Fine Tuning


Buster is finally ready to start answering questions about your data! We aren’t finished onboarding quite yet. While documentation and context increase performance, we’ve found that there is “undocumented” context that Buster can learn from.

Here are a few examples of user questions that require some “undocumented context”:

  • “Who are my top customers?”
    • What does that even mean? Top customer could be based on total revenue, total orders, average order size, etc.
  • “Show me our ROI on ad spend from the last 3 months”
    • If ROI is not a column in your database, a calculation of multiple columns is likely required. How should that be done? Sometimes data can be messy.
  • “Show me sales from the last 30 days”
    • What does a user expect when they see this? A report of each sale that occurred, a total sales number, or sales for each day from the last 30?

While Buster will attempt to make assumptions around questions, they may not always be correct. This can be frustrating to end users. Especially because they have zero context around your database and it may be difficult for them to guide Buster toward the right answer.

📌 We recommend approving 25-100 question ↔ query pairs. Each question that you approve is sent into our auto-training pipeline and Buster immediately starts learning from it.

Fine-tuning a Response

Fine-tuning a question in Buster

Fine-tuning a question in Buster

When you ask or view a question in this space, you will be able to see the question, the generated SQL, and the data returned by the query.

You are able to edit the question and the SQL directly. You also have the capability to guide Buster toward the correct answer using the “Not quite right?” input box. In the case of the example above, I could ask Buster to break out the company sizes by 100 instead of 50.

Once you click approve, Buster immediately starts learning from the question and generated SQL code. You can always come back and edit questions, SQL, or even remove an approval. Be aware that changing the training data will not change any generated SQL for users that have previously asked questions.

User Access Controls

Buster is able to provision access controls over your data through Row Level Policies (Multi Tenant) or Permission Groups (Single Tenant). In either case, Buster implements security without the use of an LLM.