Hallucination

A model generates text that is fluent and confident but factually wrong, because it is optimised for plausibility rather than truth.

Hallucination is not a bug to be fixed but a property of how language models work: they predict likely next tokens, and a likely sequence can be a false one. RAG, grounding in retrieved evidence, and structured output validation reduce but do not eliminate it. The most dangerous hallucinations are the ones that look correct to a non-expert.

Explained in depth

RAG

What retrieval-augmented generation actually is, how the indexing and query paths work, and where the evidence complicates the standard advice - hybrid search does not universally beat dense retrieval, and neither beats a bad chunking decision.

Learn it in context

AI evals - building the suite

Learning path · intermediate

Related concepts

Retrieval-augmented generation
Alignment
Reasoning
Synthetic data
Evals
LLM-as-a-judge