RAG vs fine-tuning
Fine-tuning teaches a model a style or behaviour by changing its weights. Retrieval-augmented generation puts your actual documents in front of it at question time. For keeping business facts current and citable, retrieval is the right tool — and updating means re-crawling, not re-training.
Side by side
| RubyRep | Fine-tuning | |
|---|---|---|
| Updating a price change | Re-crawl the page | Re-train the model |
| Citing a source | Yes — the retrieved page | No |
| Auditing a wrong answer | Inspect what was retrieved | Opaque |
| Teaching tone or format | Prompt only | Strong fit |
| Cost to change | A crawl | A training run |
| Handling unknowns | Detectable and reportable | Model guesses confidently |
Where Fine-tuning is stronger
- Teaches a consistent voice, format or output structure
- Can encode a specialised task the base model handles poorly
- No retrieval step at inference time
Where RubyRep is stronger
- Facts update by re-crawling, not re-training
- Every answer can cite the passage it came from
- A weak retrieval score is a usable signal to say 'I don't know'
Which one you actually want
Choose Fine-tuning if
- You need a very specific output format or house voice the prompt cannot achieve
- The task itself is unusual and the base model performs poorly at it
Choose RubyRep if
- The knowledge is facts about your business that change over time
- You need to show a customer where an answer came from
- You need the system to admit when it does not know
The honest answer
For customer support on a business website, retrieval is almost always correct. Your prices, policies and product details change; model weights should not have to. Fine-tuning is a tool for shaping behaviour, not for storing facts — and a fine-tuned model cannot tell you which page an answer came from, which is precisely what makes a support answer trustworthy.
Questions
- Can you use both?
- Yes — fine-tune for voice, retrieve for facts. In practice, prompting handles voice well enough that most teams never need the fine-tune.
- Does RubyRep fine-tune on my content?
- No. Your pages are crawled, split into passages and indexed for retrieval. No model is trained on your content, which is also why an update to your site shows up as soon as the re-crawl finishes.
Related
term
What is RAG (retrieval-augmented generation)?
RAG is what separates a chatbot that knows your business from one that guesses. Documents are split into passages and indexed by meaning; each question retrieves the closest passages, and the model is instructed to answer only from them.
term
What is an embedding?
Embeddings are why an AI representative can answer "do you deliver to Berlin?" from a page that only says "we ship across the EU". Text is converted to vectors, and similar meaning lands close together.
term
What is a knowledge base?
Traditionally a knowledge base is a help centre humans read. For AI support it is the source material a representative answers from — which means an ordinary website, well written, can serve as one without a separate help centre.
guide
How to train AI on your website content
"Training on your website" almost never means training a model. It means crawling your pages, splitting them into passages, indexing them by meaning, and retrieving the right ones at question time. Understanding that distinction tells you exactly what to fix when answers are wrong.
Judge it on your own website.
Comparison tables only get you so far. Point RubyRep at your site and read the answers it gives about your business.