Five Answers, One Bad Retrieval: When RAG Agreement Misleads

TL;DR for operators

A production RAG system returns the same answer five times. The operator still has to decide whether to release it, investigate it, or send it for review. Repeated agreement is useful evidence that generation is stable, but it does not show that the system retrieved the right material.

All five answers may have been generated from the same empty, incomplete, or incorrect context. In that case, repeated sampling does not independently test the answer; it repeatedly tests the decoder under one defective retrieval state. A wrong answer that remains effectively unchanged across samples is a silent error.

In the study’s five-sample results, silent errors account for 59% of dense-retrieval errors, 42% of adaptive-KG errors, and 84% of errors in the strict-KG stress tests. This is the operational risk of retrieval-state lock-in: consistency can remain high because every generation reuses the same bad evidence.

Answer variation, including SD-UQ, should therefore be treated as one diagnostic layer rather than a correctness certificate. Production gates should also inspect retrieved evidence and retrieval traces. When agreement is high but retrieval is empty, unsupported, or concentrated on a questionable source path, perturb retrieval or escalate the answer instead of approving it automatically.

Five identical answers can still be one repeated failure

Suppose a RAG application answers a customer, analyst, or clinician-facing question five times and produces nearly identical responses. An answer-dispersion metric will assign low uncertainty because the outputs occupy the same semantic region.

The operational decision then appears straightforward: release the answer, lower its review priority, or count it as a confident response.

Julka’s results show why that decision can be miscalibrated. Among incorrect answers, the proportion exhibiting effectively zero answer dispersion is substantial:

The pooled estimates have Wilson 95% intervals of 0.53–0.65 for dense retrieval, 0.36–0.47 for adaptive KG retrieval, and 0.79–0.89 for the strict condition.

These figures do not mean that 59%, 42%, or 84% of all answers are locked in. The denominator is wrong answered rows, not total system traffic. Nor does zero dispersion prove the proposed mechanism. It is a diagnostic signature that supplies an upper bound on the prevalence of retrieval-state lock-in.

That distinction is essential. A wrong answer can remain stable for reasons other than retrieval concentration, including strong parametric priors, narrow decoding behavior, or an evaluation label that fails to capture an acceptable alternative. The paper uses retrieval traces, stress conditions, perturbations, overlap analysis, and independent re-judging to investigate the mechanism, but the silent-error label itself is not a causal certificate.

Agreement measures the decoder under the context it received

A RAG response is generated through two linked processes: selecting context and producing an answer conditioned on that context. The paper represents the response distribution as:

Here, uncertainty can arise because retrieval returns different context bundles or because the generator responds differently to a fixed bundle.

When retrieval probability concentrates on one context $c^\star$, the mixture approaches:

Repeated sampling then probes variation in generation under that particular retrieval state. It does not independently test whether the state is complete, relevant, or correct.

Two failure forms follow.

First, empty-retrieval lock-in occurs when repeated retrieval produces no usable evidence. The generator may answer from parametric memory, yielding a stable response unsupported by the retrieval layer. In the strict RealMedQA condition, all 48 silent errors were empty-retrieval cases.

Second, presence lock-in occurs when retrieval repeatedly supplies populated but defective evidence. The answer may be coherent and consistent because the context itself is coherent. Detecting this form is harder: the system must determine not only whether evidence exists, but whether the retrieved route omitted a necessary source, selected the wrong entities, or assembled a misleading support chain.

The study provides stronger evidence for the absence case than for deep presence lock-in. That unresolved category matters most in systems where retrieved documents look plausible enough to satisfy superficial support checks.

Three diagnostic layers reveal different failures

The paper evaluates three signal families rather than collapsing confidence into one number.

Answer variation: SD-UQ

Question-conditioned spectral dispersion uncertainty, or SD-UQ, embeds five sampled responses, removes the direction associated with the question, and measures the remaining semantic variation through residual singular values.

Its intended role is answer-state diagnosis. Low SD-UQ indicates that sampled outputs have collapsed into similar semantic directions.

That is useful when errors create disagreement. It cannot rank wrong answers that all occupy one semantic cluster. Under strict retrieval concentration, its error-ranking performance can therefore collapse—not because its calculation is defective, but because the relevant uncertainty has moved upstream.

Evidence support: SEU

Support Entailment Uncertainty uses natural-language inference labels over retrieved chunks. Evidence that entails the answer reduces risk; contradictory evidence increases it; neutral evidence falls between them.

SEU asks whether the retrieved text supports the generated answer. It can expose cases where the answer is stable but conflicts with available evidence.

Its boundary is the reliability of the NLI model and the evidence set it receives. If retrieval omits the decisive document, an entailment checker cannot evaluate evidence it never sees. A set of mutually consistent but incomplete chunks may also produce reassuring support scores.

Retrieval structure: GPS

Graph Path Support measures whether entities linked from the answer are reachable from question entities at an expected reasoning depth in the provenance graph. Weak or absent structural support receives a higher risk score.

GPS is most useful when entities can be linked reliably and graph structure represents the task’s reasoning requirements. The paper reports useful behavior in its clinical calibration setting but weaker transfer across several open-domain datasets. PubMedQA’s label-style answers, for example, do not expose linkable answer entities, so the conjunctive rule selects no cases there.

GPS should therefore be read as a domain-conditional structural diagnostic, not a truth score.

