Retrieval-augmented generation (RAG)
Retrieval-augmented generation (RAG) is a technique that gives an AI model your own documents to read before it answers, so that its response is grounded in current, specific information from those documents instead of only the general knowledge it learned during training.
Submit a project Read the RAG context engineering guideReviewed by Huy Nguyen (David), CEO · Updated 5 Nov 2026 · 4 min read
This term is one of the entries in the OutsourcingVN glossary of words buyers meet while scoping an AI feature.
Why does retrieval-augmented generation matter?
A general-purpose model can write fluently about a subject it has never seen your documents on, and it will not tell you when it is guessing. Retrieval-augmented generation addresses that by searching an approved knowledge source at the moment of the question and passing the matching passages to the model as part of its instructions, so the model answers from text it can point back to rather than from pattern memory alone. OutsourcingVN is operated by Netbase JSC, and buyers who ask for an AI assistant that must stay inside company facts are usually asking for this pattern, whether or not they know its name.
What is an example of retrieval-augmented generation in practice?
A support team wants an assistant that answers from the current product manual rather than from whatever the model picked up in training. The manual is split into sections, each section is indexed, and a question triggers a search for the closest matching sections before the model drafts a reply citing them. When the manual changes, only the index updates — the model itself does not need retraining.
What are the common misreadings of retrieval-augmented generation?
RAG is not fine-tuning: fine-tuning changes the model's weights, while RAG keeps the model unchanged and swaps in fresh text at answer time, which is why a knowledge base can be updated in minutes rather than requiring a new training run. Retrieval also does not remove the need for access control: the model will read whatever the search returns, so a document a user should not see must be excluded by the retrieval system, not trusted to the model's judgement. And a RAG system is not automatically accurate — it is only as good as the source it retrieves from, which is why evaluation and human review stay part of the design. The RAG versus fine-tuning guide covers that first distinction in more depth.
What is retrieval-augmented generation most often confused with?
| Commonly confused with | What's actually true | Watch out for |
|---|---|---|
| Fine-tuning | Fine-tuning changes model weights; RAG keeps the model unchanged and swaps in fresh text at answer time | Expecting RAG to permanently update a model's trained knowledge |
| A substitute for access control | The model reads whatever the search returns, so exclusion must happen at retrieval | Trusting the model's judgement instead of filtering at the source |
| Automatically accurate | A RAG system is only as good as the source it retrieves from | Skipping evaluation and human review because retrieval is in place |
Plan the next step for your project
Common questions
No. They solve different problems and are sometimes used together: fine-tuning changes how a model writes, retrieval changes what facts it can see.
No. It grounds the answer in retrieved text, but the system still needs evaluation against real questions and a human review step for anything that matters.
Which decisions does retrieval-augmented generation affect?
Deciding whether a project needs retrieval, how a source is indexed, who may see which documents, and how the system is tested before launch is the subject of the RAG and context engineering guide, and whether the underlying data is even ready to index is checked first on the data platform readiness assessment. How the indexing and retrieval step itself works is covered on the embeddings and semantic search guide, and data and AI platform engineering is the service that builds a system like this end to end. When the term above describes work you want scoped, describe the sources, the audience and the questions the assistant must answer, and Netbase will respond through Netbase's own outsourcing-services platform; you can submit a project when you are ready.