Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation puts a retrieval step in front of a language model: the question is used to fetch relevant documents, and those documents are supplied to the model as context for its answer. It exists because a model's parameters are a snapshot of the data it was trained on, and that snapshot is stale the moment training ends. Retrieval separates what the model knows how to do from what it currently knows.

The design shifts the quality problem onto the corpus. A retrieval layer over stale, duplicated, or unattributed material produces confident answers built on it, so the properties that decide the outcome are data freshness, deduplication, and provenance good enough to cite. This is why real-time, structured, attributable data has become an input to AI systems rather than only to trading desks - it is the difference between an answer and a sourced answer.