rag-2020-parametric-nonparametric
IN premise — summaries/2026/08/24/wiki-Retrieval-augmented_generation-chunk-1.md
Created 2026-08-25T02:58:51+00:00
RAG, introduced in a 2020 paper, combines a parametric language model with a non-parametric external memory accessed via retrieval at inference time.
Summary
The 2020 RAG approach splits knowledge into two parts: the model's own trained weights handle general language ability, while an outside source of documents or data supplies specific facts that the system pulls in only at the moment it needs to answer. This matters architecturally because it lets you swap in or update what the system "knows" without retraining the model, cleanly separating the machinery of understanding language from the content being looked up.