Retrieval-augmented generation, usually called RAG, combines a generative model with a process that retrieves material relevant to a request. Lewis and colleagues' foundational work joins learned model information with an external retrieval mechanism. Read the paper.
A simple way to follow the process is to separate finding material from composing an answer. A service may search a collection, choose passages, and supply those passages to a model. The response can then quote, summarize or draw inferences from them. Those operations deserve different labels.
The policy that really exists
Imagine asking whether a club member was allowed to take equipment home last Tuesday. The archive contains both the rule in force that day and a replacement issued on Wednesday. A search retrieves the replacement because it is prominently labeled current. The generated answer accurately summarizes it.
Nothing in this invented example requires a fabricated document. The failure is the relationship between the question and the evidence selected. The relevant question was historical; the retrieved source described the present.
To check the answer, the reader needs the policy's effective date, the cited passage and enough surrounding text to see any exception. A link to the club's home page would offer less help. So would a precise citation to a document that does not answer the question being asked.
A source can be incomplete
An archive is also a selection. It may contain one person's letters and none of the replies. A generated account can sound balanced while drawing on that uneven collection. Showing sources helps a reader recognize the imbalance; it cannot manufacture the missing record.
That is why “uses your documents” should start a conversation about coverage, dates, permissions and citation quality. It should not end the conversation about accuracy. A system may be useful while still needing to say that the available material does not settle a claim.
Two fictional uses
The Evidence Was Perfect explores a genuine record applied to the wrong question. The Last Person Who Remembers turns to the emotional authority of a reconstructed family account. Both are non-canon stories. Neither describes a real RAG product or demonstrates its failure rate.
For another step, compare content provenance, which concerns a record's history, with AI memory, which separates different ways information can persist. A reader should be able to ask where an answer came from without having to become the system's engineer.
Listen with this question in mind
- Will We Be ‘Haunted’ By AI Chatbots Of Dead Loved Ones?
Who should have a say before a familiar voice begins saying something new?
- Digital Content Provenance
What has become easier to verify, and which claim still needs separate evidence?
Sources and limits
The examples and recommendations are editorial explanations. Fiction illustrates the questions; it is not evidence of real-world outcomes.
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Foundational architecture; not a description of every modern RAG product or a griefbot study.