Learning path · Evaluation & Quality · 72
Hallucination
Model outputs that sound plausible but are factually unsupported or contradict provided evidence.
Why it matters
- Top risk in customer-facing and compliance workflows.
- RAG without faithfulness checks can increase confident errors.
- Detection blends automated metrics and human audit.
Key ideas
- Unsupported claims
- Confident tone
- Faithfulness testing
Top resources
- 01PaperLewis et al.
Retrieval-Augmented Generation for Knowledge-Intensive NLP
Why this resource. Grounding as the architectural answer to fluent invention.
Covers in this concept
- grounding
- retrieval
- 02DocsExploding Gradients
RAGAS
Why this resource. Faithfulness metrics that quantify hallucination on your corpus.
Covers in this concept
- faithfulness
Hallucinations are fluent claims with no support in the context, or claims that contradict it. They show up when retrieval misses, the question sits outside the window, or the prompt forbids saying you do not know. Require citations, set a retrieval bar, and measure faithfulness. Watch support tickets: deflection looks good until people escalate because the policy was wrong.
Updated 2026-08-09 · Full learning path