---
title: "What is RAG? Why AI that cites makes fewer mistakes | ekory"
description: "RAG makes AI answer from your documents and cite them. Errors drop but don’t disappear: the evidence, the limits and what to ask of any system."
url: https://ekory.ai/blog/why-cited-answers-are-more-reliable
language: en
italian: https://ekory.ai/it/blog/why-cited-answers-are-more-reliable.md
updated: 2026-09-24
---

Guide 24 September 2026 11 min read

# Why AI that cites its sources makes fewer mistakes: RAG explained simply

An AI model that answers from memory tends to guess when it doesn’t know. With RAG it first searches your documents and answers from them, with the source alongside. Errors drop but don’t disappear: in a Stanford test on legal questions, from 43% for GPT-4 to between 17% and 33% for RAG-based tools. That is why the citation matters as much as the answer.

- **43%** of GPT-4’s answers to legal questions contained hallucinations (Stanford, 2025)
- **17%** for the two RAG-based professional tools that erred least: fewer errors, not zero
- **74.5%** of the citations of AI search engines actually supported the sentence they backed (2023)

## Key takeaways

- **RAG** (retrieval-augmented generation) makes an AI answer in two steps: first it searches a collection of documents for the relevant passages, then it writes the answer from those passages and cites them. The technique was described in 2020 by researchers at Facebook AI Research, University College London and New York University.
- **It reduces errors; it does not eliminate them.** In Stanford’s test of more than 200 questions on US law, GPT-4 without a legal database hallucinated in 43% of its answers; three professional RAG-based tools did so in 17% to 33%.
- **The citation is there so you can check.** In a 2023 audit of four AI search engines, only 74.5% of citations actually supported the sentence they were attached to. A source that opens in one click, at the right page, makes the remaining error visible.
- **RAG still fails** when the search misses the right passage, when the documents are old or wrong (in a Stanford test, models preferred a wrong document to the correct answer they knew in more than 60% of cases) and when permissions are not applied before the search.
- **Ask any system for five things**: a citation to the document and the page, an “I can’t find it” when the answer isn’t there, permissions applied before the search, sources that update themselves, and a willingness to be tested on your own questions.

## What an AI hallucination is, and why it happens

On 20 February 2026 the Court of Syracuse, in Sicily, ordered a company to pay the other side €14,103 on top of legal costs, plus €2,000 to the fund for court fines. Its brief cited four rulings of Italy’s Supreme Court with passages in quotation marks. The judge looked them up in the legal databases: none of them contained those passages, and some dealt with something else entirely. The explanation the court found the only plausible one was a generative AI tool used without checking its output. The court held that the case had been brought with gross negligence, partly for this reason.

The ruling also offers a good definition of a hallucination: generated content that is plausible in form but false or non-existent in substance. The key word is _plausible_. A hallucination can’t be spotted by its look, because it has the same confidence as a correct answer.

Why does it happen? A language model does not consult an archive: it produces the most likely text based on what it saw during training. A team of researchers from OpenAI and Georgia Tech explained it in September 2025 with a classroom image: like a student facing a hard exam question, a model sometimes guesses instead of admitting it doesn’t know, because the way models are trained and scored rewards a guess more than an honest “I’m not sure”.

How often? That depends heavily on how it is measured. The Stanford AI Index 2026 reports a recent test, AA-Omniscience by Artificial Analysis (6,000 questions on 42 topics), that measures exactly this behavior: among the questions a model does not answer correctly, how often it gives a wrong answer instead of an “I don’t know”. Across 26 leading models, **the rate ranges from 22% to 94%**.

## What RAG is, in plain words

**RAG (retrieval-augmented generation) is a way of making an AI model answer in two steps: first it searches a collection of documents you choose for the passages relevant to the question; then it writes the answer from those passages and can say which ones it used.** The name comes from a 2020 paper by researchers at Facebook AI Research, University College London and New York University, presented at the NeurIPS conference.

The Stanford researchers who studied legal research tools use an image that captures the idea well: without RAG the model takes a closed-book exam, with RAG an open-book one. Instead of relying on a vague memory of what it read in training, it works on the text in front of it.

