Why Your Biomedical RAG Is Hiding Contradictions From You – Unite.AI

Standard biomedical RAG resolves conflicting evidence silently in roughly three of every four queries. The fix is structural, not informational — and most of the upstream complexity teams add does not help.

Ask a retrieval-augmented clinical assistant a simple question — does melatonin help with jet lag? — and the literature it retrieves will not agree with itself.

The Cochrane review concluded that melatonin is remarkably effective, with eight of ten included trials showing reduced jet lag on flights crossing five or more time zones. A randomized, double-blind trial of 257 Norwegian physicians in the American Journal of Psychiatry found no benefit from any of three melatonin regimens.

Both documents are real. Both are methodologically serious. Any clinician deciding what to recommend needs to know they disagree.

In controlled testing, the synthesis usually did not say so. It produced a fluent, correctly cited paragraph that came down on one side and never signaled that another side existed.

That is contradiction blindness. On a contradiction-paired biomedical benchmark, it occurred in 72.6 percent of queries (95% CI 0.697–0.755). The output read normally. The citations resolved correctly. Standard quality checks passed. The disagreement simply disappeared.

The failure mode that looks like success

While there is no direct convergence in biomedically related evidence, studies show conflicting results between the use of observational studies versus randomized controlled trials (RCTs) and also varying methods used for different patient populations; however, a study may also include both strong and weak methodology, which would seem to carry the same weight.

This is not a unique case. Ioannidis’s analysis of highly cited clinical research found that 16 percent of the most-cited studies were later contradicted outright, and that observational studies were far more likely than randomized trials to be contradicted or to have their effects revised downward — five of six, against nine of thirty-nine. Thus, the issue is not just disagreement among studies but rather a structural part of the literature itself.

Therefore, as a critical function of any biomedical retrieval-augmented generation system, generating evidence about disagreements between studies should be a primary goal. However, most systems fail to accomplish this task. Using the HealthContradict benchmark’s 920 contradiction-paired examples demonstrated that a standard retrieval-augmented generation (RAG) failed to generate the disagreements in approximately 73% of the tested pairs. The problem is not retrieval. The system retrieves both sides. It then resolves the conflict silently, in favour of one of them.

A manual examination of 50 stratified failure cases produced a pattern that I have referred to as epistemological flattening. Both documents are cited individually by the model, and its rendering of the conflict is described in terms of confidence gradients (“anecdotal evidence … scientific studies indicate…”) as if no conflict existed. Less than 5% of these failure cases used the words “contradiction,” “conflict,” or “disagreement” at all. The output generated had a citation accuracy rate of .99. That is precisely the point: citation accuracy does not imply contradiction awareness. A system can attribute every sentence perfectly and still tell a clinician something the evidence base does not support.

Three features of RAG’s “synthesis” create a dangerous clinical environment.

Inverting the Value Proposition. The purpose of RAG is to display evidence relevant to clinicians. Therefore, by removing the aspect(s) of that evidence which are most important for clinicians to review, it actually accomplishes its opposite.

Creating Invisibility. Since there is no hedge, or truncation, or format artifact; a “flattened” synthesis and a faithful synthesis appear indistinguishable on the screen.

Compounding Silently at Scale. Nothing in a standard evaluation suite flags it, so a system can run in production for months while every conflicted query is quietly resolved in one direction.

Information is not the lever

An obvious solution to this problem would be to inform the model. Clearly, if the failure mode was that the model did not identify that two documents were conflicting, you could simply add a “SUPPORT” or “CONTRADICT” stance label to each of the retrieved documents. This should obviously improve performance for the model. It did not.

In an experiment where we added stance labels (by annotating) to explicitly give the model the information that two documents conflicted, we found that explicit stance annotation moved contradiction blindness from 93.8 percent in the unannotated control arm to 91.4 percent — a difference with overlapping confidence intervals and no practical significance. Although the model received the information that two documents conflicted, its default response distributions remained unchanged.

