LOOK UP WHEN NEEDED
Eval
Tests AI behavior against defined tasks and criteria.
An everyday example
An exam needs questions and grading rules.
See how these concepts connect
- Ask a questionQuery
- Search the catalogSearch index
- Retrieve passagesRelevant chunks
- Answer with referencesGeneration & citation
Check retrieval and generation separately. An answer can sound fluent despite wrong evidence; citations still need checking.
Where this term is useful
These are related contexts, not prerequisites.
Back in the project
When it comes up: When grounding answers in documents.
What to do next: Use known-answer questions to check retrieval, sources, and the final answer separately.
Open the related method and stepsWhat can this analogy leave out?
Similarity is not correctness; RAG does not automatically train documents into a model.
Everyday examples explain roles and relationships; actual behavior follows the tool, language, or platform.