Open components. One coherent system.
Kubernetes, Kubeflow, Argo Workflows, Harbor and MLflow form the platform layer. Model serving and retrieval connect through controlled APIs.
VFUSION AI / TECHNOLOGY
The whole chain. Made visible.
From document to answer, GPU to API. This is our reference architecture: components are selected for each engagement, with explicit boundaries and measurement points.
VF / 01REFERENCE ARCHITECTURE
- People and applications: CHAT / API / DEVELOPER TOOLS
- Access · API gateway: IDENTITY / LIMITS / ROUTING
- AI agents, retrieval and models: TOOLS / SOURCES / MODELS
- GPU · Storage · MLOps: KUBERNETES / NVMe-oF / OBSERVABILITY
INSIDE THE BOUNDARY
Every layer has a purpose.
This diagram is a reference design. Exact components, active layers and access boundaries are defined for each project.
Document processing as a foundation
Docling on GPU, extraction models for OCR and tables, and Gemma for visual interpretation. Dutch and English documents both matter.
Observe, test, operate
OpenTelemetry and Loki make processing stages visible. Quality evaluation, versioning and access controls belong in the architecture.
Agentic processing across applications
A shared AI interface connects model serving, document processing and orchestration tools to websites and apps. Data scope, allowed actions, approval and audit remain explicit for each application. Financial amounts and state transitions are validated by the application.
WHAT YOU GET
Language and code
gpt-oss, Gemma and an in-house quantised 72B coding model.
Retrieval
BGE-M3, Qwen3 embeddings, BM25 and reranking.
Vision and documents
YOLO, vision-language models, Docling and OCR.
Compute and fabric
H100, Kubernetes, Slurm, NVMe-oF and RoCE.
LAYER 01 → 10
RAG in ten layers.
A reference for design and evaluation. No promise of infallible answers; every layer has a defined responsibility.
01Ingest and normalise
Docling, OCR and orchestration turn source documents into consistent text and metadata.
02Two retrieval tracks
BM25 and dense embeddings combine exact search with semantic meaning.
03Select and rerank
HNSW search and reranking select the most relevant passages.
04Confidence gate
Relevance, freshness, authority and agreement determine whether there is enough evidence.
05Answer from context
The model works with selected sources and a versioned instruction.
06Cite the source
Claims connect to passages, documents and pages.
07Check grounding
An additional NLI check can flag missing or weak support.
08Measure quality
Evaluation sets, MLflow and alerts make changes visible.
09Cache with a policy
Exact and semantic caches with versions and expiry.
10Trace the chain
Per-stage tracing with OpenTelemetry and Loki.