Skip to content
Copyright © 2022-2026 basebox GmbH, all rights reserved.
Licensed to be used in conjunction with basebox, only.

// installation

Understand the architecture

Applies to

Product: Cloud · Server · Audience: Platform Operator · Administrator · Security reviewer

basebox is not one workload. It is three layers with very different resource profiles. Almost every sizing misunderstanding – and every sentence like "basebox requires 2 × H200" – comes from collapsing them into one.

flowchart TB
  P["<b>basebox platform</b><br/>UI · apps · authentication · permissions<br/>audit · APIs · connectors · orchestration<br/><i>little or no GPU</i>"]
  S["<b>Service models</b><br/>embeddings · RAG processing · OCR<br/>speech-to-text · TTS · long-term memory<br/><i>dedicated GPU capacity ideal</i>"]
  I["<b>Inference</b><br/>the LLM / VLM / reasoning model<br/><i>sizing depends on model, context, concurrency</i>"]
  P --> S --> I
  style P fill:#f4f2ee,stroke:#524e47,color:#1d1e1c
  style S fill:#dcefe2,stroke:#3a7a49,color:#1d1e1c
  style I fill:#dbeafe,stroke:#1e40af,color:#1d1e1c

Layer A – basebox platform

The management and application layer: the web UI, apps, authentication (Keycloak), permissions, audit log, the OpenAI-compatible and REST APIs, connector orchestration (MCP gateway), and general platform services (frontend, AISRV, storesrv, databases).

Resource profile: CPU, RAM, storage, network. The platform itself does not require large GPU resources. It can run on a CPU-only application server – the deployment topologies page shows exactly that layout.

Do not write or read "basebox needs GPU X" as a statement about this layer. It is technically misleading.

Layer B – service models

Additional AI models and services that basebox features rely on:

  • Embeddings and RAG processing – turning documents into searchable vectors and retrieving them (ragsrv, ragsrv-support)
  • OCR – text recognition for scans, photos and image-only documents
  • Speech-to-text – dictation and audio transcription
  • TTS, long-term memory and other internal AI services as they are added

Resource profile: these benefit from GPU acceleration and, as a general architecture principle, should have dedicated GPU capacity so that an upload or a transcription never competes with inference for VRAM. Typical useful capacity is in the order of 48 GB GPU memory, preferably around 96 GB for larger usage scenarios – but this depends on the customer's usage and which services are enabled.

Not a universal minimum

The figures above are architecture guidance, not a supported minimum requirement. Exact supported values are pending DevOps confirmation and are stated only on the reference configuration pages once confirmed.

Service models can run in CPU mode if no GPU is available; RAG and OCR remain functional but slower.

Layer C – inference

The LLM / VLM / reasoning-model layer. It is separate from basebox itself: AISRV talks to it over an OpenAI-compatible API, whether it runs as the bundled vLLM on the same node, on a separate GPU host, in another cluster, or – where approved – at an external provider.

Resource profile: this is where the large GPUs go, and its sizing depends on:

  • selected model and model size
  • quantization
  • context length
  • expected concurrency
  • required response latency
  • usage pattern and customer workload

This is a genuinely complex sizing topic. This documentation does not primarily answer "what hardware should I buy?" – separate sizing material exists for that. It answers: I have this infrastructure; how do I deploy and operate basebox correctly on it?

What this means in practice

Question Answer
Can basebox run on a CPU-only server? The platform can. Service models will run in CPU mode (slower). Inference must live on a GPU host somewhere.
Why does a reference configuration list two large GPUs? For inference and service models – not for basebox itself.
Can I change the model without touching basebox? Yes. Inference is a separate layer; change the endpoint or model in AISRV's configuration.
Where do I size? Service models: Resources & scaling. Inference: Models & inference. Concrete setups: Reference configurations.

Deeper reading

Next step: Deployment models