Understanding the mechanism is useful here. Information added passively to a prompt does not change a model’s default response distribution. For contradiction-laden biomedical questions, that distribution strongly favours a single consolidated position. Stance labels become extra tokens — often recycled into section headings — while the prose reverts to type.

Structure is the lever

What worked was structural, as opposed to being informational; instead of providing the model with all that is factual or correct about the documents, the structure required the synthesis to be constructed in terms of possible disagreements (Agreement, Disagreement, Evidence Quality, Conclusion). The scaffolded prompt provides the model with no more information than does the un-annotated control. The only difference is what the model has to do.

There were reductions of roughly nineteen fold — from 93.8 percent to 4.9 percent — in contradiction blindness without any retraining, without any pipeline modifications, without any increased latency, and without any increases in per query costs.

This is not improved reasoning. This is simply forced allocation. Once the model must provide a Disagreement section, it goes back through the documents it initially retrieved looking for something to include in this section.

Structured prompting is less about enhancing the model’s ability to reason and more about providing light-governance over how the generation behavior of the model allocates output space. It forces the model to spend some amount of its output space on disagreement (the difference between information that is available and information that needs to appear in order to meet the requirements).

This changes a common architectural assumption. Most proposed RAG enhancements add information to the context: more documents, retrieval scores, stance tags, evidence-grade metadata. Unless the synthesis prompt has specific requirements for that information to shape the output structure, most will not change the model behaviour. Changing input shifts mass at the margins; changing required output structures changes distributions directly

Why the expected helpers don’t help

Two elements commonly viewed as essential to contradiction-aware biomedical RAG provided very little when tested with controlled ablation.

1) PICO Decomposition. PICO (Population, Intervention, Comparison, Outcome) is an organized way of breaking down clinical questions into component parts for use in evidence based medicine. The hypothesis was that decomposing claims into PICO fields would constrain scope drift and improve contradiction detection across heterogeneous evidence. Removal of this level resulted in output nearly indistinguishable from that generated by the complete system. While PICO can be useful for organizing your thoughts during clinical decision making; as an inferred input at the time of analysis to support contradiction detection, it has no measurable contribution.

2) Explicit Stance Classification. In contrast to PICO removal, explicit stance classification performed poorly but differently. On clean academic corpora, it produced small gains on some metrics. However, when analyzing noisy web text – which is typically how consumer facing health apps will have access to information — the classifier returned NOT_ENOUGH_INFO for most documents, primarily because authors of consumer facing health content do not normally declare their position using declarative language expected by the classifier. Consequently, those labels as uninformative were then propagated into the synthesis context as noise. Removal of the stage for noisy corpora marginally improved contradiction detection.

The real bottleneck is upstream signal quality

The most important conclusion of this study is that the ability to recognize contradictions in sources is limited by how accurate the information (data) about these sources is, more so than how they were built or structured.

Evidence-quality utilization — the proportion of cases where the synthesis preferentially cites the higher-quality source when both are available — was 0.83 on clean scientific abstracts and fell below 0.15 on noisy web retrieval. Semantic citation faithfulness followed the same pattern, from 0.66 on abstracts to 0.45 on consumer-health content.

Identical pipelines, different corpora. Contradiction handling is capped by the explicitness of the signals in the source documents.

There is a structural reason for this. Real-world retrieved documents frequently lack the cues any classifier or weighting mechanism depends on: publication-type metadata, explicit stance statements, methodology descriptions. Abstracts of peer-reviewed articles make assertions declaratively regarding specific populations, methodologies, and results. The exact same assertions are made within anecdotal, persuasive, and hedging language in consumer health articles. Therefore, identical systems produce dramatically different behavior downstream based upon what they receive as input.

