Workflow nodes

Nodes are the vocabulary of plans. Each kind declares what it needs and what it produces; edges are valid only when both sides' interfaces agree. Conceptual overview in Workflows and plans.

Node kinds at a glance

Kind	Reads	Writes	Use when
Source	Integration	Records / collections	Data enters from your systems
Transform	Records	Reshaped records	Shape is almost right
Extract	Documents	Fields, labels	Input is unstructured
Map	Any shape	Mapped shape	Conversion must be inspectable
Skill	Guidance	Assisted behavior	A judgment call recurs
Query	State	Lookup results	Branching on what is known
Predict	State	Forecast / label	A decision needs likelihoods
Destination	Records	External system	Results must live outside
Ontology write	Mapped data	Entities / relationships	Structure must persist

Assertion mapping and assertion materializer nodes serve advanced governance flows; most workspaces never configure them directly.

Interface shapes

Shape	Meaning	Example
Record	One structured item	An order with id, total, date
Collection	Many records together	Yesterday's orders
Document	Semi-structured content	A ticket thread, a PDF invoice
Blob	Opaque bytes, carried not inspected	An image attachment
Ontology entity / relationship	Instances of defined types	Resolved Customer + Orders
Fact set	Grouped facts for prediction inputs	Per-account feature rows
Assertions	Checked statements with outcomes	"Total matches invoice: confirmed"

Source node

Points at an installed source integration. Holds no secrets — those live with the integration.

Field	Meaning	When to set it
Source	Which installed integration to read	Always — pick the door first
Read mode	Full scope vs what changed since last time	Incremental for schedules, full for first loads
Selection	Filters and field subset	Narrow before downstream, not after
Checkpointing	Resume where it stopped	On for anything scheduled
Batch size	Rows per pull	Raise for bulk loads
Retries	Attempts and backoff (fixed / exponential)	Raise for flaky sources
Row-count watch	Flag unexpected volume drops	On for feeds that must not silently shrink

{
  "kind": "source",
  "dataEndpointName": "crm.orders",
  "mode": "incremental",
  "flow": {
    "checkpoint": { "enabled": true },
    "batch": { "size": 1000 },
    "retry": { "maxAttempts": 3, "backoff": "exponential" }
  },
  "anomaly": { "rowCountDropThreshold": 0.5 }
}

A second example — one-time full read for seeding:

{
  "kind": "source",
  "dataEndpointName": "crm.products",
  "mode": "full",
  "flow": { "batch": { "size": 5000 } }
}

Image: source node configuration with mode, selection, and flow controls visible.

Transform node

Field	Meaning	When to set it
Function	Named transform (aggregate, join, normalize…)	Always — read its description first
Inputs	One stream, or named streams for joins	Named inputs for merges
Parameters	Group-by keys, join keys, thresholds	Where "aggregate" becomes "average order value by region"
Output	Where the reshaped stream goes	Always

{
  "kind": "transform",
  "function": "aggregate",
  "input": "orders",
  "params": { "groupBy": ["region"], "measures": [{ "field": "total", "op": "avg" }] }
}

Prefer small chained transforms over one large one — failures point at the guilty step.

Extract node

Field	Meaning	When to set it
Strategy	Language model, parser, or hybrid	Model for prose, parser for rigid formats, hybrid for laid-out text
Prompt / rules	What to pull and in what shape	Be explicit about downstream shape
Output	Structured fields onward	Always

Validate on samples before trusting at volume — quality is measured per field, not per document.

Mapping node

Declares shape-to-shape conversion; for ontology work see Ontology mappings.

Field	Meaning	When to set it
Inputs	Named upstream streams	Always
Entities / relationships	The mapping rules	The inspectable conversion itself
Mode	Update-or-insert / snapshot / delta	Upsert for schedules, snapshot for rebuilds

{
  "kind": "ontology-mapping",
  "inputs": { "orders": "clean_orders" },
  "mode": "upsert",
  "entities": [{ "class": "Order", "id": { "template": "ord-{order_id}" } }]
}

Skill node

Field	Meaning	When to set it
Skills	Installed skills by reference	One or more, each earning its place
Mode	Authoring-assist / execution-assist / validation-assist	Match where guidance should bite

Query node

Reads ontology or workflow state mid-plan without consuming it. No configuration beyond what to read and where it flows — branch on thresholds, lookups, and existence checks.

Predict node

Field	Meaning	When to set it
Forecaster	Which predictor, by id or slug	Always — one must be named
Subject	What the prediction is about	Name the decision target
Horizon	Aim point and window end	Always — a forecast without one has no test date
Parameters	Forecaster-specific settings	Per predictor docs

{
  "kind": "predict",
  "forecasterSlug": "renewal-risk-v3",
  "subjectRef": "account.top20",
  "horizon": "next-quarter",
  "horizonEndsAt": "2026-12-31T23:59:59Z"
}

Full predictor detail in Forecasts.

Destination node

Field	Meaning	When to set it
Destination	Installed integration receiving the write	Always
Write mode	Append / replace / update-or-insert	Append for history, replace for snapshots under approval
Source stream	Which upstream data gets written	Always

{
  "kind": "destination",
  "dataEndpointName": "warehouse.forecast_output",
  "source": "risk_scores",
  "writeMode": "append"
}

Gate external writes with node-level approval — see Write policies.

Ontology write node

Field	Meaning	When to set it
Definition	Governing ontology model	Always
Mode	Update-or-insert / snapshot / delta	Upsert for sync, snapshot for rebuilds
Atomicity	Transactional vs per-unit	Transactional for small critical writes, per-unit for resumable bulk
Validation	Unknown terms, invalid records, cardinality, datatypes	Quarantine at volume, reject where money moves
Conflict policy	Which source wins, in what priority	Wherever two systems disagree

Choose ontology writes where structure compounds across runs — not as the default outcome of every plan.

Choosing between lookalikes

Pair	Rule
Transform vs mapping	Reshaping data → transform; declared shape contract or ontology landing → mapping
Extract vs transform	Unstructured in, fields out → extract; fields in, fields out → transform
Destination vs ontology write	People read destinations; predictions read ontology
Query vs predict	Known → query; likely → predict

Validation

Scope	Severity	Effect
Graph / node / edge	Error	Blocks dry-run and commit
Graph / node / edge	Warning	Asks for a reading first
Graph / node / edge	Info	Context only

When an edge is invalid, select it — the message names the incompatible interfaces.

On this page