Crabhaus

What it is

Three ways to adapt a general LLM to your task. Prompting changes the request: instructions and examples go in the context window. RAG (retrieval-augmented generation) changes what the model sees: relevant documents are fetched at query time and placed in context. Fine-tuning changes the model itself: extra training on your examples adjusts its weights.

How it works

Prompting iterates in minutes and costs only tokens. RAG chunks and indexes your documents (usually as embeddings), retrieves the top matches per query, and puts them in the prompt - knowledge updates by re-indexing, not retraining. Fine-tuning runs supervised training on prompt-response pairs, often via lightweight LoRA adapters, so every later call gets the behavior without carrying examples.

A concrete example

A support bot: the system prompt sets tone and refusal rules; RAG retrieves current plan docs so answers quote today's prices, not training-data prices; and if the bot still won't emit your exact escalation JSON after prompt iteration, fine-tuning on a few hundred examples locks the format in.

Why it matters

The three fix different failures, and teams burn weeks on the wrong one. Wrong answers because the model lacks your facts? That is retrieval's job - facts change faster than any training loop. Wrong shape, tone, or format? That is what fine-tuning shapes. Escalate in cost order - prompt, then RAG when facts are missing, then fine-tune - and combine them freely.