Three ways to change an answer

Larkspur’s assistant answers from the company’s documents, and Evals checks every change before it ships. Now the support lead wants every reply to sound like Larkspur and sign off as the team. Someone in the meeting says: “Just fine-tune it.”

This chapter: What can you change to change an answer, and where does each change live?

You will learn

The three levers: the prompt, the documents and the weights
Which lever changes what the model reads, and which changes the model itself
Why a reply’s style and its facts can come from different levers

Words you’ll meet (4)

Prompt: The words sent to the model with each question, like “sign off as the Larkspur team”. Sent every time.
Documents: Passages fetched for each question and put in front of the model, as in the RAG module.
Weights: The numbers inside the model that decide what it writes. They change only when the model is trained.
Fine-tuning: Training a model that already exists a little further, on your own examples, so its weights change.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 4

Larkspur’s bench has three levers, one for each way to change what the assistant says. Pull each one and read where its change lives.