Fine-tuning
Fine-tuning changes how a model behaves. It is a poor way to teach it what is true about your business.
Why it matters
'Train it on our data' is the most common request in AI support projects and usually the wrong instinct. Fine-tuning is expensive, slow, and stale the moment your content changes — and it cannot produce a citation.
- Assuming custom training is required to answer company questions
- Expecting a fine-tuned model to stay current as content changes
- Wanting sources on answers from a model with no retrieval
What it is good at
Shaping behaviour: tone, output format, adherence to a structure, handling a specialised task consistently. If you need answers in a rigid format every time, that is a fine-tuning problem.
Why retrieval wins for facts
Facts change. Retrieval updates when you publish; a fine-tune requires retraining. Retrieval also names its source, which a fine-tuned model fundamentally cannot — the information is distributed through weights, with no page to point at.
Currency
Publish and re-index versus retrain and redeploy.
Attribution
Retrieval can cite; weights cannot.
Common questions
- Should I fine-tune a model on my documentation?
- Almost certainly not. It is slower, more expensive, harder to keep current, and cannot cite sources. Retrieval solves the actual problem — making your content answerable — better on every one of those axes.
- Can you combine fine-tuning and retrieval?
- Yes, and where both are used the division is usually behaviour from fine-tuning and facts from retrieval. For most support use cases retrieval alone is sufficient.
From the blog
All posts- EvaluatingRAG vs fine-tuning for support: the decision that sticks“Train it on our docs” sounds right and usually isn’t. For website support, RAG wins on currency, citations, and cost; fine-tuning earns its keep for behaviour, not policies.Read
- BuildingPersona tuning without turning the bot into a mascotSet a specific greeting, match your brand’s register, and write hard boundaries. Skip the forced personality that makes refusals feel like a bit.Read
- MeasuringDeflection rate is the most overstated number in support automationCounting bot conversations as deflected tickets overstates the result. The honest version is a before-and-after on ticket volume, read next to satisfaction.Read
Answer honestly. Capture the rest.
Point Matter Chat at your site and see what it can — and can't — answer. It's honest about both.
No credit card. 2 minute setup.
Every answer cites the source it came from. When there isn't one, it says so — and hands the visitor to you.
Installs on the tools you already run.