> ## Documentation Index
> Fetch the complete documentation index at: https://docs.seekr.com/llms.txt
> Use this file to discover all available pages before exploring further.

# SeekrFlow Self-Managed architecture

> What runs in your cluster, what you provision around it, and how the pieces connect.

A self-managed deployment runs the SeekrFlow platform in your Kubernetes cluster and keeps its data in services you provision alongside it. Users and API clients reach it through a single hostname.

```mermaid theme={null}
flowchart LR
  users([Users and API clients]) --> ingress[Ingress controller<br/>and load balancer]

  subgraph env [Your environment]
    subgraph cluster [Kubernetes cluster]
      platform[SeekrFlow platform<br/>services]
      gpu[GPU nodes<br/>model servers and<br/>training jobs]
      platform --> gpu
    end

    subgraph data [Services you provision]
      pg[(PostgreSQL)]
      obj[(Object storage)]
      fs[(Shared file storage)]
    end

    ingress --> platform
    platform --> pg
    gpu --> obj
    gpu --> fs
  end

  registry[Seekr artifact<br/>registry] -.->|container images| cluster
  platform -.->|optional| external[Amazon Bedrock]
```

## How it works

* **Installation** – Helm installs the SeekrFlow operator. The operator reads your values file, creates the platform services, model servers, agents, and tools, and keeps them matched to your configuration. The operator, not Helm, owns most SeekrFlow pods.
* **Access** – Users and API clients reach SeekrFlow through the ingress hostname. The authentication service validates every request.
* **Capabilities** – The platform services provide the web interface and API, agents and custom functions, fine-tuning and dataset preparation, inference, and explainability.
* **Inference** – The inference router sends each request to a model server on your GPU nodes, or to Amazon Bedrock for externally hosted models. Model servers cache their weights on shared file storage.
* **Training and data preparation** – Fine-tuning, alignment, and file ingestion run as jobs through Argo Workflows, on GPU and CPU nodes.
* **Data** – PostgreSQL stores service data. Your object storage holds datasets, model artifacts, and fine-tuning checkpoints. Retrieval-augmented generation uses OpenSearch.
* **Licensing** – Services don't report ready until the License service confirms a valid license key. The license also sets a ceiling on the number of GPUs SeekrFlow workloads can request across the cluster.

## Services in a deployment

### Services you provide

Which services fill each role depends on where you deploy SeekrFlow. Select your environment.

<Tabs>
  <Tab title="Amazon Web Services (AWS)">
    | Role | Service | Purpose |
    | - | - | - |
    | Kubernetes | Amazon Elastic Kubernetes Service (Amazon EKS) | Runs the SeekrFlow platform. |
    | PostgreSQL | Amazon Relational Database Service (Amazon RDS) or Amazon Aurora | Stores service data. |
    | Object storage | Amazon Simple Storage Service (Amazon S3) | Stores datasets, model artifacts, and checkpoints. |
    | Shared file storage | Amazon Elastic File System (Amazon EFS) | Caches model weights for GPU nodes. |
    | OpenSearch | Amazon OpenSearch Service | Stores the search indexes for file search and retrieval-augmented generation. |
    | Ingress | Application Load Balancer | Routes traffic to the platform. |
    | Cloud identity | Amazon EKS Pod Identity or AWS Identity and Access Management (IAM) roles for service accounts | Authenticates pods to cloud services. |
    | Secret store (optional) | AWS Secrets Manager or AWS Systems Manager Parameter Store | Holds credentials outside the cluster. |
    | External models (optional) | Amazon Bedrock | Serves inference alongside your GPU nodes. |
  </Tab>

  <Tab title="On-premises">
    | Role | Service | Purpose |
    | - | - | - |
    | Kubernetes | Your own Kubernetes cluster | Runs the SeekrFlow platform. |
    | PostgreSQL | PostgreSQL you provide | Stores service data. |
    | Object storage | An S3-compatible store, such as MinIO | Stores datasets, model artifacts, and checkpoints. |
    | Shared file storage | Network File System (NFS) or Longhorn | Caches model weights for GPU nodes. |
    | OpenSearch | Deployed in the cluster by the chart | Stores the search indexes for file search and retrieval-augmented generation. |
    | Ingress | NGINX or Traefik | Routes traffic to the platform. |
    | Object storage credentials | Static keys in a Kubernetes secret | Authenticates pods to your object store. |
  </Tab>
</Tabs>

### Components the chart deploys

| Component | Purpose |
| - | - |
| Argo Workflows | Runs fine-tuning, alignment, and ingestion jobs. |
| MLflow | Tracks experiments and provides a model registry. |
| Keycloak (optional) | Brokers single sign-on with your identity provider. Off by default. |
| Valkey (optional) | Caches sign-in and inference data. Sign-in rate limiting depends on it. Off by default. |
| JobSet (optional) | Runs multi-node fine-tuning. Off by default. |


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.