**RAGAS** is a framework for evaluating retrieval-augmented generation (RAG) systems. A RAG pipeline has two halves to grade: the retriever, which fetches context, and the generator, which writes the answer over that context.
The four metrics split across those halves. **Faithfulness** checks whether the claims in the answer are actually supported by the retrieved context, so it catches hallucination. Answer relevance checks whether the answer addresses the user's question, regardless of context.
The other two grade the retriever. Context precision asks: of the chunks you retrieved, what fraction were actually relevant? Low precision means you pulled in noise. Context recall asks: of everything you needed, how much did you retrieve? Low recall means you missed important chunks.
The trick is option E. Faithfulness, answer relevance, and context precision can all be computed without a gold answer, using only the question, the retrieved context, and the generated answer. But context recall needs a reference answer to measure coverage against. So RAGAS is mostly reference-free, not entirely.
Option F is the correct reading of the precision-recall pattern: high precision plus low recall means clean but incomplete retrieval.
**RAGAS** evaluates a RAG pipeline on four metrics that split across its two stages. Two grade generation. Two grade retrieval.
Generation metrics. Faithfulness checks whether every claim in the answer is entailed by the retrieved context. Answer relevance checks whether the answer addresses the user's question, independent of context. They catch different failures: hallucination versus off-topic drift.
Retrieval metrics. Context precision is the fraction of retrieved chunks that are relevant. Context recall is how much of the needed information was retrieved at all. They trade off: a precise but incomplete retriever scores high precision and low recall.
That makes options A through D correct descriptions, and option F the correct reading of the precision-recall split.
Option E is the trap. Three of the four metrics are reference-free, computed from the question, context, and answer alone. Context recall is the exception. To measure how much of the answer was covered, you compare against ground-truth claims from a reference answer. No reference, no recall. So the suite is mostly reference-free, but calling it fully reference-free is wrong.
Quick reference
TL;DR
RAGAS scores RAG on faithfulness, answer relevance, context precision, and context recall. Three are reference-free; context recall needs a gold answer, so the suite is not fully reference-free.
Imagine grading a student's open-book essay. RAGAS checks four things. Did the essay only state facts that appear in the books it was handed (faithfulness)? Did it actually answer the question asked (answer relevance)? Were the books it pulled off the shelf useful, or did it grab junk (context precision)? And did it find ALL the books it needed, or miss some (context recall)? The first three you can grade just by looking at the essay and the books it used. But to check whether it missed any books, you need an answer key showing what a complete answer should contain. That answer key is the catch: it means RAGAS is not entirely answer key free.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Pipelines that are not retrieval-augmented at all, such as pure generation or classification The retrieval metrics have no context to score, so most of the RAGAS suite does not apply.
Continuous production eval where curating a reference set per query is infeasible Context recall needs gold answers, so you can only run the three reference-free metrics live.
Safety-critical factual correctness where grounded but wrong answers are dangerous Faithfulness measures grounding in context, not truth in the world. A wrong context yields a wrong answer.
Very high query volume where per-query LLM-judge cost dominates the eval budget Each metric is an LLM-judge call. Cost scales with traffic, so sampling or cheaper judges become necessary.
Quick scan · 2–4 min
Everything important, quickly.
5 min: the four metrics mapped to retriever versus generator, why three are reference-free and context recall is not, the precision-recall reading, and how this shapes production eval design.
Real-world examples
Real products, models, and research that use this idea.
1RAGAS is the de facto open-source RAG eval library in 2026, wired into LangSmith and Langfuse dashboards for continuous retrieval scoring.
2Teams run the three reference-free RAGAS metrics on live production traffic and gate context recall on a curated golden set.
3TruLens offers a parallel RAG triad (groundedness, context relevance, answer relevance) that mirrors the reference-free side of RAGAS.
4RAGAS uses an LLM judge such as GPT-5.5 or Claude Opus 4.7 to decompose answers into atomic claims for faithfulness scoring.
5Patronus and Galileo ship hosted RAG evaluators whose metric taxonomy closely tracks the original four RAGAS axes.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does RAGAS compute faithfulness without a reference answer?
Sign in or go Pro to see the approach for this follow-up.
QIf three metrics are reference-free, how would you architect a continuous RAG eval that still measures recall?
Sign in or go Pro to see the approach for this follow-up.
QWhy might a high faithfulness score still ship a wrong answer to users?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Calling RAGAS fully reference-free. Three metrics are, but context recall needs a gold answer to compute, so the full suite still depends on curated references.