND/OR
└── (nested group)
├── field_name operator value
└── field_name operator value
THEN → [Then branch]
ELSE → [Else branch]Each condition compares a left operand (typically a field reference from an upstream node) against a right operand (a literal value or another field reference) using a comparison operator.
Supported operators
| Operator | UI label | Applies to |
|---|---|---|
eq | == | All types |
neq | != | All types |
gt | > | Numeric, Rating |
gte | >= | Numeric, Rating |
lt | < | Numeric, Rating |
lte | <= | Numeric, Rating |
contains | contains | String types (Short Text, Long Text, Datetime, Pairwise, single-select) |
startsWith | starts with | String types (Short Text, Long Text, Datetime, Pairwise, single-select) |
endsWith | ends with | String types (Short Text, Long Text, Datetime, Pairwise, single-select) |
isEmpty | is empty | All types |
isNotEmpty | is not empty | All types |
The available operators filter dynamically based on the selected field's type. For example, selecting a numeric or rating field only shows comparison and equality operators, while selecting a select field with Allow Multiple enabled (array type) shows only equality and empty checks.
Use cases
- Auto-rejection: Route tasks with low quality scores directly to an end node.
- Cross-field comparison: Compare two field values (e.g., ensure a rating exceeds a threshold).
- Category branching: Send tasks down different paths based on a select field value.