Test a prompt

Run prompts over your data; compare versions in the Playground before shipping to production.

Testing a prompt means running it against your data and scoring outputs with evaluators. You can replay production traces, compare versions side by side, and catch regressions before anything ships.

Once a prompt is saved to Prompt Hub, attach a dataset, run the model, and open View Experiment for a full breakdown of results.

For the conceptual model — what the Playground is for, the three modes (run on dataset, replay on spans, side-by-side compare) — see The Prompt Playground.

Workflow

By Arize Skills

Coming soon!

By Alyx

1. Load or generate data

   Open Alyx and load your data, or ask it to generate examples from your use case. You can also start from a production span—Alyx has the full context automatically.

2. Load your prompt

   Ask Alyx to load a prompt from Prompt Hub. For dataset runs, make sure to align `{variables}` referenced in your prompt with dataset columns.

3. Iterate until satisfied

   Refine in Alyx until the prompt matches what you want to test.

4. Create an evaluator

   Ask Alyx to draft an evaluator. Describe what you want to evaluate and it sets up the labels and criteria.

5. Run the experiment

   Ask Alyx to run the experiment.

6. Review results

   Open Datasets & Experiments to compare outputs, latency, tokens, and scores from evaluators. You can also ask Alyx to summarize the results for you.

By UI

On a dataset

When modifying a prompt in the Playground, you can test your new prompt across a dataset of examples to validate that the model is hill climbing in terms of performance across challenging examples, without regressing on core business use cases.

1. Create a dataset

   Follow Build a dataset to upload your dataset to Arize AX.

2. Select your dataset in the Playground

   Go back to the Prompt Playground and choose your dataset from the Select a Dataset dropdown.

3. Load or write your prompt

   Load a prompt from Prompt Hub using the Select a template from Prompt Hub dropdown. Or fill in a new prompt (see Build a prompt). Include variables from your dataset in the prompt inside curly braces (for example `{destination}`).

4. Attach evaluators

   Evaluators score each row. Attach an evaluator: code-based for deterministic checks, or LLM-as-a-Judge for qualitative scoring.

5. Run and review

   Click Run. The Playground fills each row’s variables, calls your model, and scores the output. Open View Experiment to compare results and look for patterns. For each row, the Playground substitutes values into the template, calls your model, and runs every attached evaluator on the generated output.

On a span (replay)

Load any production span directly into the Prompt Playground. All parameters, messages, variables, and function definitions are auto-populated so you can iterate on the exact call that went wrong. See production replay for the full walkthrough.

Span replay bridges production traces and prompt iteration: you get the real execution context instead of guessing what caused a response—the same inputs, variables, and function calls you saw in production, ready to experiment on.

With span replay, you can:

• Debug in context — Reproduce the precise LLM invocation that occurred in production, including all inputs and configurations.
• Iterate instantly — Adjust prompts, parameters, or models and re-run them on the same real data, no manual reconstruction needed.
• Validate improvements — Compare responses side by side and ensure changes lead to measurable quality gains.
• Accelerate experimentation — Turn traces into actionable testing environments for faster, data-driven prompt refinement.

In short, span replay turns tracing data into an interactive feedback loop—connecting observation (what happened) with optimization (how to make it better).

Multiple prompts at once

Compare up to three prompts side by side against the same dataset and evaluators. Use the + button to add prompt objects. See compare prompts side by side.

With tool calls

Tools drive agentic workflows, but debugging them is fiddly: unclear descriptions or parameter specs, wrong argument values, the LLM skipping or mis-picking a tool, or a missing tool for the user input.

The Prompt Playground can pull tools from span data automatically (easiest when function calls are traced), so you iterate on prompts in the same tool setup you saw in production. Open Functions to edit or add tool definitions as JSON (use a list of objects for multiple tools) and configure Function Selection so the model gets clearer instructions on when to invoke each tool.

See using tools in the Playground.

With image inputs

Multimodal models such as gpt-4o accept image URLs as variables. Add an `{image}` variable to your prompt, paste a URL into Input Variables, and run. Supported formats: JPG, PNG, GIF, SVG, WebP. Supported URL types: https://, gs://, s3://, base64. See image inputs in the Playground.

By Code

To run the same kind of checks in code - over every dataset row, with scores from evaluators logged and compared in Datasets & Experiments - use the Experiments docs:

• Set up an experiment - overview of datasets, runs, and comparisons
• Run an experiment - execute a task against a dataset
• Log experiment results - record outputs and metrics

For task-based runs that go beyond a single LLM call in the Playground, see Experiment in code.

Next up

Once you have runs and eval signals you trust, Optimize a prompt to turn that feedback into an improved template.