Online evaluations

​

Introduction

Online evaluations is the concept of running some code or LLM as a judge on the results of your LLM calls in real-time. This is useful to gather statistics about the performance of your LLMs.

Laminar achieves that by analyzing the inputs and outputs of your LLM spans.

​

Why do we need online evaluations?

Online evaluations allow you to gather statistics about the performance of your LLMs in production. We like to think of online evaluations as almost like canaries or metrics in traditional applications, but for unstructured data.

This approach allows you to constantly monitor your apps without having to trigger evaluations on your own.

​

Key concepts

• Span path – this is an identifier of your LLM function. It is constructed from the location of the call within your code, and must ideally be unique. For more information about what a span is, see the tracing documentation.
• Span label – a label attached to a span. It can be a boolean or a categorical label. The label can be both set manually or by an evaluator. For more information about span labels and label classes, see the label spans documentation.
• Evaluator – the function evaluating the input and output of your LLM call. It can be a code snippet or a prompt for LLM as a judge.

You can register one evaluator per label class. For example, you could have a code evaluator check for the presence of a certain keyword in the output of your (label class keyword) and an LLM evaluator that checks the sentiment of the output (label class sentiment). In that case, you would not be able to add another evaluator for neither the sentiment nor the keyword label class.

Every span path can have multiple span labels, but only one evaluator per label class.

​

Creating an online evaluation

Once you have sent your first trace to Laminar, you can create labels to attach to spans.

1. Go to “Traces” in the Laminar dashboard.
2. Open the span you want to evaluate and click “Add Label”.
3. Create a new label class or choose an existing one.
4. Click the 3 vertical dots next to the label class and select “Add evaluator”.
5. You will see the span input and span output
6. You can select between LLM as a judge and Python, and configure the prompt and code respectively.
   • When working with prompt, the output schema is controlled by BAML, which you can read about in their docs
   • For code, make sure that the output is one of the possible label values.
7. Click the play button to test the evaluator.
8. Once you are happy, click “Save online evaluator”.
9. If you want to disable online evaluator on this path, switch the toggle under “Run on current span” in the “Add label” menu.