ITSpedia – Your IT Encyclopedia / March 2026

Welcome to the March installment of contributions to your IT password collection. As always, we’re bringing you three more entries; you can find the complete list available for download as a PDF below.

We wish you enjoyable self-study!

Latest entries in the collection

NPU

Neural Processing Unit – specializovaný procesor navržený pro úlohy umělé inteligence a strojového učení. Na rozdíl od CPU a GPU je optimalizovaný hlavně na efektivní inference a akceleraci AI funkcí při nižší spotřebě energie, což je důležité zejména v noteboocích, edge zařízeních a AI PC.

Our tip: An NPU makes the most sense when you need to run AI locally, cost-effectively, and with low latency. However, it’s not enough to simply check whether a device “has an NPU”—you should also ask about supported frameworks, performance for specific use cases, and actual software support.

RAG

Retrieval-Augmented Generation – architektonický přístup, při kterém generativní AI před vytvořením odpovědi nejprve dohledá relevantní informace z externích zdrojů, například z interní dokumentace, databází nebo znalostní báze, a teprve potom je použije jako kontext pro odpověď. RAG pomáhá zlepšit relevanci výstupů, pracovat s aktuálními nebo interními daty a snižovat riziko halucinací.

Our tip: RAG isn’t a magic bullet for poor-quality data. More often than not, it fails not because of the model itself, but due to outdated sources, poor chunking, missing permissions, and ineffective search. Start by ensuring the quality of your knowledge base, and only then focus on prompt design.

Vector DB

Vector databases – specialized databases designed for storing, indexing, and searching vector representations of data, known as embeddings. Instead of traditional searches based on exact matches or SQL queries, they enable searches based on similarity and meaning, which is essential for semantic search, recommendation systems, multimodal AI, and RAG architectures.

Our tip: A vector database alone does not guarantee high-quality results. The quality of results also depends on how you create embeddings, how you combine vector and traditional full-text search, and how you manage metadata. Without a hybrid approach, the results are often worse than expected.