Human Evaluation for RAG Systems: Groundedness, Citations and Relevance

Why RAG Systems Need Their Own Evaluation Framework

A RAG system is not just a language model — it is a model plus a retrieval pipeline. Failures can originate at either stage: the retrieval layer can return irrelevant or misleading documents, and the generation layer can fail to stay grounded in whatever documents were retrieved. Standard LLM evaluation, which assesses generated outputs against a quality rubric, does not decompose these failure modes.

This decomposition matters because the remediation is different. A retrieval failure requires work on the embedding model, chunking strategy, or retrieval index. A groundedness failure requires work on the generation model, context window handling, or prompting strategy. An evaluation that only scores the final output — without tracing whether the problem lies upstream in retrieval or downstream in generation — cannot guide either fix.

The result is that teams working on enterprise RAG deployments — legal document review, medical knowledge bases, internal policy assistants, financial advisory tools — need a dedicated evaluation protocol that assesses all three layers. Our RAG system evaluation services are structured around this three-layer decomposition from the start of every engagement.

The Three Core RAG Evaluation Dimensions

Dimension 1: Retrieval relevance

Retrieval relevance measures whether the documents returned by the retrieval pipeline actually contain information relevant to the query. This is assessed before the generation layer is even reached: given the user's query, did the retrieval system return documents that a knowledgeable person would consider useful?

Retrieval relevance evaluation requires raters who understand the query domain well enough to judge what "relevant" means. For a legal knowledge base, a retrieved document about a different jurisdiction may be topically related but legally irrelevant. For a medical knowledge base, a document about a similar but distinct condition is similarly misleading. Domain expertise is required to make this distinction.

Automated retrieval relevance metrics — typically based on embedding similarity or BM25 scores — measure surface-level topical overlap, not domain-appropriate relevance. They consistently overrate documents that are terminologically similar to the query but substantively unhelpful, particularly in specialised domains where jargon is dense and semantic overlap between relevant and irrelevant documents is high.

Dimension 2: Groundedness and faithfulness

Groundedness (also called faithfulness) evaluates whether each factual claim in the generated response is supported by the retrieved documents. A fully grounded response contains no claims that cannot be traced to the content of the documents that were retrieved. An ungrounded response introduces information — even accurate information — that the model has generated from its parametric knowledge rather than from the retrieved context.

The RAGAS framework (Es et al., 2023 — "RAGAS: Automated Evaluation of Retrieval Augmented Generation") introduced automated faithfulness scoring for RAG systems using an LLM to decompose responses into atomic claims and then assess each claim against the retrieved context. This approach works well for clear-cut groundedness failures, where a response introduces information that is obviously absent from the retrieved documents.

Where automated faithfulness scoring fails is on the boundary cases that matter most in high-stakes deployments: claims that are paraphrased slightly beyond what the source document says, implicatures that follow plausibly but not conclusively from the retrieved text, and scope extensions where the model applies a finding from a specific context to a broader claim than the document supports. These require human judgement about the meaning and scope of the source, not just string-matching against retrieved text.

Dimension 3: Citation accuracy

Citation accuracy evaluates whether explicit source attributions in the generated response are correct. When a RAG system says "according to section 4.2 of the policy document" or "the 2024 clinical guidelines state that...", citation accuracy evaluation checks whether that attribution is accurate — not just whether a citation is present.

This dimension is independent of groundedness. A response can be highly grounded — all its claims are supported by the retrieved documents — but cite them incorrectly, attributing a claim from Document A to Document B. Conversely, a response can correctly cite its sources while still paraphrasing them inaccurately. In regulated industries where source traceability is a compliance requirement — legal AI, medical AI, financial AI — citation accuracy is a first-order quality criterion, not a secondary concern.

Citation accuracy assessment requires a human to read the cited document and verify that the attributed claim actually follows from it. This is not automatable with current techniques, because the assessment requires understanding of the source document's meaning and scope, not just its text. It is also where domain expertise matters most: assessing whether a legal citation is accurate requires understanding the legal instrument being cited, not just comparing strings.

What Automated RAG Metrics Miss: The Evidence

The gap between automated and human RAG evaluation is not hypothetical. Research on the reliability of automated RAG metrics consistently finds that they correlate well with human judgement on straightforward cases and diverge significantly on the cases that matter most for production deployments.

A systematic analysis by Saad-Falcon et al. (2023) in "ARES: An Automated Evaluation Framework for Retrieval-Augmented Generation Systems" found that automated faithfulness metrics correlate moderately with human judgements on general-domain tasks (Pearson correlations in the 0.6–0.75 range) but substantially less well on domain-specific tasks where expert knowledge is required to assess claim accuracy. The same study found that the weakest correlation was on "boundary paraphrase" cases — exactly the subtle groundedness failures that matter most in high-stakes deployments.

