Overview

@arizeai/phoenix-evals provides evaluator building blocks for TypeScript workflows. It includes LLM-based evaluators, code-based evaluators, prompt templating helpers, and compatibility points for Phoenix experiments.

Install#

@arizeai/phoenix-evals depends on model adapters from the AI SDK ecosystem. Install the package plus at least one provider adapter for the models you plan to use.

npm install @arizeai/phoenix-evals

Common Setups#

npm install @arizeai/phoenix-evals @ai-sdk/openai

npm install @arizeai/phoenix-evals @ai-sdk/google

You can also pair it with Phoenix experiments:

npm install @arizeai/phoenix-evals @arizeai/phoenix-client @ai-sdk/openai

Runtime Expectations#

• Node.js 22.12+ (required by AI SDK v7, which is ESM-only)
• an AI SDK v7-compatible provider package such as @ai-sdk/openai (v4+)
• credentials required by your chosen provider

Minimal Example#

import { openai } from "@ai-sdk/openai";import { createFaithfulnessEvaluator } from "@arizeai/phoenix-evals";const faithfulness = createFaithfulnessEvaluator({  model: openai("gpt-4o-mini"),});const result = await faithfulness.evaluate({  input: "What is Phoenix?",  context: "Phoenix is an open-source AI observability platform from Arize.",  output: "Phoenix is an open-source AI observability platform from Arize.",});

Docs And Source In node_modules#

After install, a coding agent can inspect the installed package directly:

node_modules/@arizeai/phoenix-evals/docs/node_modules/@arizeai/phoenix-evals/src/

The bundled docs cover evaluator creation, LLM evaluators, classification metrics, templates, classification, and Phoenix integration.

Where To Start#

• Create evaluator for custom and code-based evaluator flows
• LLM evaluators and Classification for model-backed evaluation
• Classification metrics for precision/recall/F-beta code evaluators
• Templates and Phoenix integration for prompt helpers and experiment wiring

Source Layout#

• src/index.ts re-exports the package surface you usually import from @arizeai/phoenix-evals
• src/llm/ contains classification helpers and built-in LLM evaluator factories
• src/code/ contains deterministic classification-metric evaluators (precision, recall, F-beta)
• src/helpers/ contains createEvaluator and evaluation-result helpers
• src/template/ contains formatTemplate and getTemplateVariables
• src/types/ contains shared evaluator and prompt types