RAG (Retrieval-Augmented Generation)
Retrieval-Augmented Generation (RAG) is a method that connects a language model to an external knowledge base: before answering, the model retrieves the relevant passages from your documents, then writes its response grounded in that retrieved content. The goal is to base answers on verifiable sources rather than on the model's memory alone.
Updated on July 10, 2026 · Bertrand Dumast
How a RAG system works
A RAG system combines two distinct steps. First, retrieval: the user's question is turned into a query and matched against indexed documents (contracts, product sheets, internal procedures, support tickets) to pull out the most relevant passages. Second, generation: those passages are passed to the LLM (large language model), which writes the answer based on that content instead of relying solely on what it learned during training.
Why RAG cuts down on made-up answers
A language model on its own answers from what it learned during training, with no access to your proprietary data or your current business activity. It fills gaps with plausible-sounding but sometimes wrong statements. RAG changes that: the answer draws on real excerpts from your documents, so you can trace the source and check accuracy. This does not eliminate the risk of error entirely, but it turns the problem into one of document quality rather than pure invention.
When it's worth building
- Customer support: answer questions from your product documentation, terms of service, and ticket history, without rewriting an FAQ article every time something changes.
- Internal documentation: give your teams one entry point to procedures, process sheets, and technical notes scattered across several tools.
- Content that changes often: product catalogs, pricing, internal policies, where a static model quickly loses accuracy.
How to scope a RAG project
This kind of project usually sits inside a broader business automation and AI initiative: RAG alone will not help if the source data is unreliable or hard to reach.
- Inventory your source documents and how current they are: a RAG connected to outdated documents reproduces the errors it's meant to prevent.
- Define how documents are chunked and tagged with metadata, so retrieval finds the right passage, not just the right document.
- Plan for continuous updates to the document base, not a one-time load frozen in time.
- Scope which questions the system is meant to cover, so the model doesn't answer off-topic questions with the same false confidence.
Does RAG cost more than connecting directly to a language model?
RAG adds an indexing and retrieval step, so setup costs run higher than a plain connection to a language model. In exchange, it cuts the need to reprompt or retrain the model every time your content changes, which lowers ongoing costs over time.
Should I use RAG or a fine-tuned model?
Fine-tuning changes the model's own behavior and fits a recurring tone or format. RAG leaves the model unchanged and feeds it current content at question time, which suits cases where your documents change often, like a product catalog or a support ticket base.
What's the main risk in a RAG project?
The main risk isn't technical, it's documentary: if your sources are incomplete, contradictory, or poorly organized, RAG surfaces those flaws with the same confidence as a correct answer. Auditing the document base before rollout limits this risk.
Related terms.
LLM (large language model)
An LLM (large language model) is an artificial intelligence system trained on vast amounts of text to understand, summarize, and generate natural language.
Learn moreAI agent
An AI agent is a program built on a language model that carries out a task end to end: it reads data, decides on a course of action, and triggers steps through tools (APIs, business software, databases) without waiting for approval at every step.
Learn moreGEO (Generative Engine Optimization)
GEO (Generative Engine Optimization) is the set of practices that make web content citable and picked up by AI answer engines like ChatGPT or Perplexity, as well as by AI summaries built into search results.
Learn moreA project where RAG comes into play?
Send us your context. We tell you where to start, with a concrete first opinion and no commitment.