Authoring and graph connections
Turn a business task into reviewed nodes, ports and executable bindings
A pipeline graph is a set of tasks and their dependencies. Start with the expected output and the data that can establish it. The assistant can draft the graph, but installed capabilities, accessible endpoints and explicit bindings determine what can run.
You need a Semogram account with workspace membership, a project in that workspace and permission to create and run pipelines. Reading or writing also requires access to the selected endpoints and external systems. Creating a pipeline does not grant those permissions.
Describe a bounded task
Read the two records from demo_orders. Keep order_id and total, rename total to order_total, and show the resulting records. Use the installed select-project transform. Do not write to an external target or add a schedule. Show the proposed source and transform bindings before applying the draft.Open Pipeline Studio → New pipeline in the project, describe the task and review the proposed graph. Select the ingress node and choose its Data endpoint and Load mode. Select the transform node, choose its installed capability and fill the operation's parameter controls. Inspect the connection between their output and input ports before saving.
If the assistant reports a missing endpoint, plugin, MCP server, skill or ontology definition, create that actual resource in its owning surface. A suggested name in a graph does not install a connector or create a business table.
Understand ports and edges
| Part | Example | Meaning |
|---|---|---|
| Node ID | select-orders | Unique graph-local identity, not a database UUID |
| Port key | out / in | Exact connection key on the node |
| Port interface | collection | Declared data shape |
| Output pool | projected_orders | Named intermediate result available to consumers |
| Edge | ingress out → transform in | Dependency between two declared ports |
| Resource binding | An endpoint or capability UUID | Actual saved workspace resource |
The graph must have unique IDs, existing edge endpoints, correctly directed ports and no dependency cycles. Match collection/record/fact interfaces to the downstream task. Drawing an edge defines dependency; operation configuration still has to identify the right inputs and output names. For multi-input transforms, use each manifest's named inputs rather than guessing in.
Review changes
Check changed targets, operation parameters, write modes and resource versions. Accepting an AI proposal applies a change to the draft; saving a version and running it are separate decisions. Use undo/redo and comparison to inspect edits before commit.
Programmatic authoring
A complete serialized graph uses kind-specific sections such as ingress, transform, egress and ontologyMapping, not one universal config property. The document reference includes a complete two-node graph. The input methods reference shows HTTP creation, draft save, validation, version commit and execution, with separate MCP equivalents where exposed.
Individual node examples in this section are definition fragments. They belong inside a complete graph with real ports, edges, resource IDs and project/workspace metadata; they are not independent HTTP or MCP calls.