RAG and public health: 99% right on multiple-choice, but barely two-thirds of answers faithful to the official guidance

A UK team extends PubHealthBench, a set of 7,929 questions drawn from 687 UK Government public health guidance documents, to test Retrieval-Augmented Generation (RAG) — a technique that "grounds" a language model's answers in a corpus of documents fetched on the fly. On multiple-choice questions, supplying the right passages pushes accuracy to around 99%, so much so that small open-weight models beat a large proprietary one used without retrieval. But as soon as you move to free-form answers and check their faithfulness to the guidance, the best configuration holds up only 64% of the time — and the automated scoring tool collapses exactly on the "factual consistency" dimension.

The context

Large language models (LLMs: ChatGPT-style networks trained to predict text) score excellently on medical multiple-choice quizzes. But two obstacles make them risky in public health. First, hallucination: the model produces a plausible but false statement with the same confidence as a correct one. Second, staleness: official guidance changes — sometimes week to week during an outbreak — while a model's knowledge is frozen at its training date.

RAG (Retrieval-Augmented Generation) is the most common answer to both problems. The principle: before answering, the system searches an up-to-date corpus for the most relevant passages, then supplies them to the model as context. The answer is thus "grounded" in explicit sources rather than in the model's opaque memory. Attractive in theory; in practice, the final quality depends entirely on what you retrieve and how you evaluate it. An earlier benchmark, PubHealthBench, had shown that no model exceeded 75% on free-form answering to these questions, against a human score of 88% (with cursory search-engine use) and a ceiling of 92.5% for the best proprietary model, GPT-4.5, without retrieval. The paper decrypted here starts from that and asks: what do we actually gain by adding retrieval, and can we even measure it correctly?

The method

The authors turn PubHealthBench into a RAG benchmark. The 687 guidance documents are converted to markdown and split by header levels, producing a corpus of 5,358 fragments ("chunks"). The retrieval part compares three families of methods. Dense retrieval turns each fragment and each question into an embedding — a vector of numbers meant to capture meaning — and retrieves the fragments whose vector is closest to the question's; eight embedding models are compared, from NV-Embed-v2 (8 billion parameters) down to small encoders like MedCPT or EmbeddingGemma. Sparse retrieval relies on exact words, via TF-IDF and BM25, two classical methods that weight rare terms. Hybrid retrieval merges both rankings via Reciprocal Rank Fusion (RRF), with a tunable weight between dense and sparse. Three corpus variants are tested: full, summary (each fragment replaced by a summary produced by GPT-4o-mini), and reduced (only fragments over 512 tokens are summarized).

Retrieval quality is measured by Recall@k (does the right source appear in the top k results?), MRR (Mean Reciprocal Rank, rewarding a good result placed high), and nDCG@10 (a ranking-quality measure over the top ten). On the generation side, eleven LLMs are evaluated, from Llama-3.3-70B and MedGemma-27B down to Gemma-3-1B, first on multiple-choice questions. To get closer to real use, the authors isolate 760 expert-reviewed questions presented without options, in free form (up to 15 context fragments, capped at 10,000 tokens). Since one cannot hand-score thousands of free-form answers, they introduce an LLM-as-a-judge: a model (GPT-OSS-120B) grades each answer on four criteria — faithfulness (add nothing not in the source), completeness, clarity, factual consistency. Crucially, this judge is validated against two human annotators on 100 answers, with Cohen's kappa (κ, a chance-corrected agreement: 0 = chance, 1 = perfect agreement). One bias the authors flag: the benchmark was itself generated by Llama-3.3-70B, one of the models then evaluated.

The results

On retrieval, the ranking is clear. The best dense encoder, NV-Embed-v2, reaches a Recall@10 of 0.98, an MRR of 0.85, and an nDCG@10 of 0.88; hybrid pushes it to 0.99 / 0.88 / 0.91. Small encoders fall off badly — EmbeddingGemma and MedCPT drop to around 0.24–0.27 precision at rank one — but hybrid gives them back 5 to 10 points. Two practical lessons: beyond 700 to 800 words per fragment, ranking quality drops, and no model gets its best result on the "summary" corpus. Summarizing to condense degrades retrieval.

On multiple-choice questions, the effect of retrieval is spectacular. Providing the right context brings almost all models to around 99%: Llama-3.3-70B at 0.995, MedGemma-27B at 0.992, and even the small Gemma-3-4B at 0.956; only Gemma-3-1B lags (0.664). All but that one exceed the human score of 88%, and several beat GPT-4.5's ceiling of 92.5% without retrieval — a well-fed 8-billion-parameter open model outperforms a poorly-fed large proprietary one. The gain appears with a single context fragment and peaks around three to five.

