Retrieval-augmented generation
The model reads your files afresh to answer each question. It works, but nothing accumulates. Each answer starts from a blank page.
Retrieval-augmented generation, RAG, is the technique behind most tools that let you chat with your documents. When you ask a question the system retrieves the passages most likely to be relevant, hands them to the model alongside the question, and the model answers from what it was handed. It grounds the answer in your material rather than the model’s general knowledge, which reduces hallucination and is why enterprise AI mostly works this way.
Its limit, for account planning, is that nothing accumulates. Each answer starts from a blank page; the synthesis you got yesterday is gone today unless you saved it, and the system has no view of the account, only of the passages that matched the last question. The LLM wiki pattern exists to fix exactly that, and a living plan uses retrieval as a component rather than as the whole design.
Set out in full in: Lewis et al. (2020); Creating Customer Value Propositions with AI (Kogan Page, 2027), chapter 9.
Related terms
The LLM wiki
Karpathy's pattern: the AI keeps a set of interlinked pages current as sources arrive, so knowledge is compiled once and maintained, not rebuilt on every question.
Context window
The model's working memory: everything it can see at once, including your prompt, the documents you attach and the conversation so far.
The living account plan
Something closer to a brain than a binder: an account plan an AI maintains from raw sources, against your methodology, so it is current when you open it.
See the method run on one real account in three days: the One Page Proposition.