A further limitation identified in Gao et al. (2023) "Retrieval-Augmented Generation for Large Language Models: A Survey" is that automated faithfulness metrics struggle with what the authors term "retrieval hallucinations" — cases where the generation model correctly retrieved a document but then confidently applied information from it to a context the document did not address. This is a particularly common failure mode in enterprise RAG, where users ask questions adjacent to but not directly covered by the knowledge base, and the model extrapolates from the closest retrieved material rather than acknowledging the limit of its knowledge.

The practical implication for enterprise teams: automated RAG metrics are appropriate for high-volume regression testing — catching when a system that was passing a quality bar starts failing it after a change. They are not appropriate as the final quality gate before deploying a RAG system that handles medical, legal, financial, or other domain-specific content where hallucination has real consequences.

Case Study: Evaluating a Legal Knowledge Base RAG System

An anonymised project for a professional services firm deploying a RAG-based internal policy assistant illustrates how automated and human evaluation produce different quality pictures — and how different the remediation is.

The initial automated evaluation using a RAGAS-style faithfulness scorer rated the system at 88% faithful and 84% relevant on a 500-item test set. These scores were considered acceptable for a pilot rollout. A supplementary human evaluation — 300 items evaluated by two qualified legal professionals, each working independently — found a different picture.

Human raters identified three classes of failures that automated scoring had missed: (1) scope extension errors, where the system correctly cited a policy clause but applied it to a broader situation than the clause addressed — present in 14% of responses; (2) jurisdiction misattributions, where the system retrieved a document from a different regulatory jurisdiction and generated a grounded but inapplicable response — present in 9% of responses; (3) citation inversions, where the system attributed a claim to the correct document but to the wrong section — present in 7% of responses.

The overall human-assessed groundedness was 69% — compared to the automated score of 88%. More importantly, the failure categorisation allowed the engineering team to address each class separately: scope extension errors responded to prompting changes; jurisdiction misattributions required metadata tagging of documents by jurisdiction; citation inversions required changes to the citation-generation logic.

After targeted fixes, a follow-up human evaluation rated the system at 84% grounded — a material improvement that moved it into the range appropriate for internal use. The automated score had approved the system before it was ready; the human evaluation identified what specifically needed to change.

Designing a Human RAG Evaluation Protocol

A human RAG evaluation protocol differs from standard LLM evaluation in one critical respect: raters must be shown the full context — query, retrieved documents, and generated response — not just the response. A rater assessing groundedness without access to the retrieved documents cannot determine whether a claim is grounded or hallucinated. This three-part evaluation item (query + retrieved documents + response) is the fundamental unit of RAG evaluation.

Separate rubrics for each dimension

The three evaluation dimensions require separate rubrics because they assess different things. A retrieval relevance rubric asks: "Given this query, does this set of documents contain information that would help answer it?" A groundedness rubric asks: "For each claim in the response, can you find the supporting passage in the retrieved documents?" A citation accuracy rubric asks: "Does the cited source actually say what the response attributes to it?"

Collapsing these into a single overall quality score produces a number that cannot guide improvement. A composite score of 75% could reflect excellent retrieval with poor groundedness, or excellent groundedness with poor citation accuracy — and these have completely different fixes.

Rater qualifications per domain

Retrieval relevance and coarse-grained groundedness assessment can often be performed by trained generalists with strong reading comprehension skills. Fine-grained groundedness and citation accuracy, in any specialised domain, require subject-matter expertise. A legal RAG evaluation needs raters who can identify a scope extension error in a contract clause. A medical RAG evaluation needs raters who can tell whether a clinical claim stays within the scope of the cited guideline.

For the enterprise RAG evaluation projects we run, domain-expert raters are matched to content type: legal professionals for legal and compliance content, registered clinicians for healthcare content, CFA-qualified or equivalent for financial content. Generalist raters handle the high-volume retrieval relevance layer; domain experts handle the fine-grained groundedness and citation accuracy layer.

Integrating RAG evaluation into your release cycle

The highest-value use of human RAG evaluation is as a release gate, not a one-off validation exercise. Each change to the RAG pipeline — a retrieval model update, a new document corpus, a prompting change — can shift groundedness or citation accuracy in unexpected directions. A standing evaluation protocol that runs against a consistent test set at each release surfaces these regressions before they reach users.

For data quality processes upstream of RAG evaluation — ensuring the documents in your knowledge base are themselves high quality and correctly formatted — our data QA and validation services apply similar structured quality processes to the source document corpus before it enters the retrieval index.

For teams building the broader evaluation infrastructure around their LLM and RAG systems, our post on building LLM evaluation rubrics that annotators can consistently apply covers rubric design principles that apply equally to RAG evaluation tasks. For teams also evaluating model quality beyond the RAG layer, our LLM evaluation vendor checklist provides a structured framework for assessing evaluation partners.

The combination of automated screening (RAGAS-style metrics for high-volume regression testing) and human evaluation (for release gates and domain-specific quality assurance) is the production-grade approach. Neither replaces the other: automated metrics catch the obvious failures quickly, and human evaluation catches the subtle failures that matter.

Frequently Asked Questions