Technology
What is RAG?
A language model on its own answers from memory: knowledge frozen when it was trained, with no sources. Retrieval-augmented generation, RAG, lets it look things up first and answer from what it found.
In short. RAG has two parts: a retriever that searches a collection of documents, the web or your own files for passages that fit the question, and a language model that writes the answer with those passages in front of it. Answers can be more current, and the sources can be shown. In the 2020 paper that named it, the method produced more specific and factual text than a model answering from memory. It still goes wrong when the search finds the wrong thing, and what it reads can carry hidden instructions.
Memory is not enough
A large language model stores a great deal of knowledge in its parameters: what a large language model is. But, as Lewis and colleagues noted in 2020, its ability to access and precisely handle that knowledge is limited, and showing where an answer came from, or updating what the model knows, were open research problems. Retraining a model every time the world changes is slow and costly.
Look it up, then answer
- Search. The question is turned into a search. In the original paper, a neural retriever looked through an index of Wikipedia; today it may be a web search, a company’s documents or your own mail.
- Fetch. The passages that fit best are pulled out.
- Answer. The language model writes its answer with those passages in front of it, and can point to them.
The authors report that models built this way set new top results on three open-domain question-answering tests at the time, and wrote more specific, diverse and factual text than a comparable model answering from memory alone.
What it fixes
- Fresher answers. Updating the collection it searches is enough; the model does not have to be retrained.
- Sources you can check. An answer can show the passages it came from.
- Your own information. With your permission, an assistant can answer about your calendar or your mail by retrieving from them.
What it does not fix
- A bad search. If the retriever finds a wrong or outdated page, the answer is built on it, just as fluently.
- Misreading. The model can still misstate what a source says, so the sources are worth a look when it matters.
- Hidden instructions. A page or an email it reads can contain text written to steer the AI: prompt injection.
- Privacy. To search your files, something has to read and index them. Ask where that index lives, and who can see it: AI assistant memory.
How Lychnos Ring uses it
When you repeat a claim and ask whether it holds, Lychnos checks it against live sources on the web and the news, gives you the reason on the ring’s screen and keeps the sources in the app. It checks the claim, not the person. It is in development: is that true?
Source, read 4 November 2026: Lewis and colleagues, retrieval-augmented generation for knowledge-intensive NLP tasks, 2020 (arXiv).
Questions
What does RAG mean in AI?
Retrieval-augmented generation: before a language model answers, a search step finds relevant passages in documents, the web or your own files, and the model writes its answer from them, so it can be more current and cite its sources.
Does RAG stop AI from making things up?
It reduces the problem but does not end it. If the search finds the wrong or an outdated page, the answer can still be wrong, and the model can still misread what it found. Checking the cited sources remains worthwhile.