Then comes free-form answering, and the picture flips. Completeness and clarity stay high (often 0.88 to 0.97), but faithfulness — adding no claim absent from the cited source — is met, for the best model, only 0.64 of the time; overall, only half to two-thirds of answers pass this criterion. And it degrades as the right source is retrieved lower in the list. In concrete terms: of three apparently well-written and "sourced" answers, about one adds material the cited guidance does not support. In public health, where advice reaches whole populations, an unfaithful addition — one extra dose, a forgotten exclusion — cannot be caught case by case. Worse, the automated judge meant to measure factual consistency does not reproduce human judgments there: its agreement with the annotators is near zero (κ of 0.06 and 0.17), whereas it is decent on faithfulness (κ 0.64 to 0.71) and completeness (κ 0.57 to 0.59). The instrument is blind exactly where the clinical stakes are highest.

What is good

An evaluation grounded in real guidance, and beyond multiple-choice. The benchmark rests not on artificial questions but on 687 official public health documents covering ten domains. Above all, the authors do not settle for the comfort of multiple choice: they build an expert-reviewed free-form set, the only form that resembles real use. It is that shift that reveals the gap between 99% on the quiz and 64% faithfulness.

A clear, reproducible practical result. Retrieval is the primary lever; a small, well-fed open model rivals, even beats, a large proprietary one without retrieval. The settings that work are documented (hybrid rank fusion, weighting, TF-IDF, unsummarized corpus), and the whole is published under an open CC BY 4.0 license — enough to replay and challenge the measurements.

Rare methodological self-criticism. The authors validate their automated judge against two human annotators, publish the agreements criterion by criterion, and state in black and white that the "factual consistency" score is unreliable and must therefore be read with caution. In a literature that readily sells "LLM-as-a-judge" as truth, this honesty about the instrument's limits is to their credit.

What is less good

A misleading metric up front. The spectacular figure — ~99% — is a multiple-choice score, and the authors themselves acknowledge it overestimates real capability: on multiple choice, a model can eliminate wrong options and scrape partial credit without really mastering the content. The proof is internal to the paper: high multiple-choice accuracy persists even when the retrieved fragment is not the right source. The real story is not the 99%, it is the 0.64 faithfulness in free form.

Evaluation circularity and a biased comparator. The benchmark was generated by Llama-3.3-70B — which then tops the ranking at 0.995. Evaluating a model on questions it produced itself is a form of leakage: the questions are "answerable by construction" from the corpus, never testing reasoning across several fragments or the ability to say "I don't know" on an out-of-corpus question. Add a biased comparator on the evaluation side: the LLM judge is unstable on factual consistency (near-zero agreement with humans), so the most critical dimension in health is, at bottom, not reliably measured.

A population bias and deployment fragility. Everything rests on the guidance of a single government — the UK — in English, at a population level: nothing guarantees these results transfer to another country, health system, or language. And the dominant failure mode is telling: the models tend to say too much, incorporating guidance beyond what is asked, which reduces precision and multiplies chances of misunderstanding. Controlling the scope of the answer — adding no unsolicited advice — remains the unsolved challenge, and it is precisely what separates a prototype from a deployable tool.

What it changes

For the research community, the message is methodological: evaluating a medical LLM on multiple choice is misleading, and free-form faithfulness evaluations, backed by judges themselves validated against humans, should be preferred. The paper maps the useful work ahead: automated judges reliable on factual consistency (currently missing), retrieval able to reason across several fragments, and test sets that include questions with no answer in the corpus, to check that the system knows how to refuse.

For clinicians and health agencies, RAG grounded in an up-to-date corpus is a credible path to align an assistant with official guidance without retraining the model — a real asset when guidance changes fast, notably in an outbreak. But nothing is deployable as is: as long as faithfulness plateaus around two-thirds and scope control is unresolved, such an assistant remains a supervised aid, not an autonomous source.

For patients and the general public, the lesson is reading hygiene. An assistant that "cites the official guidance" is not thereby faithful to what it says: in this study, roughly one free-form answer in three adds elements not supported by the displayed source. An excellent quiz score says nothing about the safety of advice written out in full — and it is the latter that matters when you seek health information.

To go further

The preprint is available on arXiv (10.48550/arXiv.2607.06641), under a CC BY 4.0 license. On the limits of LLMs in sensitive clinical contexts, see our decryptage on the clinical safety of LLMs at the edge of decision-making; on an LLM applied to diagnosis from clinical notes, that of GenoSolver for rare diseases; and on the effect of a model's answer format, that of GPT-4 in radiology and the formatting of its explanations.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.