What "good" actually means before you can measure it

Context Relevance asks a question about the lookup step, before a single word of the answer is written: of the material the system pulled up to answer with, how much of it actually relates to what the user asked?

• Plain English: "Did we fetch the right pages?" If someone asks about refund timing and the system retrieves the shipping policy, context relevance is low - the answer was doomed before it started.
• The failure it catches: the system confidently answering from the wrong source material. A perfect writer cannot save a wrong document.

Groundedness is the anti-hallucination check. It asks: is each statement in the answer actually supported by the material that was retrieved, or did the model add things from thin air?

• How it is scored, in plain terms: break the answer into its individual claims, check each claim against the retrieved context, and score the fraction that are supported. Nine of ten claims backed by the documents is a groundedness of 0.9. This "break into claims, check each" method is worth remembering - it is exactly how the machine does it too.
• The failure it catches: the fluent invention - a plausible detail, number, or policy that appears nowhere in the source. This is the failure that quietly misinforms customers.

Answer Relevance closes the loop back to the user. An answer can be built from the right documents and be perfectly grounded in them, and still miss - because it answered a slightly different question, or buried the point, or trailed off into a related topic.

• Plain English: "Was this a reply to what they actually asked?" Correct and grounded but off-target is still a bad answer.
• The failure it catches: the technically-true non-answer - accurate, sourced, and unhelpful.

It is tempting to measure just the final answer - that is what the user sees, after all. But a single answer score cannot tell you which half to fix, and worse, it can lull you.

• Great retrieval + bad answer = still bad. The system found the perfect document and then wrote a muddled, off-target reply. The retrieval score looks great; the customer is unhappy.
• Bad retrieval dooms even a great generator. The best writer in the world, handed the wrong documents, produces a confident wrong answer. No amount of writing skill recovers from being fed the wrong facts.
• Two halves, two teams, two fixes. Fixing retrieval means better search and indexing. Fixing generation means better prompting and instructions. If you only know "the answer was bad", you do not know which team to send it to.

Here is the takeaway to write on a sticky note. When an answer is bad, the Triad tells you where to look:

• Low groundedness - the answer contains claims the documents do not support - means the generation half is the problem. The right info may have been there; the model invented or drifted. Fix the writing side.
• Missing information - the right material was never retrieved, so low context relevance - means the retrieval half is the problem. Fix the search side; no prompt tweak will conjure a document that was never found.

The RAG Triad framing was popularized by DeepLearning.AI's short course on building and evaluating advanced RAG, taught with the open-source TruLens tool, which scores Context Relevance, Groundedness, and Answer Relevance on real traffic. A parallel open-source framework, RAGAS, offers a similar family of metrics under slightly different names - faithfulness for groundedness, and so on. You do not need to pick a tool today. You need the concept: three checks, two halves, and a diagnostic that tells you which half broke. The builder track puts real numbers on all three.