What is fine-tuning? Retrain the AI or hand it the docs
Fine-tuning means taking an AI model that already exists and training it further on your own examples, to change how it responds. Most people who ask “what is fine-tuning?” want something else: an AI that knows their company. For that, handing it documents is nearly always enough.
What is fine-tuning? The pitch a campground gets
Three Junipers Campground, in the mountains of northern New Mexico, changes its rates every season. A vendor offers to “fine-tune a model on your data.” That means collecting hundreds of examples of a question and its ideal answer, then training the model again on them until it answers that way out of habit. What comes out is a copy of the model with different habits.
What fine-tuning in AI is good for
For locking in a way of doing things. Think of a very distinctive writing style, a format that has to come out the same every time, or a narrow task repeated thousands of times, like sorting reservations. It changes the how, not the what.
Fine-tuning vs. RAG: which one for what
This is the most common mix-up. With RAG, the AI looks in your documents before answering and can show where the answer came from. With fine-tuning, what it learned is dissolved inside the model: there’s no document to open. It’s not the same as embeddings either, which only arrange your texts so they can be found.
Why a small business rarely needs to fine-tune an LLM
Because what you want the AI to know (prices, terms, hours) changes. The July site rate goes up and the retrained model keeps quoting the old one until someone repeats the process. On top of that, a large language model (LLM) doesn’t store individual facts reliably: it can blend two seasons.
When to fine-tune a model
When there’s a high volume of a single task, more than enough good examples and someone technical to maintain it. Before you accept a proposal, ask what happens when a fact changes. The guide to training AI on your own data goes through the options.
What happens when a fact changes?
Two AIs that know the campground’s rates: one reads them from a document, the other learned them in a retraining run. Change the rate and compare.
Season change A tent site in July goes from $32 to $36 a night.
With documents
1 of 1 step
- Upload the new rate sheet under the same nameIt comes in as a new version and the previous one is kept
Campground rates.pdf version 2 · July: $36
How much is a tent site in July?
In July, a tent site is $36 a night. Now gives the new figure
Retraining
4 of 4 steps
- Prepare new examplesQuestions and answers with the right rate
- Train againThe vendor adjusts the model once more
- CheckThat it gives the new figure and hasn’t broken anything else
- DeploySwap the old model for the new one
The fact lives inside the model: there’s no document to open.
How much is a tent site in July?
In July, a tent site is $36 a night. Now gives the new figure
Both end up giving the new figure. The difference is what it takes to get there: one step versus four, every time a rate changes.
This is a diagram, and a simplification. Real retraining takes hours or days and its steps vary by vendor; here they take seconds so you can see the difference. The rates are made up.
Fine-tuning vs. RAG: which do I pick?
| Retrain (fine-tuning) | Hand it documents (RAG) | |
|---|---|---|
| What it’s for | Changing the style or format of the answers | Getting answers based on your company’s facts |
| What you need | Hundreds of well-made examples and someone technical | Your documents, kept current |
| When a fact changes | You have to train again | You change the document |
| Does it show the source? | No: what it learned stays inside the model | Yes, it can show the document it checked |
The two aren’t mutually exclusive: some teams retrain for style and hand over documents for facts.
How it looks in Flipa
Nothing gets retrained in Flipa. You give it documents, your website and what you tell it, and it looks there before answering. When a rate changes, you upload the document under the same name: it comes in as a new version and the previous one is kept. If an answer is off, you give it a thumbs-down and write how it should have answered. And the draft of our terms says your content isn’t used to train models. The app’s interface is in Spanish today. The details are in knowledge.
Before you sign a fine-tuning proposal
What is fine-tuning in AI?
It’s continuing the training of an existing model on your own examples, so that its internal parameters adjust to a task or a style. You don’t start from scratch: you retouch a general model.
Fine-tuning vs. RAG: what’s the difference?
RAG doesn’t touch the model: with each question it passes along the passages of your documents the model needs, which is why you update it by changing a document. Fine-tuning does modify the model, and it’s for changing its behavior, not for keeping facts current.
Do I need fine-tuning for the AI to know my company?
No. For it to know your products, prices and policies, what works is giving it those documents and having it search them. Fine-tuning makes sense later, if you need a style or a format you can’t get with instructions and examples.
When should you fine-tune a model?
When instructions and examples in the request aren’t enough, the task is narrow and repeated at high volume, and you have hundreds of good examples plus someone to maintain the result. If the goal is up-to-date facts, it’s the wrong tool.
How much does fine-tuning an LLM cost?
It depends on the provider, the model and how many examples you use, and not every provider offers it. Ask for a quote with the line items separated: preparing the examples, which is people’s work; the training; and using the model afterward, which is sometimes billed at a higher rate than the standard one.
Keep going
What is RAG
The usual alternative: the AI looks in your documents before answering.
Train AI on your own data
What “training” means in practice, and what your options are.
What is an LLM
The model that gets retrained, and why it isn’t a database.
Knowledge
How you give Flipa your documents and how you can tell how much it has understood.