Fine-tuning teaches behavior and style, not facts. For a knowledge base that updates daily, RAG is the architecture; a light fine-tune only adds voice on top.
Picture a brilliant new hire on their first day. Fine-tuning is the onboarding that shapes how they speak, format reports, and follow company policy. It changes their habits, not the facts in the filing cabinet. RAG is letting them open the cabinet and read the current file before answering each question. If the prices changed yesterday, you do not re-onboard the hire overnight; you just update the file they read. Trying to cram every fact into their head through training is slow, gets stale instantly, and they start half-remembering details and making things up. So you train the habits once and let them look up the live facts every time.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
This MCQ is a misconception detector. Three of the four options propose elaborate ways to push facts into model weights, and only one puts knowledge where it belongs. The belief under test is that fine-tuning is how you teach a model new information. That belief is wrong, and it is wrong in a way that costs teams weeks of effort and ships systems that confidently lie.
The core distinction is simple once stated. Fine-tuning, whether full, LoRA, or QLoRA, runs gradient updates that change the model's output distribution. It encodes BEHAVIOR: tone, format, refusal rules, reasoning patterns. Retrieval-augmented generation leaves the model frozen and changes the PROMPT at query time, supplying fresh text the model reads. One edits parameters; the other edits context. They are operations on different artifacts and they solve different problems.
The knowledge base in this question updates frequently. That single fact decides the architecture. Anything that stores facts in weights inherits a retraining cadence measured in days, while the data shifts in hours. The only structures that stay current are retrieval-based, which is why option B is correct and the three fine-tuning for knowledge options all fail.
It helps to see why the misconception is so sticky. Both techniques make the model better at your task, so from the outside they look interchangeable, and courses teach them in parallel chapters that invite a false either-or. The fix is to stop thinking of them as rival approaches and start thinking of them as edits to different parts of the system. Fine-tuning edits the parameters; RAG edits the prompt. Two operations, two artifacts, two categories of result. This deep dive explains what gradient updates actually learn, why facts in weights confabulate, why each distractor is a tell, how to decide on the spot, and what the real production stack looks like.
What gradient updates actually encode
Fine-tuning is gradient descent over labelled examples. The optimiser nudges weights so the model's output distribution matches your targets, averaged across the whole dataset. The thing it captures well is a PATTERN that recurs: a tone that appears in thousands of replies, a JSON schema followed in every example, a refusal that fires on a class of inputs.
This is why fine-tuning is sample-efficient for behavior. A style is reinforced by every example that exhibits it, so a few thousand samples lock it in. The gradient has a consistent direction to follow across the data.
A single fact is the opposite case. The price of one product, or one clause in one policy, appears in a handful of examples at most. The gradient signal is weak and idiosyncratic, so the model does not cleanly store it. Instead it interpolates toward semantically nearby facts it saw more often. That interpolation is exactly the mechanism behind confident wrong answers.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Concern | Fine-tuning for facts (wrong tool) | RAG (right tool) |
|---|---|---|
| What it actually teaches | Behavior, style, format, reasoning patterns | Nothing new in the model; supplies fresh text to read |
| Daily knowledge update | Full retrain cycle (data, train, eval, deploy) | Re-embed changed chunks in seconds |
| Fact reliability | Partial memorisation, confident confabulation | Grounded in retrieved source with citations |
| Side effects | Catastrophic forgetting on narrow corpora | Retrieval latency and token cost per query |
| Cost to stay fresh | Repeated expensive retraining, still stale between runs | Cheap reindex, always current |
Real products, models, and research that use this idea.
- Perplexity grounds answers in live retrieval over web results rather than fine-tuning facts into a model, pairing retrieval with lightly tuned Claude Opus 4.7 and GPT-5.5 for citation-style synthesis.
- Glean and other enterprise search copilots index company wikis, tickets, and docs for RAG, then apply a thin fine-tune only for house voice and refusal policy.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy is a single fact sample-inefficient to learn via fine-tuning while a style is sample-efficient?
Frame gradient descent as pattern generalisation. A style appears across thousands of examples so the gradient reinforces it; a lone fact is one point the model interpolates toward neighbours, which is exactly how confabulation arises.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Believing fine-tuning is how you teach a model new facts. Gradient updates teach behavior and format; facts crammed into weights go stale and turn into confident confabulation.
60 second bullets to scan on the way to the call.
Why fine-tuning teaches behavior rather than facts
Why facts in weights go stale and cause confabulation
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.