clinicians.build · interactive · august 18, 2026

The Retrieval Tax

Eight frontier models answered 304 pediatric cardiology board questions. The top score was 98.4% with retrieval switched off. Turning on a single‑textbook RAG pipeline added almost nothing — and at that ceiling, on that many questions, it could not have shown you if it had.

Primary source: “Evaluating retrieval‑augmented large language models for pediatric cardiology knowledge using standardized questions,” npj Digital Medicine, 17 Aug 2026
Everything on this page below the two source figures is computed live in your browser from binomial statistics. No per‑model results are shown, because the paper’s per‑model table was not available to verify.

Boston Children’s wrapped eight frontier models in a pediatric cardiology chatbot and ran them against 304 board-style questions. The top models hit 98.4% with no retrieval at all. Enabling a single-textbook RAG pipeline barely moved the needle.

The interesting part is not that RAG failed. It is that on a benchmark shaped like this one, RAG was never going to be able to succeed visibly. Here is the whole benchmark, one square per question.

304 q
98.4%
+1.0 pp
Answered correctly Wrong — all the headroom RAG has
Questions
304
Wrong answers left
5
95% CI half‑width
±1.5
Smallest lift detectable
+1.6 pp

Where the claimed lift disappears

The curve is the smallest true improvement a benchmark of a given size could actually detect — 80% power, two-sided, at the baseline accuracy you set. The horizontal line is the lift you are claiming. Anywhere the line sits under the curve, your benchmark cannot tell that improvement apart from noise, no matter what the table says.

And the questions are the real problem

Statistical power is the polite objection. The impolite one is that a board question and a patient are not the same object.

A board-style question

One pre-digested paragraph. Every fact you need is present, nothing irrelevant is. One correct answer, written by someone who knew it before they wrote the stem. No missing labs, no contradictory notes, no family in the room.

~1 document · 0 conflicts · 1 right answer

A real encounter

Years of notes from clinicians who disagreed, labs drawn at the wrong time, imaging read twice, an outside-records blob nobody reconciled, a medication list that is wrong in a way the chart cannot show. The answer may not exist yet.

hundreds of documents · conflicts everywhere · answer unknown

Retrieval is a machine for finding the relevant thing in a pile. A board question has no pile. So a benchmark of board questions is close to the worst possible instrument for deciding whether retrieval is worth building — it removes the exact condition retrieval exists to handle, then reports that retrieval did not help. The npj result is real and worth knowing. It is evidence about this benchmark, not about your chart.

If you can’t name the failure mode your retrieval layer fixes, you built a cost center with a citation feature.

Which leaves the honest case for RAG intact and much narrower than the pitch deck: you build retrieval when you need the answer to point at a source someone can open during a chart review. You are buying defensibility, not accuracy. Price it that way and the decision gets easy.