# IDS platform

> Services on the ids3 Kubernetes cluster at Maastricht University's Institute of Data Science, and how to use them. Every page listed here is plain Markdown; fetch it directly.

The platform is: a workspace service (Coder), an OpenAI-compatible LLM gateway (LiteLLM), S3 object storage (MinIO), and a portal for GPU booking and keys. Coder, the portal and SigNoz share one identity (Keycloak realm `ids3`). LiteLLM does not: its API is authenticated by a virtual key (`Authorization: Bearer sk-...`), and its dashboard by a separate Keycloak realm (`litellm`) for admins or email+password for invited users.

## Pages

- [Workspaces](https://docs.k8s.semanticscience.org/workspaces.md): Coder workspaces — template parameters, the 6-CPU rule for GPU workspaces, CLI, REST API, network policy from inside a workspace
- [Models](https://docs.k8s.semanticscience.org/models.md): the LiteLLM gateway — served model names (`gemma4-g1`, `nomic-embed-text`, `claude-*`), endpoint, auth header, examples, how to request a model
- [Data](https://docs.k8s.semanticscience.org/data.md): MinIO S3 storage — endpoint, how keys are issued, client examples
- [Rules](https://docs.k8s.semanticscience.org/rules.md): the constraints that are not obvious — public hostnames only, GPU CPU cap, no scale-to-zero, ephemeral /mnt
- [Help](https://docs.k8s.semanticscience.org/help.md): who to ask and what to include

## Key facts

- LLM endpoint: `https://litellm.k8s.semanticscience.org/v1` — OpenAI-compatible, `Authorization: Bearer sk-...`
- Chat model: `gemma4-g1` (32k context). Embeddings: `nomic-embed-text` (768-dim; prefix inputs with `search_query: ` or `search_document: `)
- Workspaces API: `https://coder.k8s.semanticscience.org/api/v2`, header `Coder-Session-Token`
- S3 endpoint: `https://data.k8s.semanticscience.org`
- From inside a workspace, use public hostnames only; internal cluster DNS is blocked
- GPU workspaces: keep `cpu` <= 6 (not enforced by the template yet; an oversized request Pends silently); one A40 per workspace; stop them when idle
- Coder `workspace` template parameters: cpu, memory_gb, gpu_count, gpu_model, home_gb (sizes the /mnt scratch disk, not /home/coder), image