Fine-tuning changes a language model by training it further on the brand's own examples. Retrieval-augmented generation (RAG) leaves the model as it is and gives it the right documents or records at the moment it answers. For an ecommerce brand, RAG is usually the right first choice, because the facts an AI tool needs, such as products, prices, stock and order status, change every day. Fine-tuning earns its place when the brand needs a consistent voice or format, or wants a smaller, cheaper model to do one task well. This answer sets out the difference, the costs at published prices and how to choose. It is from Brand Media Works, an AI consulting and marketing data engineering firm for ecommerce brands.
What is the difference between fine-tuning and RAG?
| RAG | Fine-tuning | |
|---|---|---|
| What it changes | What the model can see when it answers | The model itself |
| Best for | Facts that change: catalog, prices, stock, policies, orders | Style and behavior: brand voice, output format, one repeated task |
| Keeping it current | Update the documents or the database | Retrain when the examples change |
| Showing its sources | Can cite the document it used | Cannot point to where an answer came from |
| What it needs | Clean, searchable product and policy data | Thousands of good examples for most tasks |
A fine-tuned model does not reliably learn facts such as a price list. It learns patterns from its examples. If a price changes after training, the model still has the old pattern. That is why a support or shopping assistant that quotes products and policies should look them up, not remember them.
What does each cost, and when is a custom build worth it?
The compute is the small part of either approach. Together AI lists LoRA fine-tuning at $0.34 per million training tokens for small Qwen models and $1.05 for Gemma 4 26B or Qwen3.5 27B, with a $4 minimum per job (read 31 October 2026). Training on 5,000 examples of 500 tokens each for three passes is 7.5 million tokens, about $7.88 on Gemma 4 26B (our arithmetic). Running an open model is also cheap: Amazon Bedrock lists Gemma 4 26B A4B at $0.13 per million input tokens and $0.40 per million output tokens in US regions (read 15 October 2026).
The cost is in the work around the model. Published build prices from AI development firms give the range:
- One firm prices a production AI feature with RAG integration at $25,000 to $50,000 over 8 to 12 weeks (read 18 October 2026).
- Another puts the retrieval layer alone, with embedding pipelines, a vector database and content filtering, at $15,000 to $50,000 (read 18 October 2026).
- A third puts the setup cost of fine-tuning an open-weight model at $30,000 to $90,000, in return for tighter data control and lower token prices at high volume (read 18 October 2026).
Fine-tuning a closed model is also becoming less of an option. OpenAI says it is winding down its fine-tuning platform: new users can no longer access it, existing users can create training jobs for the coming months, and fine-tuned models stay available until their base models are retired (read 31 October 2026). For a brand that wants a fine-tuned model it controls, the route now runs through open-weight models, which can be trained and run inside the brand's own cloud account.
A custom build is worth it when the brand's AI use touches customer or order data at volume, when answers must cite the brand's own policies, or when one task, such as writing product attributes, runs thousands of times a day. Private LLMs on open-weight models describes how Brand Media Works builds both inside a brand's AWS account, and what a private LLM costs to run sets out the running costs.
When should an ecommerce brand use RAG?
- Product questions: sizes, materials, compatibility and stock, answered from the live catalog.
- Policy questions: returns, shipping and warranties, answered from the current policy with a link to it.
- Order questions: status and delivery dates, read from the order system for the signed-in customer only.
- Internal search: letting the team ask questions of past campaigns, reports and supplier documents.
RAG depends on the data being findable. Product data with missing attributes or policies spread across old pages gives poor answers, however good the model is. The first step is usually cleaning and structuring that data.
When is fine-tuning worth it?
- A consistent brand voice across thousands of product descriptions or replies, where prompting alone drifts.
- A fixed output format, such as product attributes in the exact structure a feed or catalog needs.
- Classification at volume, such as tagging tickets or reviews, where a small fine-tuned model can replace a larger, more expensive one.
The examples are the real cost. One AI firm says simple tasks such as classification and extraction need 1,000 to 10,000 examples, and complex generation tasks 10,000 to 100,000 (read 15 October 2026). Many brands already have them in past tickets, approved product copy or tagged reviews, but they need checking before training.
Can you use fine-tuning and RAG together?
Yes, and many production systems do. The fine-tuned model handles tone and format, and retrieval supplies the facts. A common order is to start with RAG and a well-written prompt on an open model, measure where it falls short, and fine-tune only for the gaps that prompting cannot close. Private AI for business covers when a marketing team needs its own model at all.
Frequently asked questions
Is RAG better than fine-tuning?
For facts that change, yes: RAG keeps answers current without retraining and can cite its source. For voice, format and narrow repeated tasks, fine-tuning is better. They solve different problems, so the question is which problem the brand has.
Can I still fine-tune OpenAI models?
Only if the account already used OpenAI's fine-tuning platform. OpenAI says the platform is being wound down and is no longer accessible to new users, though existing fine-tuned models stay available until their base models are retired (read 31 October 2026). Open-weight models can still be fine-tuned on services such as Together AI, or on the brand's own GPUs.
Does fine-tuning teach a model my product catalog?
Not reliably. Fine-tuning teaches patterns, not a lookup table, and a catalog changes faster than a model can be retrained. Use retrieval for catalog facts and fine-tuning for how the model writes about them.