Semogram for AI assistants
A compact scope and lifecycle guide for assistants using Semogram
Semogram connects operational records, models their business meaning, executes pipelines, answers queries and produces evidence-backed predictions. Use this guide to choose the correct resource and verify the result of each supported operation.
Resource scope
Workspace resources include plugin installations, data endpoints and standalone skills. Project resources include pipelines/runs, ontology definitions/bindings/queries and forecasters/predictions. Assertion workflows carry the appropriate workspace/project scope. Projects reference shared endpoints; they do not own separate endpoint copies.
Discover current capabilities and schemas on the actual MCP endpoint. On a workspace endpoint, select projectId from project_list for project actions. A project-scoped endpoint binds the project already. Permissions, policy and external credentials remain separate from resource discovery.
Identify the requested operation
| Intent | Resource/action |
|---|---|
| Configure a connector | Plugin installation |
| Select a reusable source/destination | Workspace data endpoint |
| Process or transfer records | Pipeline definition, then separate run |
| Define business terms | Ontology definition, then compiled package |
| Connect terms to values | Active read bindings |
| Make a reusable read | Query definition, validation and published release |
| Predict a future event/value | Forecaster, then asynchronous prediction |
| Record what happened | Evaluation with actual observation time/evidence |
A prompt describes intent. A tool call uses the discovered schema. A UI field expects its own representation. Do not invent API routes, tool names, resource UUIDs or missing datasets.
Lifecycle checks
- Review exact proposed asset revisions and their validation before committing.
- Saving a definition does not execute it or ingest data.
- Pipeline preflight validates configuration; it does not preview real processed rows.
- Scheduling is separate from configuring a connection or graph.
- Query publication pins logic; live records can still change.
- Forecaster versions pin published evidence releases. Prediction completion and observed outcome are separate.
- Assertion review, identity correction, action approval and outcome evaluation have different effects.
- Feedback does not guarantee automatic model retraining.
Verify before answering
Name the workspace/project and preserve returned run, artifact, release and prediction identifiers when relevant. Inspect asynchronous job states and execution evidence. Verify external receiving systems after writes. Explain whether the answer is a source observation, recorded claim, model prediction or proposal awaiting an action.
If evidence is partial or a capability is absent, explain the actual limitation and prerequisite. Do not turn a saved asset into a claim that ingestion occurred, a graph label into an enforced approval gate, or a completed model call into measured predictive accuracy.
Documentation map
Getting started → Core concepts. Detailed workflows: MCP, Plugins, Data Endpoints, Data Pipelines, Ontology and Forecasting and predictions.