To what extent can RAG improve the factual accuracy of generative AI?

Lewis et al.'s RAG combined the model's internal parametric memory with an external memory consisting of a dense vector index of Wikipedia. It generated responses conditioned on search results for knowledge-intensive tasks, achieving state-of-the-art performance on three open-domain question-answering benchmarks at the time, and reported that the generated texts were more specific and factual than models relying solely on parameters.[1]

These results demonstrate that updatable knowledge sources and citation candidates can be brought into the generation process. However, the experiments were based on models of that era, a Wikipedia-centric index, and research tasks; the same effects are not guaranteed in business environments containing internal abbreviations, permission-restricted documents, mutually contradictory regulations, and procedures under revision.[1]

Transferability is determined not by the similarity of knowledge sources but by operational conditions. We must first confirm whether necessary documents exist in the index, can be retrieved with the questioner's permissions, whether a single valid version can be selected, and whether contradictory documents can be identified. In areas where any of these cannot be satisfied, we treat the issue as a lack of knowledge curation rather than RAG answer quality and halt the expansion of answer generation.[1]

RAGTruth collected approximately 18,000 responses generated by multiple LLMs from standard RAG settings including question answering, data summarization, and news generation, and performed manual annotation at both the instance and phrase levels. It is a corpus designed on the premise that even when search context is provided, responses may contain claims unsupported by the retrieved content or contradictory to it.[2]

Therefore, evaluation tables should not end with a single accuracy rate. We distinguish whether necessary documents entered the candidates, whether the selected passages support the question, whether each claim in the answer can be derived from those passages, and whether the citations are actual evidence. Simply counting search failures as generation model failures prevents us from identifying the components that need correction.

To eliminate alternative explanations for errors, we separate tests that fix search results to compare only the generator from tests that fix ground-truth evidence to score only the answers. If differences appear in the former, the issue lies in generation or citation; if unsupported claims remain in the latter, faithfulness is the primary cause. Rather than relying solely on the product of retrieval success rate and answer accuracy, we treat unsupported and contradictory claims as separate stop indicators.[2]

Liu et al. examined long-context models in two control tasks: multi-document question answering and key-value search, varying the position of relevant information. Many models performed well when relevant information was at the beginning or end but dropped significantly when placed in the middle. Having a long context window does not mean it can be used with equal accuracy across its entire span.[3]

In RAG, packing more top candidates may seem reassuring, but irrelevant fragments, duplicates, and positional biases become new failure factors. Chunk size, number of candidates, re-ranking, and evidence ordering should not be fixed values; they must be adjusted based on question type using both search recall and final faithfulness.

The impact of evidence position can be measured via counterfactual tests that move the same fragment to the front, middle, or back of the candidate list. If answers or citations change solely due to order, the improvement may not come from adding candidates but from position dependency. For such questions, we reduce candidates, re-rank, or revert to extracting evidence first, excluding them from automatic answers until they withstand order changes.[3]

RAGTruth addresses the persistence of unsupported and contradictory claims despite available evidence, while long-context research shows conditions where utilization results vary by position even with the same evidence.[2] [3] Therefore, in addition to scoring accuracy, we implement counterfactual evidence perturbation tests that intentionally alter retrieved candidates to verify whether answers follow the evidence. Test versions are also linked to answers.

In perturbation tests, we apply conditions such as removing a primary evidence piece, adding irrelevant fragments, swapping candidate order, and providing only ground-truth evidence to the same question. We first specify which conditions should change the answer and which should not, and we do not summarize results into search success or final accuracy rates but keep them separate by condition. Each condition is executed with the same model version and generation settings to isolate the source of differences.

If the assertion does not change even when primary evidence is removed, multiple explanations remain, such as internal model knowledge, other fragments, or question guidance. We do not immediately label it as hallucination but verify the correspondence between cited fragments and claims. Questions where conclusions reverse solely due to candidate order are re-tested under different conditions, including candidate reduction and evidence extraction, before being returned to automatic answers. If explanations compete, we withhold the answer.

We divide answers into individual claims and classify each as directly supported, conditionally supported, contradictory, or unsupported. For claims using multiple fragments, we record which combination is necessary for the claim to hold. We do not treat increasing citation counts as quality improvement and also check that unnecessary citations do not obscure assertions. If a claim is too large, we split it and re-evaluate.

In conditional support, we retain qualifiers in the answer such as period, target, exception, and unit, which change the meaning if removed from the evidence. Claims wider than the range indicated by the evidence fail even if the sentence is natural. If mutually contradictory candidates are provided, we do not have the generator integrate them but withhold the answer as a state where neither claim can be determined. Summaries without qualifiers are treated as separate claims.

After adjudication, we compare answers with only unsupported claims removed against answers regenerated from the necessary support set. If the former passes, suppression of extra claims suffices; if not, reconstruction of evidence utilization is needed. Instead of returning corrections to the searcher in bulk, we return the perturbation conditions and adjudication results as pairs to the responsible process, also saving the claim differences before and after correction.

• PERTURB: Fix conditions for removal, addition, order change, and ground-truth evidence
• EXPECT: Specify claims that should change and those that should remain invariant for each condition
• ADJUDICATE: Adjudicate direct support, conditional support, contradiction, and no evidence
• MINIMIZE: Save the necessary support set and qualifiers to establish each claim

The first gate is search. We confirm that the correct version of the document is retrieved for representative questions and that unauthorized or expired documents are not mixed in. The second gate is evidence utilization. We confirm that, given the correct fragments, no unsupported supplements are added and contradictory documents can be identified. The third gate is business outcomes, measuring the answerer's confirmation time, rejections, major errors, and appropriateness of withholding.

In production, we record the question, search candidates, adopted evidence, answer, model and index versions, and reviewer corrections in the same log. When an error is found, we classify where it broke: document omission, search ranking, context placement, generation, or citation display. The maturity of RAG is evaluated not by the rate of plausible answers but by the speed at which errors can be localized and corrected alongside the evidence.

We separate the responsibilities for document owners, search operators, and answer quality managers for production changes. When major errors increase, we restore not only the immediate previous model but also the index version, permission rules, and candidate order to reproduce which gate deteriorated. If the cause cannot be identified, we set rolling back the relevant question group to withholding as the rollback condition, and do not lift the hold based solely on the recovery of the overall average.

• Search: Ensure the correct valid version enters candidates and unauthorized/expired documents are not mixed in
• Faithfulness: Map each claim in the answer to adopted evidence, counting unsupported and contradictory claims separately
• Withholding: Do not answer questions lacking necessary evidence, having conflicts, or unable to reproduce citations
• Recovery: Re-run failure samples by aligning questions, indices, candidate order, models, and correction history