AI Answer Library

How accurate can an enterprise RAG knowledge base actually be?

Short answer

The honest answer: an accuracy number without an evaluation set behind it means nothing. "Accuracy" has to be split into at least four metrics — retrieval recall, answer correctness, hallucination rate and refusal behaviour — each measured differently. The same system that looks excellent at looking up a policy clause degrades sharply on cross-document synthesis or anything requiring calculation. When a vendor quotes a percentage without naming the evaluation set, the question types and the scoring method, ask for those before you discuss acceptance criteria.

Key points

  • 01"Accuracy" is a composite. Split it into at least four: retrieval recall (Recall@k), answer correctness, hallucination rate (or citation groundedness), and refusal versus over-refusal. Collapsing them into one number hides which stage is failing.
  • 02The ceiling is set by corpus quality, not by the model. Stale documents, conflicting versions, scanned-image PDFs and mixed corpora with no permission boundaries will defeat any model you swap in.
  • 03Question type dominates difficulty: single-document fact lookup is easiest, multi-document synthesis is harder, and anything needing calculation, temporal judgement or inference is hardest. Stratify the evaluation set by type instead of reporting one blended score.
  • 04The evaluation set belongs to the buyer, not the vendor. Have the business side write 100–300 real questions with reference answers and accept against those; a vendor-supplied benchmark score is not an acceptance criterion.
  • 05Refusing is usually more valuable than guessing. In clinical, legal and compliance settings, treat a healthy refusal rate as a feature rather than a defect — forcing the system to always answer just converts errors into hallucinations.

Break "accuracy" apart first

The four metrics below measure genuinely different things, and hitting any one of them alone does not make a system usable. The most common mistake in practice is looking only at final-answer correctness: retrieval never surfaced the right passage, the engineer keeps tuning the prompt, and the problem never gets fixed.

MetricThe question it answersHow to measure itCommon mistake
Retrieval recall (Recall@k)Did the passage containing the answer make it into the context at all?Manually label a gold passage per question and measure top-k hit rateJudging only the final answer while retrieval already lost the passage, then endlessly tuning prompts
Answer correctnessIs the generated answer factually right?Score against reference answers — correct / partially correct / wrong — by human or strong-model judgeAuto-scoring by text similarity, so "sounds like the reference" counts as "correct"
Hallucination rate / groundednessDoes the answer contain claims the cited sources do not support?Check sentence by sentence whether each claim is directly supported by the cited snippetTreating any citation as proof of grounding, without checking whether it supports the sentence
Refusal rate and over-refusal rateDoes it say "not in the knowledge base" when that is the truth?Seed the evaluation set with questions the corpus genuinely cannot answer and see whether it answers anywayCounting every refusal as a failure and forcing an answer, which raises the hallucination rate instead

Why answers go wrong, ordered by how often it happens

First, the corpus itself: three versions of the same policy with no effective dates, key content trapped in un-OCRed scans, internal jargon with no glossary. Second, a mismatch between chunking, retrieval and content: tables cut in half, long documents stripped of their section context, or pure vector search that cannot recall exact identifiers such as part numbers and clause references. Third, questions that fall outside the retrieval paradigm — synthesising across a dozen documents, performing arithmetic, or deciding which version is current — none of which are fixed by raising top-k. Fourth, permissions and recency: answering with content the asker should not see, or from a superseded revision. Model capability comes last; in most enterprise deployments it is not the binding constraint.

What to fix first, ordered by return on effort

One: fix the corpus — deduplicate, stamp every document with a version and effective date, OCR the scans, build a glossary. This usually yields the largest gain at the lowest cost. Two: fix retrieval — hybrid keyword plus vector search with a reranker, chunk along document structure rather than fixed character counts, and carry heading hierarchy into each chunk. Three: add constraints — require citations, forbid claims beyond the cited text, and refuse plus escalate to a human when retrieval confidence falls below a threshold. Four, and only then: move to a larger model or fine-tune; done before the corpus is clean, this has the worst return of the four. Throughout, keep the evaluation set running after every change, or you will not know whether you improved things or made them worse.

Where this applies

When this answer does not hold

  • This answer deliberately quotes no percentage. An accuracy figure detached from its evaluation set, question-type distribution and scoring rules is neither comparable nor testable, and publishing one would only mislead.
  • Public RAG benchmarks are built mostly on general-domain corpora, whose distribution differs sharply from internal policies, process documents or clinical records. Leaderboard scores do not transfer to your deployment.
  • High accuracy does not mean it is ready to face customers. Clinical, legal and safety-critical workflows still need a human review step — that is a process design decision, not a model metric.
  • Any corpus update invalidates prior evaluation results. Re-run the suite with each corpus version and keep the evaluation set under version control, or in six months nobody will be able to explain where the original number came from.

People also ask

  • What accuracy do RAG systems reach in practice?
  • Can an enterprise knowledge base answer questions without ever being wrong?
  • How do you deal with hallucination in a RAG system?
  • How do you evaluate the accuracy of a knowledge-base assistant?
  • A vendor claims 95% RAG accuracy — should I believe it?
Written by: YGG Solutions TeamPublished: 2026-08-01Last reviewed: 2026-08-01