What is Retrieval-augmented generation?
Answering from your documents rather than from the model's memory. The retrieval half is where it succeeds or fails.
The pattern is: find the relevant passages, then have a model write an answer using only those. It is on essentially every vendor site in this market, which is why leading with the acronym conveys nothing. What distinguishes implementations is entirely in the retrieval step and in what happens when retrieval returns the wrong passage confidently.
The other thing that decides it is traceability. On a delivered public-facing assistant, every answer was traceable to approved published content, and same-session task completion went from 27% to 68%. Traceability is the sentence that ends the hallucination objection, and it is a design property rather than a model property.
| On this | Retrieval-augmented generation | A model answering from its training |
|---|---|---|
| Where the answer comes from | Your documents, cited | Whatever the model absorbed, uncited |
| When your policy changes | Update the document; the answer changes | The old answer persists until the model is replaced |
| Failure mode | Retrieves the wrong passage - visible, and fixable | Invents a plausible answer - invisible until acted on |
- When it is the right answer
- Whenever the answer must be attributable to a document you control. Which, in a regulated operation, is most of the time.
- The service line that delivers this
- Workflow Copilots & Agents
They already have four systems open. Nobody wants a fifth place to go and ask.
Start here
If a term here is the one your board paper turns on, ask us about it.
We will send the entry, the evidence behind it, and the honest note about where it does not apply.
You get a reply within one working day, from the engineer who would do the work - not a sales sequence.