The 2020 paper already named two open problems of models that answer from memory: saying where an answer comes from, and updating what the model knows. The authors showed the second with a simple experiment. They took 82 world leaders who changed between 2016 and 2018 and asked the same question (“Who is the President of Peru?”) to the same model with two different archives. With the archive from the right year, 70% and 68% of answers were correct; with the wrong one, 12% and 4%. The model was the same. The documents had changed.

That is also the difference with the other approach you hear about, fine-tuning (additional training). Fine-tuning changes the model, and updating it means training again. RAG changes the documents the model reads, and updating it means updating the documents.

## How much less it errs: what the measurements say

The right question is not whether RAG works, but how much and where. Here are the most solid measurements we found, read at the source, with their limits.

Errors with and without searching the documents

| Study | What it compares | Result |
| --- | --- | --- |
| Lewis and others, 2020 | Human raters compare 452 pairs of generated texts, with and without search | The text with search was more factual in 42.7% of cases, the one without in 7.1% |
| Shuster and others, 2021 | Conversation replies with at least one false statement, rated by people | 68.2% without search; from 7.9% to 21.5% for the versions with search |
| Magesh and others, 2025 (Stanford) | Answers with hallucinations, over 200 questions on US law | 43% for GPT-4; from 17% to 33% for three professional RAG tools |
| Liu, McCoy and Wright, 2025 | Meta-analysis of 20 studies in medicine and biology | Better performance with RAG: odds ratio 1.35 (between 1.19 and 1.53) |
| Vectara, leaderboard of 22 September 2026 | Summaries of a given text judged unfaithful, 108 models | From 1.8% to 24.2%; half the models above roughly 9.5% |

Sources: the works listed in the references. The studies measure different tasks, models and years: compare numbers within a row, not across rows.

The most instructive case is the Stanford test (RegLab and the Institute for Human-Centered AI), published in 2025 in the Journal of Empirical Legal Studies. Some legal research providers presented RAG as the end of hallucinations; the researchers prepared and pre-registered more than 200 questions and put them to three professional RAG-based tools from LexisNexis and Thomson Reuters, and to GPT-4. The tools erred less than GPT-4 (43%), but one answer in six, or one in three depending on the tool, still contained a hallucination: a false statement, or a source cited for something it does not say. The authors’ conclusion is blunt: the providers had promised more than their tools deliver.

The second figure shows the underlying limit. Vectara’s leaderboard, updated on 22 September 2026, gives 108 models the same task: summarize a text using only what it says. Even with the document in front of them, no model reaches zero: the summaries judged unfaithful range from 1.8% to 24.2%. The maintainers add two caveats themselves: the judge is their own automatic evaluation model, not a person, and it measures faithfulness to the text, not the quality of the summary.

In medicine, a meta-analysis of 20 studies published in 2025 in the Journal of the American Medical Informatics Association finds a real but moderate improvement: an odds ratio of 1.35, meaning the odds of a better result rise by about 35% with RAG. The authors also flag a possible publication bias, the tendency to publish positive results first.

The honest summary: working from documents cuts errors substantially, but does not eliminate them. What changes is the nature of the error that remains. It is no longer an invention with nothing behind it; it is a statement next to a source you can open. That is where the citation comes in.

## Why the citation matters as much as the answer

An answer with a source is not correct by definition. It is checkable, which is different. A 2023 audit of four AI search engines widely used at the time shows it: the answers were fluent and looked informative, yet only 51.5% of the generated sentences were fully supported by their citations, and only 74.5% of the citations actually supported the sentence they were attached to. The authors judged these numbers worrying for tools that look reliable and that many people use as their first source of information.

There is also a decisive difference between a citation that was retrieved and one written from memory. When a model produces references on its own, without searching for them, it can make them up. A study published in 2024 in the Journal of Medical Internet Research asked three widely used models to find the literature of 11 systematic reviews: among the references they proposed, 28.6% were hallucinated for GPT-4, 39.6% for GPT-3.5 and 91.4% for Bard, meaning at least two of title, first author and year were wrong. In a RAG system the citation points instead to a document the system actually found, and that you can open.

