Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Regression testing (LLM apps)

Regression testing for LLM apps means re-running a fixed set of evaluation cases after every change to a prompt, model, retrieval setup or tool, to catch cases that used to work and now fail.

Also called: LLM regression testing, prompt regression testing, eval regression

By Divya Sharma · Checked Nov 15, 2026 · How we review

In ordinary software, a regression is a bug in something that used to work. LLM apps regress easily: a reworded prompt fixes one answer and breaks three others, a provider updates a model, or a new retrieval setting changes what context arrives. Regression tests catch this before users do.

How it works

Start with a golden dataset of real inputs and expected outcomes. After each change, run every case through the app and grade the outputs: with code where an exact answer exists, with LLM-as-a-judge for open-ended answers, and with people for a sample. Compare the scores with the previous run and fail the build if they drop. OpenAI's cookbook shows this for an app that summarises push notifications, checking whether a prompt change regresses its behaviour. Open-source tools such as promptfoo let you write the cases declaratively and run them in CI.

What makes it different from unit tests

Model output varies from run to run, especially at higher temperature. Run each case more than once, track pass rates rather than single results, and set thresholds instead of demanding exact matches. Pin model versions where you can, so you know whether a change came from your code or the provider.

In production

Offline tests cover known cases. LLM observability catches the ones you did not think of; add those failures to the test set so they stay fixed. Together they form the testing side of LLM evaluation.

Related terms

LLM evaluation
LLM evaluation, often called evals, is measuring how well a model or AI application performs a task by running it on a set of test cases and scoring the outputs against defined criteria, by code, by humans or by another model.
Golden dataset
A golden dataset is a fixed, carefully checked set of example inputs paired with the answers or behaviour you expect, used as the reference when you evaluate or regression-test an AI system.
LLM-as-a-judge
LLM-as-a-judge is an evaluation method where a strong language model grades the outputs of another model or application, using a rubric, a reference answer or a side-by-side comparison, in place of or alongside human reviewers.
LLM observability
LLM observability is recording what an AI application does in production, including prompts, responses, tool calls, latency, token use and cost, as traces you can search, so you can debug failures and see how the system behaves over time.
Temperature (sampling)
Temperature is a sampling setting that controls how random a language model's word choices are. Low values make it favour the most likely tokens; higher values spread choices across less likely ones.

Sources