[Ecosystem] OGX

[franciscojavierarceo]

Contact emails

farceo@redhat.com, shan@redhat.com, cdoern@redhat.com

Project repo URL

https://github.com/ogx-ai/ogx

GitHub handles of the project maintainer(s)

@leseb @mattf @franciscojavierarceo @cdoern @skamenan7

Project license

https://github.com/ogx-ai/ogx/blob/main/LICENSE

Project summary

Open-Source, Vendor-Neutral Generative AI Application Server

Project description

OGX, formerly Llama Stack, provides an open-source API layer for building agentic AI applications on user-controlled infrastructure. It exposes OpenAI-compatible APIs for chat completions, Responses, vector stores, files, tools, and retrieval-augmented generation, while supporting multiple model and infrastructure backends.

OGX is relevant to the Kubeflow community because it gives platform teams and ML application builders a consistent application-facing API on top of Kubernetes-native AI infrastructure. It can be deployed with Kubernetes operators, connected to OpenAI-compatible inference services such as vLLM/KServe deployments, and integrated with Kubeflow Pipelines / Data Science Pipelines for data ingestion workflows.

Kubeflow Subprojects Integration

OGX integrates with Kubeflow Pipelines / Data Science Pipelines through multiple RAG demos that ingest multimodal data into OGX/Llama Stack vector stores. These pipelines process source data, upload files with the Llama Stack client, create vector stores, attach files, and make the resulting indexes available for RAG queries through the Responses file_search API.

Concrete examples:

• PDF ingestion pipeline: https://github.com/opendatahub-io/rag/tree/main/demos/kubeflow-pipelines/pdf-conversion
• ASR/audio ingestion pipeline: https://github.com/opendatahub-io/rag/tree/main/demos/kubeflow-pipelines/asr-conversion
• OCR/image ingestion pipeline: https://github.com/opendatahub-io/rag/tree/main/demos/kubeflow-pipelines/ocr-image-conversion
• Spreadsheet ingestion pipeline: https://github.com/opendatahub-io/rag/tree/main/demos/kubeflow-pipelines/spreadsheets-conversion
• End-to-end bank RAG demo: https://github.com/opendatahub-io/rag/tree/main/demos/redbank-demo

Documentation

• OGX documentation: https://ogx-ai.github.io/docs
• OGX repository: https://github.com/ogx-ai/ogx
• Quick start: https://ogx-ai.github.io/docs/getting_started/quickstart
• OpenAI API compatibility: https://ogx-ai.github.io/docs/api-openai
• Kubernetes/operator support: https://github.com/ogx-ai/ogx-k8s-operator
• Open Data Hub RAG/Kubeflow Pipelines demos: https://github.com/opendatahub-io/rag

Testing infrastructure and CI/CD

OGX uses GitHub Actions for unit tests, integration tests, OpenAPI/SDK validation, docs builds, provider builds, and OpenResponses conformance validation. The repository publishes CI status through README badges and maintains active automation for testing and release workflows.

The Open Data Hub RAG integration demos include Kubeflow Pipelines source definitions and compiled pipeline YAMLs for PDF, audio/ASR, OCR/image, spreadsheet, and end-to-end RAG ingestion flows. These examples demonstrate practical integration between Kubeflow Pipelines / Data Science Pipelines and OGX/Llama Stack vector store ingestion.

Additional information

OGX was previously known as Llama Stack, so some existing integrations and demos still use the llama-stack and llama-stack-client names. The ecosystem application should treat those artifacts as part of the same project lineage and integration story.

[andreyvelich]

Thanks for submitting this application @franciscojavierarceo!

As @kubeflow/kubeflow-steering-committee voted to accept OGX as part of Kubeflow Ecosystem, we can move this forward 🚀
Recording from the vote: https://youtu.be/8vXQtZTNLyo?t=925

@franciscojavierarceo Please go ahead and update the remaining sections as described here: https://github.com/kubeflow/community/tree/master/ecosystem#ecosystem-project-process

/area ecosystem-accepted