Evals for AI Engineers

Chapter 3. Error Analysis

A Note for Early Release Readers

With Early Release ebooks, you get books in their earliest form—the author’s raw and unedited content as they write—so you can take advantage of these technologies long before the official release of these titles.

This will be the 3rd chapter of the final book. If you’d like to be actively involved in reviewing and commenting on this draft, please reach out to the editor at [e-mail].

The Analyze step of the evaluation lifecycle is where we look at what the system is actually doing and figure out where it fails. This chapter covers the practical methods for doing that.

In this chapter, you will learn:

• How to read and inspect LLM traces systematically

• How to identify and categorize failure modes

• How to build a failure taxonomy by reading traces and clustering observations

• How to generate synthetic data for testing

The process of developing robust evaluations for LLM applications is inherently iterative. It involves creating test cases, assessing performance, and refining the system based on those observations. High-level guides, such as Anthropic’s documentation on creating empirical evaluations for Claude Anthropic 2024, often depict the evaluation process as a cycle of developing test cases, engineering prompts, testing, and refining (Figure 3-1).¹

Figure 3-1. Anthropic’s visualization of the iterative ...