The strict tests expose a regime; they do not isolate one cause

The dense and adaptive-KG comparisons hold the corpus, embeddings, generator, and question subsets fixed. Across the six snapshots, adaptive KG does not produce a statistically detectable accuracy difference from dense retrieval. Its point estimate is lower on every dataset.

This prevents a common misreading: the paper does not demonstrate that KG-RAG is more accurate.

The strict graph-only experiments serve a different purpose. They force a more concentrated retrieval regime and reveal how answer-only uncertainty behaves when retrieval repeatedly collapses into empty or restricted states. They are mechanism-oriented stress tests.

They are not clean ablations of graph structure. The strict condition changes multiple components, so the resulting 84% silent-error rate cannot be attributed to one isolated design choice. It identifies a failure regime in which stable wrong answers become common.

The robustness work—route logging, retrieval-overlap analysis, perturbation replays, sensitivity checks, bootstrap intervals, and independent re-judging—supports the diagnostic interpretation. These checks strengthen the claim that retrieval state contributes material information. They do not convert the design into a causal estimate of any single retrieval component.

A precise gate can still miss most errors

The paper combines answer, evidence, and retrieval signals into a conjunctive audit rule for adaptive-KG runs.

Across 1,117 answered rows, the rule selects 86 answers:

RealMedQA produces the strongest result: 48 of 223 answered cases are selected, with 100% automated precision and a 95% interval of 0.926–1.000. Other datasets select as few as two, four, or six answers, making their precision estimates trace-level rather than stable deployment statistics.

The rule is best understood as a selective audit mechanism. It may justify automatically blocking or escalating a narrow subset when false accusations are expensive and review capacity is limited.

It does not justify labeling the remaining 92.3% of outputs as safe. Low coverage and low recall mean that most cases pass through untouched, including most errors.

A production team evaluating such a gate should report all three metrics together. Precision without coverage hides how rarely the rule acts. Coverage without recall hides how many failures remain. Recall without precision hides the burden imposed on reviewers.

Production systems need retrieval observability, not another confidence scalar

The paper directly shows that answer agreement alone misses many wrong outputs in the tested system and that evidence- and retrieval-side diagnostics recover additional signal. It also shows that one transparent multi-signal rule can identify a small subset of errors with high precision.

Cognaptus draws four operational implications.

First, preserve the retrieval state as an auditable object. Store retrieved chunk identifiers, rankings, entity links, graph paths, route type, fallback decisions, empty results, and abstentions. Without these records, teams can inspect the final answer but cannot reconstruct the state that stabilized it.

Second, treat conflicting signals as a trigger for intervention. A low-dispersion answer with contradictory evidence, an empty retrieval result, or an unsupported graph path should not inherit a generic confidence label. The appropriate action may be retrieval perturbation, a second retriever, a broader search policy, or human review.

Third, separate certification from diagnosis. A high-precision gate may be suitable for deciding which answers to block. A broader collection of lower-precision signals may be more useful for incident investigation, retriever tuning, and dataset discovery. One threshold rarely serves both purposes.

Fourth, cascade checks according to cost and exposure. Embedding-based answer dispersion is comparatively cheap. NLI across retrieved chunks adds inference cost. Graph replay and entity linking add infrastructure and may fail outside calibrated domains. Teams can run inexpensive checks broadly and reserve deeper retrieval audits for cases with higher business impact or conflicting signals.

These are design inferences from the reported evidence, not validated return-on-investment estimates. The paper does not measure review costs, latency budgets, user harm, or the financial value of prevented errors.

The remaining uncertainty sits in populated retrieval states

The study covers six fixed QA snapshots and 1,180 planned questions using OntoGraphRAG v1.0.0, GPT-4o-mini, all-MiniLM-L6-v2 embeddings, an NLI evidence scorer, and Neo4j provenance graphs. Provider failures reduce some clean-analysis denominators.

This scope supports a fixed-system diagnostic claim. It does not establish that the reported rates transfer to other generators, embedding models, retrieval stacks, sampling budgets, or production corpora.

Three boundaries affect practical adoption.

First, the correctness labels are automated. The independent re-judge adds robustness, but high-stakes certification still requires human validation against domain-specific acceptance criteria.

Second, both SD-UQ and SEU depend on model and configuration choices. Changing the response encoder, sampling temperature, sample count, NLI model, or chunking policy may change their calibration.

Third, graph support is unavailable or weak when answers lack linkable entities or when the graph does not encode the task’s relevant structure. The uneven cross-domain GPS results make system-specific calibration mandatory.

The hardest open problem is populated retrieval that is plausible, internally consistent, and still wrong. Empty retrieval leaves an observable absence. Deep presence lock-in may supply enough apparent evidence to satisfy answer, entailment, and structural checks while omitting the decisive fact. The paper identifies the category but does not establish a general detector for it.

Stable answers should begin an audit, not end one

Repeated agreement remains informative. It tells an operator that the generator is stable under the states it repeatedly encountered.

The error is attaching a larger meaning to that observation. In RAG, answer consistency does not establish evidence adequacy or retrieval correctness.

Julka’s decomposition gives teams a more defensible inspection model: assess the answer surface, the supporting evidence, and the retrieval state separately. The reported gate shows that these signals can support precise selective rejection. Its low coverage and recall define the boundary just as clearly.

For operators deciding whether to release, investigate, or review a stable answer, the relevant question is not only whether the model changed its mind. It is whether the system ever changed what the model was allowed to see.

Cognaptus: Automate the Present, Incubate the Future.