Series 3 of 4 · AI Evals · Level 2 · Topic 14

The RAG Triad

The Practice
IN THIS NOTE YOU WILL LEARN
  • 01.The three independent metrics that decompose RAG quality -- context relevance, faithfulness, and answer relevance -- and why keeping them separate is the entire point.
  • 02.The failure pattern most teams misread: high faithfulness plus low answer relevance means the system is accurately summarizing the wrong thing.
  • 03.Why enterprise RAG quality is inseparable from access control, document freshness, and multi-tenant architecture.

The Story

Consider a large employer that launched an internal HR assistant over benefits, leave, and travel policy documents. Employees started getting answers that were articulate, cited policy language, and were often slightly wrong on the newest rules. The team's first instinct: upgrade the model -- "it's hallucinating."

After trace review, the pattern was obvious. The retriever kept surfacing older policy versions because the index sync lagged behind the source-of-truth document system. The answer stayed faithful to retrieved context -- the model wasn't making things up. The context itself was stale. The team thought they had a model problem. They actually had a retrieval freshness problem.

A SaaS support bot asked: "Can I cancel and get a prorated refund on the annual enterprise plan?" It retrieved the refund policy for self-serve annual plans, summarized it accurately, and cited it correctly. The answer was grounded. It was also wrong -- the user was on an enterprise contract with different terms. High faithfulness, low answer relevance. The team initially celebrated because citation quality improved. In reality, the bot had become a better summarizer of the wrong document.

Three systems. Three different "RAG failures." None of them were model problems. All of them were invisible to a single blended "RAG quality" score.

The Core Idea

RAG failures are usually not "the model was wrong." They are "the retriever fetched the wrong thing," "the model stayed grounded to the wrong thing," or "the answer stayed grounded but didn't answer the user's actual question."

The RAG Triad -- Three Questions, Three Fix Paths
Keep them separate. That separation IS the diagnostic value.
CONTEXT RELEVANCE
"Did the retriever fetch the right docs?"
METRICS
Context Precision + Context Recall
WHEN IT FAILS
Stale index, wrong tenant, incomplete evidence
Fix chunking, reranking, freshness, ACLs
RETRIEVER
FAITHFULNESS
"Did the model stick to what was retrieved?"
METRICS
Groundedness / Faithfulness
WHEN IT FAILS
Model invents claims not in retrieved context
Fix prompt constraints, add citation checks
GENERATOR
ANSWER RELEVANCE
"Did the answer address the user's question?"
METRICS
Answer Relevancy
WHEN IT FAILS
Well-sourced answer to the wrong question
Fix query understanding, retrieval scope
END-TO-END
High faithfulness + low answer relevance = accurately summarizing the wrong thing.

Where This Hits in Production

Access control is part of retrieval quality. In enterprise RAG, "did the retriever fetch the right documents?" has a security dimension. AWS's guidance is the cleanest statement: unlike search engines, RAG returns results directly from the LLM, which can bypass permission checks at the original data source. From a PM standpoint, "retrieval relevance" without "authorization correctness" is not enterprise retrieval quality.

Multi-tenant document freshness is a silent killer. If Tenant A's index is stale and Tenant B's is current, the same user question can look like a model failure in one account and a retrieval freshness issue in another. Triad dashboards need version metadata -- which knowledge base version, which sync timestamp, which tenant environment.

Multi-hop queries break simplistic triad thinking. When a question requires combining evidence from multiple documents, single-hop context precision can look fine while the system misses evidence scattered across three different policy documents. Context recall becomes the metric that catches these failures -- and most teams don't track it.

!

Common Mistake

Collapsing the triad into one blended "RAG quality" score.

Why it feels right: a single number looks executive-friendly, easy to trend, and easy to compare across experiments.

What actually happens: a system with context precision 0.88, faithfulness 0.96, and answer relevance 0.63 averages to 0.82 -- which looks fine. But the answer relevance score means users aren't getting their questions answered.

The fix: never collapse the triad into a single release number without also preserving the per-metric view. Gate release decisions on the worst individual metric, not the average.
Good Run vs Deceptive Run
Good Run
Context Precision0.88
Context Recall0.84
Faithfulness0.97
Answer Relevance0.89
Users get accurate, grounded answers to their actual questions.
Deceptive Run
Context Precision0.79
Context Recall0.58
Faithfulness0.96
Answer Relevance0.63
Accurately summarizing incomplete, wrong evidence. Citations look clean.
Average: 0.74 -- looks OK. Reality: users aren't getting answers. Never blend the triad.

Remember This

1. The RAG triad asks three independent questions: did the retriever fetch the right documents (context relevance), did the model stay grounded (faithfulness), and did the answer address the user's question (answer relevance)? Keep them separate. That separation IS the diagnostic value.

2. The failure pattern most teams misread: high faithfulness plus low answer relevance means the system is accurately summarizing the wrong evidence. The citations look clean. The user is still unserved.

3. In enterprise RAG, access control and document freshness are part of retrieval quality, not separate concerns. A triad dashboard without version metadata and authorization checks is incomplete.

Previous Topic Back to the Deep Dive