This is also supported by the largest expert assessment of medical RAG ever published, in which eighteen medical experts contributed 80,502 annotations across 800 model outputs. Only 22 percent of top-16 retrieved passages were relevant. Standard RAG degraded factuality and completeness relative to non-RAG baselines. Retrieval and evidence selection, not generation, were the dominant failure points — an echo of what benchmark work on medical RAG has been reporting for two years.

Although there is a relatively limited implication of this finding in terms of application, it can be said conclusively that the ability of your system to handle contradictions when it retrieves from PubMed abstracts cannot be transferred to a consumer facing website regardless of how advanced the rest of your system is.

Evaluation’s own blind spot

Measuring contradiction handling is more difficult than it seems, and even a standard approach to measuring contradiction handling has its own set of problems.

LLM-judge scoring represents the most common way that open-ended RAG output is scored for conflict handling and also diverges from structural acknowledgment checks in terms of how they are developed. Prompt strategies which organize synthesis into PICO fields receive high scores from judges while failing explicit acknowledgment tests outright. This is expected: LLM judges have been documented as favouring longer, more elaborate responses and will also view a buried caveat in the outcome section as thoroughness when nothing in the prose level references the conflict.

Retrieval strategy introduces a second pitfall. Random retrieval produces a contradiction blindness rate of zero — a synthesis cannot fail to acknowledge a contradiction its retrieved set does not contain. Maximal marginal relevance retrieval, on the other hand, reduced semantic citation faithfulness down to .11. Each looks acceptable under one single contradiction-handling metric, but both fall short under paired evaluation.

So pair the metrics. Run three checks on every synthesis: a deterministic structural check for expressed language that acknowledges a conflict; an LLM judge for depth and balance; and a semantic citation check confirming that cited content matches its source. Each identifies what the others do not. None can be trustworthy by themself as a benchmark for contradiction handling.

Four actions for this quarter

1. Test your baseline. Every biomedical, clinical, regulatory RAG system in production must be tested for contradiction blindness before further deployment. To perform the test, you need a contradiction-paired test set and a per query check that the output signals substantive disagreement. A 72.6 percent baseline is not a rounding error.
2. Implement structured synthesis prompts. The Four (required) sections – agreement, disagreement, evidence quality, conclusion – force output bandwidth on disagreements. There’s no new infrastructure required; no training required; no per query cost; one change produced nineteen-fold reduction!
3. Don’t use MMR retrieval for contradiction-sensitive queries. Although MMR uses diversified documents for topical coverage, it does not optimize for stance coverage — which is incorrect when the goal is to retrieve both sides of a disagreement. Therefore, bm25 plus embedding retrieval should be used as a safer default. However, stance aware diversification would be the direction this work points toward.
4. Pair your evaluation metrics. Retire the single LLM-judge score. Combine it with a structural check for substantive content under acknowledgment headings and a semantic citation check verifying that the cited evidence supports the stated claim.

The broader point

The dominant instinct in RAG development is additive: better retrievers, better rerankers, richer metadata, longer context windows. The industry conversation about why RAG pipelines still fail in production has largely followed the same shape. For contradiction handling, the effective move was subtractive — constraining what the system is allowed to output rather than expanding what it is given. A single four-part synthesis prompt outperformed a five-stage pipeline. It did so by changing what the model had to produce, not what the model could see.

That does not make architecture irrelevant. Retrieval sets a ceiling, random retrieval makes synthesis meaningless, and poor evidence quality caps everything downstream. This is why the question of whether RAG systems solve the problem of reliability continues to resurface. However, what determines whether conflicting evidence is represented as conflicting turns out to be later in the pipeline and cheaper to change than many of the other components, which means that changes needed to improve reliability can occur sooner.

The expensive work — better contradiction datasets, explicit disagreement training signals, stance-sensitive retrieval, contradiction-native benchmarks — still needs doing, and will take considerably longer.

Therefore, if you want to know whether your system surfaces contradictory evidence as contradictory, you should ask yourself the uncomfortable question and find the answer this week: Does your system tell users when its evidence contradicts?