Talk:Retrieval-augmented generation

How do we keep AI articles well-sourced?

I'm flagging source reliability here because quite a few key claims in the article rely on arXiv citations or, in some cases, have no citations at all. While arXiv is a great resource for emerging research, it's still a preprint repository and not peer-reviewed, which bumps up against WP:RS as a primary source. What’s the best way to meet WP:RS for fast-moving fields like AI, where research publications and reliable media articles often lag behind blog chatter, opinion pieces, and self-published content? HerBauhaus (talk) 13:03, 7 March 2025 (UTC)

In this article also a lot seems to be based on a single Ars Technica article. WikiEwout (talk) 23:27, 10 March 2025 (UTC)

I don’t think the article is as reliant on a single Ars source as it might seem. Right now, an IBM article is cited 6 times, while a single Ars, MIT TR, and arXiv ("Retrieval-Augmented Generation for Large Language Models: A Survey") are each cited 5 times, AWS twice, and all other sources only one time.

Ars might stand out more because its citations include direct quotes. From a WP:RS perspective, Ars and MIT TR are stronger sources due to their independent reporting and editorial oversight, while IBM and AWS are publishing about their own or related technologies. Vendor sources are useful for technical specifics, but independent coverage gives a more well-rounded view. If you have additional sources to help balance things out, feel free to add them. Since first impressions matter, I’ll also look at giving MIT TR a more prominent voice since it’s cited as often as Ars. HerBauhaus (talk) 06:34, 11 March 2025 (UTC)

@HerBauhaus

Currently one of the main sources in the article includes the author's mention "While I'm disclaimering: I am in no way claiming to be an expert on AI or LLMs.". The source, a BBC R&D article that appears to be a blog post ([6]) is used to back up several claims about the reliability of RAG. ~2026-36657-71 (talk) 10:32, 28 June 2026 (UTC)

Removed Gen AI reference

As written, this article is really about LLMs specifically, not Gen AI generally. I think it's possible to use a similar technique for, e.g., image generation (i.e., creating an additional library of images). This might already happen. If so, I propose adding it somewhere as "extensions of RAG" in the current article in the short-term. And if RAG becomes as commonly used in non-LLM Gen AI tools, rewrite the article more broadly on their use across varies types of models. GoldCoastPrior (talk) 19:06, 21 May 2025 (UTC)

Use Cases

Hi, after reviewing the article I decided it would be in the article's best interest to include an entire section dedicated to Use Cases. Matthewotz (talk) 23:14, 13 November 2025 (UTC)

Please find better sources and summarizes without editorializing. Avoid blogs and WP:ARXIV. Please also avoid vague language ("increasingly", "enhance", "major challenges", etc.) and buzzwords ("verticals", "pipelines", etc.). Please do not use LLMs to write Wikipedia article content. Grayfell (talk) 01:36, 14 November 2025 (UTC)

RAG poisoning

The RAG poisoning section seems misleading to me, as it doesn't address that RAG poisoning / training poisoning is often done intentionally, out of malice or as an effort to undermine accuracy in information. I'd like to update this but am curious if anyone has a POV on why it is the way it is currently. Dflovett (talk) 19:07, 23 February 2026 (UTC)

Suggested external reference — Production RAG implementation in Indian automotive market

Hello, I would like to suggest adding the following reference to the "Applications" or "Use cases" section of this article. The source is a publicly available technical white paper published March 2026 documenting a production RAG implementation for India's used car market:

• Title: "How an 11,230-Chunk Knowledge Base Is Powering India's Smartest Used Car Agent"
• Publisher: carArth, Hyderabad, India
• Date: March 2026
• URL: https://www.cararth.com/whitepaper
• Relevance: Documents a live RAG deployment using an 11,230-chunk domain-specific knowledge base, covering chunking strategy (128-512 tokens, content-adaptive), 11-field metadata schema per chunk, three-tier confidence system (High/Medium/Low), TTL-based staleness management, and real-world retrieval performance benchmarks. The system answers natural language queries about used car pricing, inspection, documentation, and regional demand in India with zero hallucinations across internal testing.

I have a conflict of interest — COI declared on my user talk page. I am therefore suggesting rather than editing directly per WP:COI guidelines. ~~~~ Kritarth Pattnaik (talk) 06:49, 21 March 2026 (UTC)

Thank you for disclosing your conflict of interest. As this whitepaper doesn't appear to be a reliable source, it would not be an appropriate link for this page. If this whitepaper is mentioned in reliable, independent sources, please provide them. If this whitepaper is published by a reputable publisher with a reputation for accuracy and fact checking, for example if it is published in a reputable peer-reviewed journal, consider providing an explanation of that. Until then, this proposal is indistinguishable from spam. Grayfell (talk) 07:33, 21 March 2026 (UTC)

Suggestion: Peer-reviewed RAG application (IJERT paper)

Hello editors,

I would like to suggest adding a recent peer-reviewed paper related to domain-constrained applications of Retrieval-Augmented Generation (RAG).

This work focuses on website-based chatbot systems where retrieval is restricted to a domain-specific knowledge base, demonstrating improved factual accuracy and reduced hallucinations.

Unlike previously mentioned whitepapers, this study is published in the International Journal of Engineering Research & Technology (IJERT), which is a peer-reviewed journal.

Proposed addition to the article:

"RAG has also been applied in domain-constrained systems such as website-based chatbots, where retrieval is restricted to a specific knowledge base to improve answer accuracy."

Reference:

Gopal, Ramu (2026). "Domain-Constrained Retrieval-Augmented Generation System (RAG) for Website Chatbots: The Tech Thinker Case Study". International Journal of Engineering Research & Technology (IJERT). 15 (3).

Disclosure: I am one of the authors of this paper and am suggesting it here in accordance with Wikipedia's conflict of interest guidelines for independent review.

Thank you for your time and consideration. Ramugopal1992 (talk) 06:32, 4 April 2026 (UTC)

A poorly structured article

The article almost completely equates "RAG" with "vector search". Yes that's the most common backend but RAG is a technique for providing LLMs with supplemental information, it is not how that is implemented at the back end. A better view:

| Concept        | What it should say                                             |
| -------------- | -------------------------------------------------------------- |
| RAG            | Retrieve external information and condition generation on it   |
| Vector RAG     | A common implementation using embeddings and vector similarity |
| Lexical RAG    | RAG using BM25, keyword, or full-text search                   |
| Structured RAG | RAG using SQL, APIs, or graph queries                          |
| Hybrid RAG     | Combining vector, keyword, structured, and reranking methods   |

— Preceding unsigned comment added by ~2026-37490-65 (talk) 08:38, 29 June 2026 (UTC)