Retrieval-augmented generation (RAG)
Also: RAG, retrieval-grounded generation
What is Retrieval-augmented generation (RAG)?
RAG is a pattern where a system fetches relevant passages from your own documents or data first, then has a language model write an answer grounded in those passages rather than in its training memory.
What Retrieval-augmented generation (RAG) means
Retrieval-augmented generation splits question answering into two steps. First, a retriever searches your content (policies, manuals, tickets, contracts, database rows) and returns the handful of passages most relevant to the question. Second, a large language model receives the question plus those passages and writes an answer that cites them. The model's job is composition, not recall; the facts come from the retrieved text.
In practice a RAG pipeline is an ingestion side (parsing, chunking, embedding, indexing into a vector database or a hybrid index) and a query side (query rewriting, retrieval, reranking, prompt assembly, generation, citation). Most quality problems live in the ingestion and retrieval half, not in the model.
RAG is not fine-tuning. Fine-tuning changes how a model behaves; RAG changes what it can see at answer time. It is also not "upload a PDF to a chatbot": a production RAG system enforces access rights, keeps the index fresh as source documents change, and is measured with recall, precision and groundedness evals rather than a demo.
Who it really matters to
- CTO / Head of Engineering: RAG is the default architecture for putting private knowledge behind an LLM without training a model, and the place where most production failures originate.
- Product manager: It decides whether an assistant answers from your actual help centre and policies, with citations, or from a model's general impression of them.
- Compliance officer: Answers can be traced to a source document and a version, which is what an auditor or regulator will ask to see.
- CFO: Indexing and retrieval are cheap relative to training; the ongoing cost is inference tokens and keeping the corpus current.
Why it exists
Language models are trained on a snapshot of public text and know nothing about your contracts, price list or last week's policy change. Asked anyway, they produce fluent guesses. RAG exists to replace guessing with lookup: retrieve the evidence, then write from it. The trade-off is that the system is only as good as its retrieval. A poorly chunked corpus, a stale index or a query the retriever misreads yields a confident answer built on the wrong passage, which is harder to spot than an obviously wrong one. That is why retrieval quality has to be measured separately from answer quality.
Where it is applied
- Internal knowledge assistant over Google Drive, Confluence and Slack for a SaaS company's support and sales teams
- Policy and product-terms assistant for an NBFC's collections and customer-service staff, citing the exact clause
- Clinical protocol lookup for hospital staff over discharge summaries and treatment guidelines, kept inside the hospital's own infrastructure
- Course-grounded AI tutor that answers only from the institution's approved curriculum
- Returns and warranty assistant for a retailer that quotes the applicable policy for the customer's order and region
Is Retrieval-augmented generation (RAG) a skill?
Technique / practiceAn architecture pattern rather than a product. Building it well requires retrieval engineering, eval design and access-control work. Eazyware's Retrieval & Knowledge Engineering service covers ingestion, hybrid search, reranking and groundedness evals end to end.
Eazyware service that covers it: Retrieval & Knowledge Engineering. Starting prices are on the pricing page.
Frequently asked questions
Does RAG stop hallucinations completely?
No. It reduces them sharply by giving the model evidence to write from, but the model can still misread a passage or answer when retrieval returned nothing useful. Groundedness evals and a "no answer found" path are still needed.
How long does a RAG system take to build?
A grounded prototype over one document set fits inside a three-week ProofRun. A production system with permissions, freshness and evals typically takes six to ten weeks depending on how messy the source documents are.