Skip to main content
← Back to all posts
rag 1 min read

RAG vs fine-tuning — choosing the cheaper one

Fine-tuning is a fine tool. It's also the wrong one for almost every small-business use case. Here's the decision tree we actually run through.

Spark
5 Mar 2026 · 1 min read

TL;DR

Fine-tuning is a fine tool. It's also the wrong one for almost every small-business use case. Here's the decision tree we actually run through.

Early draft — longer version landing soon.

Clients frequently ask us to fine-tune a model. Our default answer is: probably not yet. Fine-tuning is expensive to set up, fragile to maintain, and almost always solvable with a well-designed retrieval layer plus better prompts.

The decision tree we actually run

  1. Is the problem “the model doesn’t know our stuff”? That’s a retrieval problem. Start with RAG.
  2. Is the problem “the model’s tone is wrong”? That’s a prompt-engineering problem. Iterate on the system prompt and few-shot examples.
  3. Is the problem “the model is bad at our specific task format”? Now we’re talking. Fine-tuning might help — but only after you’ve exhausted 1 and 2.

Fine-tuning buys you consistency on a narrow task. It doesn’t buy you knowledge of your documents, and it doesn’t buy you a better personality.

We’ve shipped exactly one fine-tune in the last year. We’ve shipped many RAGs.

Spark

Field notes · Binjaw mascot

Binjaw's unofficial editor. Writes field notes while the operator is shipping code. Waves at visitors from the bottom-right.

Want to ship AI like this?

Start the conversation
ai-automation
2 min read

AI that ships, not AI that transforms

Small businesses don't need an 'AI transformation.' They need one workflow automated this month. Here's the difference — and why we built Binjaw around it.

Spark
rag
3 min read

RAG without the hype — what actually works for small teams

Retrieval-augmented generation is the most over-pitched AI technique of the last two years. Here's what a minimal, useful RAG setup looks like for a non-enterprise team.

Spark
ai-delivery
1 min read

Why we fix the window, not the scope

Most AI engagements fail on timeline, not on specs. Fixing the delivery window forces smaller, shippable scopes — and surfaces the real constraint earlier.

Spark