It is the same point the Syracuse ruling makes: according to the court, using these tools without checking their output against primary sources is gross negligence. A citation does not remove the duty to check; it lowers its cost. That is why the detail matters: a citation to a 300-page manual sends you searching again, a citation to the right page can be checked in a few seconds.

## Where RAG still fails

If RAG reduces errors without eliminating them, it pays to know where the remaining ones come from. A study presented at the CAIN 2024 conference, drawn from three real projects in research, education and biomedicine, lists seven failure points, mostly technical. For a company using these systems, four weak spots matter most.

1. **The search misses the right passage.** The answer exists, but in a document the search does not rank near the top, or in words different from the question’s. In the Stanford test, poor retrieval contributed to almost half (47%) of one tool’s hallucinations. Stuffing the model with text does not fix it: with very long inputs, models make worse use of information that sits in the middle.
2. **The document is old or wrong.** RAG repeats what it finds. In a Stanford study presented at NeurIPS 2024, six leading models, given a document containing a wrong figure, adopted it over the correct answer they already knew in more than 60% of cases. If the 2024 and the 2026 price lists sit in the same folder, the system can cite the wrong one with complete confidence.
3. **The answer isn’t there, and the system makes one up.** It is the first of the seven failure points: when the question is on topic but the documents don’t contain the answer, the system can be fooled into answering anyway. A good system says “I can’t find it in the documents”.
4. **Permissions come last.** If the index holds everything and access is checked only on the answer, a well-phrased question can surface what it shouldn’t. OWASP, the nonprofit foundation known for its lists of application security risks, includes this risk in its 2025 list for applications built on language models and recommends search indexes that take permissions into account.

The authors of the CAIN study add a remark useful to anyone buying: a RAG system’s robustness is built through use, not declared at the start. It is worth testing one on your own questions before trusting it.

## What to ask of any company AI system

The evidence leads to five questions worth asking before you choose an assistant that works on your company’s documents, whatever the product.

1. **Does it cite the document and the page?** A file name is not enough: the citation should take you to the passage in one click, so checking takes seconds.
2. **Does it say when it doesn’t know?** Ask a question the documents don’t answer. The right answer is “I can’t find it”, not a well-written guess.
3. **Are permissions applied before the search?** Someone who can’t open a document should not find its content in an answer, even by rephrasing the question.
4. **Do the sources stay up to date by themselves?** When a file changes where it lives, what the system reads should change too, and old versions should be able to leave.
5. **Will it let you test it?** Prepare 20 or 30 real questions whose answer and source you know, and count: right with the right source, right without a source, wrong, “not found”.

One concrete example of how this translates: in ekory every answer of an Eko, the AI agent that works on the company’s documents, cites the document and the page, and says so when the answer isn’t in the documents. An Eko reads only the documents you have given it, and permissions are checked before the search, not after. The details are on the [Product](https://ekory.ai/product) and [Security](https://ekory.ai/security) pages.

Be wary of anyone promising zero errors: no serious measurement supports it. RAG makes errors rarer and, above all, visible, which is the difference between trusting an answer and being able to check it. If you are working out how to organize your company’s documents for a system like this, our [guide to the company knowledge base](https://ekory.ai/blog/company-knowledge-base-with-ai) starts there.

## Frequently asked questions

### What is RAG, in simple terms?

It is a way of making an AI answer in two steps: first it searches a collection of documents for the relevant passages, then it writes the answer from them and says where they come from. It is like moving from a closed-book exam to an open-book one. The technique was described in 2020 by researchers at Facebook AI Research, UCL and NYU.

### Does RAG eliminate hallucinations?

No, it reduces them. In Stanford’s test of more than 200 legal questions, GPT-4 hallucinated in 43% of its answers and three professional RAG-based tools in 17% to 33%. Even with the document in front of them, models write unfaithful summaries 1.8% to 24.2% of the time, according to Vectara’s September 2026 leaderboard.

### What is the difference between RAG and fine-tuning?

Fine-tuning, or additional training, changes the model itself, so updating it means training again. RAG leaves the model as it is and changes the documents it reads, so updating it means updating the documents. For company knowledge, which changes often and has to be checkable, RAG is the more direct route.

