tier II · practitioner · 9 min · interactive article ✦
Ask ChatGPT about the waiting period set in your company agreement: it doesn't know — and it may well invent a plausible answer. Your procedures, your contracts, your customer histories don't exist inside the model. RAG is the standard answer to this problem, and it's probably the first serious AI project your company will run. Best to understand it before it lands in a committee meeting.
Three letters, one simple principle
RAG: Retrieval-Augmented Generation. Instead of letting the model answer from memory, the system first searches your documents for the relevant passages, then provides them to the model along with your question, and the model writes its answer from those excerpts — with cited sources to back it.
the RAG pipeline: search first, generate second
Picture a brilliant consultant who, before every answer, is handed the right file, open to the right page. That's RAG.
Why it's the right first step
No retraining: you don't modify the model, you feed it. Cheaper and faster than fine-tuning.
Always up to date: a procedure changes? You update the document, not the model.
Traceable: the answer cites its sources — the user can verify. For enterprise use, that's decisive.
Partitionable: you choose what goes into the base — and you can respect existing access rights.
What a manager must watch
RAG's Achilles' heel isn't the AI — it's the quality of your documentation. A RAG connected to outdated, contradictory or poorly organized documents will produce outdated, contradictory and well-phrased answers. The golden rule: a RAG's value is 20% model and 80% document governance. Good news: governance is your job.
test yourself — just like in the path
Your IT department proposes a RAG assistant on the HR document base. What is THE first question to ask?
✓ The question that kills RAG projects is never the choice of model: it's the governance of sources and access. You just saved six months of committee time.
Important, but not first. Reread the 20/80 rule and try again.
Going deeper
RAG or fine-tuning: how to choose?
RAG brings knowledge (your documents, up to date). Fine-tuning brings a behavior (a tone, a format, a specialty). Need the AI to answer about your procedures? RAG. Need it to write exactly in the house style across thousands of cases? Fine-tuning — and often, RAG first anyway.
The 3 scoping questions before launching
1) Which documents go into the base — and which are excluded (sensitive HR, legal)? 2) Who is responsible for keeping the sources updated? 3) Does the system respect existing access rights — can an employee obtain, via the AI, a document they aren't allowed to open?
The warning sign to know
If the vendor's demo answers perfectly… on THEIR demo data, insist on a pilot using YOUR real documents, with your duplicates, your contradictory versions and your scanned PDFs. That's where RAG projects are won or lost.
How do you know if your AI is improving?
A RAG (or any assistant) isn't frozen: it's steered over time. The simplest lever is right in front of you — the thumbs up / thumbs down your users click. It's not cosmetic: it's improvement data. Three building blocks to put in place: observability (trace what the system answered, and from which sources), feedback collection (the thumbs vote, plus a comment), and a failure-review routine (what do the negative ratings say? missing source, off-topic excerpt, made-up answer?). The useful loop: observe → measure → fix. Without it, quality drifts silently and you don't see it.
Frequently asked questions
What's the difference between RAG and fine-tuning?
RAG provides your documents to the model at question time, without retraining it. Fine-tuning retrains the model on examples to fix a tone or a format. RAG is often the right first choice.
Does RAG retrain the AI on my data?
No. It retrieves the relevant excerpts on the fly and passes them to the model; your data doesn't modify the model itself.
What do you need before launching a RAG project?
Clean, up-to-date documents with clear access rights. Documentation quality is the bulk — about 80% — of the project's success.
Does RAG prevent hallucinations?
It strongly reduces them, since answers are sourced and verifiable, but doesn't fully eliminate them: verifying what matters remains necessary.
This notion is a full level of the IAPLC path.
In the course, you practice it on YOUR context: lessons generated for your job, interactive exercises, gamified progression.