| name | rhoai-kfp-pipeline-authoring |
| metadata | {"author":"rhoai3-demo","version":"1.0.0","platform-family":"rhoai","platform-baseline":"repo","ocp-baseline":"repo","skill-group":"RHOAI Platform"} |
| description | Author, refactor, and review Kubeflow Pipelines v2 pipelines for the rhoai3-demo stages once active KFP content exists; during the reimplementation, use this skill to rebuild KFP standards from legacy references. Use when editing stage-*/**/kfp/**/*.py, KFP components, pipeline runner scripts, DSPA client code, pipeline PVC/RBAC integration, Output[Metrics]/Output[Model]/ Output[HTML] artifacts, caching behavior, Docling data-processing pipeline code adapted from Red Hat-documented examples, reusable component catalog adoption, modular KFP component anatomy, or RHOAI Dashboard pipeline visibility. Do NOT use for AI Pipelines product lifecycle, pipeline server setup, dashboard import/version/run/schedule operations, Elyra runtime configuration, or DSPA troubleshooting (use rhoai-ai-pipelines). Do NOT use for generic GitOps changes unless paired with project-gitops-authoring.
|
KFP Pipeline Authoring
Use this skill when working on KFP v2 pipelines in the RHOAI demo.
Use rhoai-ai-pipelines first for official product behavior around pipeline
servers, KFP SDK prerequisites, Kubernetes API storage, pipeline versions,
caching, experiments, runs, schedules, logs, Elyra, workspaces, and DSPA
troubleshooting. Use this skill when the task reaches repo-specific pipeline
Python, components, compiled artifacts, or runner scripts.
Before authoring a new KFP component, check whether an existing component in
the Kubeflow or Red Hat Data Services component catalogs can be reused or
adapted. Capture why a local component is needed when no reusable component
fits.
For Docling data-preparation pipelines, use
rhoai-model-customization-training for the official RHOAI
data-preparation chapter and the opendatahub-io/data-processing stable
branch examples before adapting code into this repo. Compare the current
main/kubeflow-pipelines tree when a newer reference implementation is
intentionally selected, and record the branch decision in the stage plan.
Reimplementation Status
The active implementation is being rewritten. No active KFP pipeline
implementation or stage runner scripts exist yet. Treat references to previous
stage folders as legacy examples for rebuilding pipeline standards, not as
active-project paths.
Do not run or modify scripts from backup/legacy-implementation-2026-06-09/
unless the user explicitly asks to restore or inspect the legacy implementation.
Workflow
- Read the affected stage README and existing KFP implementation.
- Treat active
stage-*/kfp/ content as the implementation target. Until it
exists, use
backup/legacy-implementation-2026-06-09/steps/step-12-mlops-pipeline/kfp/
only as a legacy reference.
- For Docling pipelines, read
rhoai-model-customization-training first and
capture whether the standard or VLM data-processing example is being
adapted.
- Read
references/kfp-patterns.md before editing pipeline definitions,
components, runner scripts, artifacts, or DSPA client code.
- For reusable component work, record selected catalog source, stability
level, input modes, image posture, and validation data set.
- Keep pipeline infrastructure in GitOps and compiled/uploaded pipeline
definitions in stage scripts.
- Keep component functions hermetic, typed, Dashboard-visible where possible,
and aligned with
rhoai-ai-pipelines and the active official documentation.
Validation
- Compile changed pipelines locally when possible.
- Validate runner scripts with shell syntax checks.
- If cluster execution is needed, follow the OpenShift safety guard in
AGENTS.md before suggesting or running live commands.
References
references/kfp-patterns.md../rhoai-model-evaluation/references/kfp-advanced-patterns.md
