GlossaryHow a model is chosen and adapted

Retrieval-augmented generation

Grounding a model's answer in documents fetched at request time, so the source stays inspectable and changing knowledge does not require retraining.

What it is actually for

Retrieval answers a question about data, not about the model: how fast does this knowledge change, and does the answer have to cite its source? If the knowledge changes weekly, retrieval is the only option that keeps up — training takes hours to days and the model is stale before it finishes. If the answer must be traceable to a specific passage in a specific document, retrieval is the only option that can produce one.

Retrieval can also run in several steps, searching again when the first pages are not enough; that is still retrieval: it brings evidence and grants no authority.

What it does not fix

Retrieval does not change how the model behaves. If the output format drifts, if the tone is wrong, or if the model ignores a tool contract, retrieval will not help; those are behavior problems and behavior is changed by post-training. Retrieval also does not make an ungrounded model honest — it makes grounding possible, and the system still has to check that the answer used the retrieved material.

Where it degrades

A retrieval index is fast on day one and quietly wrong later. Relevance decays when the embedding model is updated underneath existing vectors, when source documents change without re-indexing, or when the query distribution moves to a use case the index was never tuned for. Each of those returns plausible answers that are worse than they look, which is why source freshness and embedding-version consistency are operating metrics rather than setup steps.

Often confused with

How it differs from similar terms.

retrieval-augmented generation compared to adjacent terms
retrieval-augmented generation	Compared with	What's different
Retrieval	Fine-tuning	Retrieval grounds changing, citable knowledge. Fine-tuning changes behavior, format, and tool use.
Retrieval	Scoped memory	Retrieval reaches approved source material. Memory carries permitted state for this run or case only.

Follow-up questions.

Is RAG better than fine-tuning?: They answer different questions. Retrieval wins when the knowledge changes faster than a training cycle or when the answer must cite a specific source. Fine-tuning wins when the output format, tone, or tool-use contract has to hold across volume. Many institutional systems need both, and some need neither — a careful prompt with a structured output contract and a validator is often the whole design.
Why do RAG systems get worse over time?: Relevance decays for three separate reasons: the embedding model is updated underneath vectors that were embedded with the previous version; source documents change without re-indexing; or the query distribution moves to a use case the index was never tuned for. Each returns plausible answers that are quietly worse, so embedding-version consistency and source freshness are continuous operating metrics rather than setup steps.
Does retrieval stop a model inventing things?: It makes grounding possible, not automatic. Retrieval supplies source material the answer can be traced to; the system still has to check that the answer actually used it. That check is a guardrail outside the model, not a property the model acquires by being given documents.

Want these terms worked through on one of your own decisions? Start with a Decision Blueprint

Read further

Where it shows up in the work.

https://eqlx.ai/glossary/retrieval-augmented-generation · Printed copy; the website is the current version.