> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sutro.sh/llms.txt
> Use this file to discover all available pages before exploring further.

# Build your first Function

> Teach Sutro a binary decision from a representative dataset.

# Build your first Function

This walkthrough creates a binary classifier in the Sutro app. The same loop
applies to judges and extractors.

## 1. Prepare a dataset

Create a CSV with one row per decision. Start with representative, unlabeled
production data; around 1,000 rows is a useful default.

```csv theme={null}
text
"The customer asks to reset a password."
"The customer reports an unrecognized charge."
```

Sutro accepts CSV, JSON, JSONL, and Parquet datasets.

## 2. Create the Function

In [the Sutro app](https://app.sutro.sh), open **Functions**, select
**Create**, and configure:

* **Type:** Binary classification
* **Task definition:** `Should this request be escalated to a human?`
* **Dataset:** your uploaded file
* **Input column:** `text`

Keep the task definition short and use your own language. Your annotations
will teach Sutro the detailed decision rules.

## 3. Review the hard cases

Sutro runs an ensemble over the dataset and opens an annotation workspace.
Review:

* **Low confidence:** cases where the models disagree
* **High confidence:** cases where the models agree, included as a check
* **Held-out:** fixed evaluation cases, if you enabled a held-out set

Choose the correct label and add a rationale when the decision depends on
policy, context, or expert judgment.

## 4. Optimize and repeat

Complete the iteration to generate an improved prompt. Compare the prompt and
metrics with the previous iteration, then run another iteration if important
errors remain.

When the Function is ready, use **Model Sweep** to compare deployment models,
then open **Integrate** to copy a working SDK example.

<Card title="Run the Function" icon="play" href="/functions/run">
  Invoke the published Function from Python.
</Card>