### How do you reduce AI hallucinations in a company?

Make the AI answer from the company’s documents instead of from memory, require a citation to the page, keep old versions of documents out, and reward “I can’t find it” over an invented answer. Then measure: a test with 20 or 30 questions whose answers you know tells you more than any promise.

### Why not just paste all our documents into ChatGPT?

For three reasons. With very long inputs, models make worse use of the information in the middle; nobody controls who can read what; and when a document changes, the pasted copy stays old. A RAG system searches the up-to-date documents for the right passages each time, and can apply permissions before it searches.

### How do you check an AI answer?

Open the cited source and read the passage: if the citation goes to the right page, it takes a few seconds. An answer with no source, or with a source that doesn’t say what the answer claims, should be treated as a hypothesis. A 2023 audit found that only 74.5% of the citations of AI search engines actually supported their sentence.

## Sources

1. Lewis P. and others, [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401), NeurIPS, 2020.
2. Shuster K. and others, [Retrieval Augmentation Reduces Hallucination in Conversation](https://aclanthology.org/2021.findings-emnlp.320/), Findings of EMNLP, 2021.
3. Magesh V. and others, [Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools](https://doi.org/10.1111/jels.12413), Journal of Empirical Legal Studies, 2025.
4. Liu N. F., Zhang T., Liang P., [Evaluating Verifiability in Generative Search Engines](https://arxiv.org/abs/2304.09848), Findings of EMNLP, 2023.
5. Wu K., Wu E., Zou J., [ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence](https://arxiv.org/abs/2404.10198), NeurIPS Datasets and Benchmarks, 2024.
6. Barnett S. and others, [Seven Failure Points When Engineering a Retrieval Augmented Generation System](https://arxiv.org/abs/2401.05856), CAIN, 2024.
7. Liu N. F. and others, [Lost in the Middle: How Language Models Use Long Contexts](https://aclanthology.org/2024.tacl-1.9/), Transactions of the ACL, 2024.
8. Liu S., McCoy A. B., Wright A., [Improving large language model applications in biomedicine with retrieval-augmented generation](https://academic.oup.com/jamia/article/32/4/605/7954485), JAMIA, 2025.
9. Chelli M. and others, [Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews](https://www.jmir.org/2024/1/e53164), Journal of Medical Internet Research, 2024.
10. Vectara, [Hallucination Leaderboard](https://github.com/vectara/hallucination-leaderboard), version of 22 September 2026.
11. Kalai A. T. and others, [Why Language Models Hallucinate](https://arxiv.org/abs/2509.04664), OpenAI and Georgia Tech, 2025.
12. Stanford HAI, [AI Index Report 2026, Responsible AI](https://hai.stanford.edu/ai-index/2026-ai-index-report/responsible-ai), 2026; Artificial Analysis, [AA-Omniscience](https://artificialanalysis.ai/articles/aa-omniscience-knowledge-hallucination-benchmark), 2025.
13. Tribunale di Siracusa, second civil section, judgment no. 338 of 20 February 2026 ([text](https://iusletter.com/wp-content/uploads/tribunale-siracusa-sez-ii-sent-20-02-2026-n-338.pdf)).
14. OWASP, [LLM08:2025 Vector and Embedding Weaknesses](https://genai.owasp.org/llmrisk/llm082025-vector-and-embedding-weaknesses/), OWASP Top 10 for LLM Applications, 2025.

## More research and articles.

- [The cost of starting from scratch](https://ekory.ai/blog/the-cost-of-starting-from-scratch): Research 17 September 2026 Onboarding a new hire costs €5,300 to €9,300: four months to full productivity and a hundred hours of colleagues’ time. In Italy, €16 to €28 billion a year.

- [AI in law firms: professional secrecy, Law 132/2025 and the uses that hold up](https://ekory.ai/blog/ai-in-law-firms): Industries 15 September 2026 Secrecy applies to AI providers too, the client must be told, and AI-invented citations have already reached the courts, in Italy as well.

- [All research and articles.](https://ekory.ai/blog): Resources Data, regulation and cases, in English and Italian.

## Tell us about a process you still do by hand. We'll show it to you on your own materials.

For the first call, your website or a catalog is all we need.