This article is Part 1 of my RAG in Practice series, where I explain retrieval-augmented generation in practical, production-oriented terms.
In this part, we cover why AI models get things wrong and why they can't use your private data — the core problems RAG was designed to solve.
TechNova is a fictional company used as a running example throughout this series.
The Confident Wrong Answer
A customer contacts TechNova support. They want to return their WH-1000 headphones, bought last month, barely used. The AI assistant replies immediately. Friendly. Confident. Thirty days, no problem.
The policy changed to fifteen days last quarter. The return window closed two weeks ago. The customer escalates. A support agent has to intervene, apologize, and explain that the AI was wrong.
Nobody on your team wrote the wrong answer. The model was not confused. It gave the answer it had learned, from a document that was accurate at the time of training and wrong by the time it mattered.
The most dangerous AI answer is not nonsense. It is the smooth, confident answer that sounds right but was never connected to your system.
Why Models Get This Wrong
There are two causes, and they are separate.
The first is frozen knowledge. A model is trained on data up to a point in time. After that cutoff, it knows nothing new. Every fact the model holds is a snapshot, accurate when captured, increasingly stale after.
The WH-1000 return policy was thirty days when the model was trained on TechNova's documents. The model learned that fact correctly. The fact changed. The model did not.
The second is that the model cannot see your live systems. Even setting aside the training cutoff, the model has no connection to your actual systems at query time. It cannot open your policy database. It cannot query your CMS. It cannot retrieve the document that was updated last quarter. It answers from what it learned during training, a fixed internal state with no path to the live source of truth.
A model is not a connected system. It is a compressed representation of knowledge from a particular point in time.
It is worth being precise about what this means, because the language shapes the fix. The TechNova model did not make something up. It stated a real policy accurately. The problem is not that it generated fiction; it is that it was too faithful to a document that had stopped being true. Calling this a hallucination can lead people to fix the wrong thing. They may try to make the model sound less certain or lower its confidence.
A model that says "I'm not sure, but I think the return window is around thirty days" is still wrong. It is just more politely wrong. The customer still gets denied.
Fine-Tuning Does Not Fix This
The obvious fix is retraining. Update the model on TechNova's current documentation, including the new return policy, the latest specs, and the updated warranty terms.
Fine-tuning changes how a model behaves, including its tone, format, and reasoning patterns within a domain. A fine-tuned model is still a frozen model. Its knowledge is fixed at the point the fine-tuning data was collected. When TechNova's return policy changes next quarter, the fine-tuned model will have the same problem the base model had this quarter. You would have to retrain again. And again. The problem of outdated knowledge does not go away. It just gets pushed into a retraining schedule.
Fine-tuning addresses behavior. It does not keep knowledge current.
What Would Fix This
The model does not need to memorize every version of TechNova's return policy. It needs access to the current policy when the question is asked.
In the headphone example, the assistant would first look up TechNova's current return policy. It would see that the return window is now fifteen days, not thirty, and answer using that current information instead of relying on what the model learned during training.
That pattern (retrieve current knowledge first, then generate an answer based on it) is called Retrieval-Augmented Generation, or RAG.
Three Takeaways
AI models are trained on snapshots. On their own, they cannot see your live data.
The TechNova model learned the return policy correctly; it just never learned that it changed.The problem is not model intelligence; it is disconnection from your current systems.
The model did not reason poorly. It stated a fact it learned correctly. Precision without access is what makes confident wrong answers possible.Fine-tuning changes how a model behaves. It can't keep what it knows current.
Retraining on current documents is a scheduled snapshot, not a live connection. The same staleness problem reappears as soon as your data changes again.
Next: what changes when retrieval enters the loop, and why does retrieval quality determine the answer? That is Part 2: What RAG Is and Why It Works.
Part of the RAG in Practice series, originally published on AI in Practice Hub alongside the full series and companion code.
Top comments (3)
Mate, I read all 7 parts and I must say that is great content! You should make a book or a Udemy course. Bind those with real world hands-on projects and its top notch.
Some parts missing would be parent-child hierarchy chunking. i.e. for educational textbooks.
Quick update, John. I moved parent-child chunking out of Part 8 during the rewrite. It now lives in Part 4 and on the Patterns page, which fits the textbook use case better. I also added more hands-on material. Part 5 builds a small RAG system, and there are now two labs: RAG Debugging Lab and RAG Evaluation Lab. Thanks for the suggestion.
Thank you, John. I’m covering parent-child chunking in Part 8 on RAG in production, publishing this week. Educational textbooks are exactly the use case.