Teach it my docs, or let it read them?
RAG vs fine-tuning · Builder · 5 min · rag, embeddings, context · after Context window, Embeddings
RAG puts the right documents into the window at answer time (open book); fine-tuning changes the model by training on examples (study). Facts that change go in the book; habits, formats and style go into the model.
1 Guess first
2 Look
Two ways to teach it your stuff
1 / 3Open book (RAG)
- 1. Question comes in
- 2. Find the relevant pages
- 3. Put them in the window
- 4. Answer, citing them
Study (fine-tuning)
- 1. Collect examples
- 2. Train the model on them
- 3. Model changes
- 4. Answer from memory
Same question, two lanes. The model is the same at the start of both.
Two lanes, same model at the start: one adds pages to the window, the other changes the model itself.
RAG is embeddings doing a job: turn the question into a point, find the pages whose points sit nearest, put them in the window. The model then answers from text it can actually see — and can cite. Fine-tuning instead trains on hundreds of examples until a behaviour is inside the numbers.
4 You are the model
Route each of the four situations to a lane, then Check.
Open book (RAG)
- 1. Question comes in
- 2. Find the relevant pages
- 3. Put them in the window
- 4. Answer, citing them
Study (fine-tuning)
- 1. Collect examples
- 2. Train the model on them
- 3. Model changes
- 4. Answer from memory
Answer questions about our 400-page internal policy that changes monthly
Always reply in our brand’s terse, lowercase style
Cite the exact clause from a customer’s contract
Parse our odd internal ticket format reliably
Route each situation to a lane, then Check.5 Check
6 Say it back
7 Mission · 5 min
Take three short notes of your own, embed them with a free embedding key, embed a question, and find the nearest note by cosine similarity. You just built the retrieval half of RAG.
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Sources
Published Oct 20, 2026 · updated Oct 20, 2026