RAG Triad

What it means

The RAG triad is a way of evaluating a Retrieval-Augmented Generation system by breaking it into three questions, one for each edge of the pipeline:

• Context relevance - did retrieval fetch chunks that are actually related to the query?
• Groundedness (faithfulness) - is the generated answer supported by those chunks, or did the model make things up?
• Answer relevance - does the final answer address what the user actually asked?

Together these three cover the full path from question to retrieved context to answer. Each is typically scored by an LLM-as-judge on a labeled dataset, giving you three numbers per example instead of one opaque pass or fail.

Why it matters

A single end-to-end quality score tells you a RAG system is broken but not where. The triad makes failures diagnosable. Low context relevance points at the retriever, the embedding model, or chunking. Low groundedness with good context points at the generation prompt or a model that ignores its sources. Low answer relevance with good groundedness usually means the prompt or output format is off. Because the three legs decompose cleanly, you can fix one without guessing about the others - which is the difference between iterating with intent and tweaking at random.

In practice

The triad maps onto metrics you can compute today. Context-precision covers the retrieval leg and answer-relevancy covers the generation leg, with a faithfulness metric handling groundedness. DeepEval bundles all three as first-class RAG scorers, and Arize Phoenix lets you trace a query through retrieval and generation so you can attach each score to the exact span that produced it. Start by running the triad over a small golden dataset, read the lowest scorer first, and fix that leg before touching the others. The how to evaluate RAG guide walks through a full setup.