Architecture1

• RAG: Should the Vector Store Live on the Model Host?The decision is about RAM, not VRAM — and the two failure modes look identical from the outside

Sizing1

• How Much VRAM Does an LLM Actually Need?A sizing method you can recompute for any model — and the reason your weights are usually not what ran out of memory

Throughput1

• Why Your GPU Sits at 30% While Serving RequestsLow utilisation is usually an unbatched workload, not an underpowered card