Grounding, Citations, and RAG Evaluation

Key jargon

Term	Plain-language meaning
Context recall	Whether the retrieved evidence includes the information needed to answer.
Context precision	How much retrieved material is actually relevant.
Faithfulness	Whether answer claims are supported by the supplied context.
Citation correctness	Whether each citation points to evidence that supports its associated claim.

Key concepts

An answer can be factually correct but unsupported by retrieval, or well-grounded in a source that is itself wrong.
Measure retrieval quality before blaming generation, then check claims and citations at sentence level.

Concept map

flowchart LR
    A["Evaluate retrieval coverage"] --> B["Generate from evidence"]
    B --> C["Match claims to citations"]
    C --> D["Score and diagnose failure stage"]

Evaluation matrix

Retrieval	Answer	Interpretation
Good	Good	Desired grounded response
Good	Bad	Generation/prompt/validation failure
Bad	Good-looking	Likely unsupported prior knowledge or lucky guess
Bad	Abstains	Correctly bounded but retrieval needs work

For each material claim, record supporting source span, whether the span entails the claim, source authority, and freshness. A citation that merely discusses the topic is not support.

Test set

Include answerable, unanswerable, conflicting-source, outdated-source, permission-boundary, table, and multi-document questions.

Exercise

Build a ten-question set for five documents. Grade retrieved evidence, claim support, citation correctness, and appropriate abstention separately.

Operational rule

Do not let the same model both generate and certify correctness without independent deterministic checks or human sampling.