@arizeai/phoenix-evals 0.2.1 → 0.3.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@arizeai/phoenix-evals",
-  "version": "0.2.1",
+  "version": "0.3.0",
   "description": "A library for running evaluations for AI use cases",
   "main": "dist/src/index.js",
   "module": "dist/esm/index.js",
@@ -34,23 +34,25 @@
   ],
   "keywords": [
     "evals",
+    "arize",
     "phoenix",
     "llm",
     "evaluation"
   ],
   "author": "oss@arize.com",
-  "license": "ELv2",
+  "license": "Apache-2.0",
   "devDependencies": {
     "@ai-sdk/openai": "^2.0.0",
     "@arizeai/openinference-instrumentation-openai": "^2.3.0",
     "@types/mustache": "^4.2.6",
     "@types/node": "^24.0.10",
+    "msw": "^2.11.6",
     "nock": "^14.0.5",
     "tsx": "^4.19.3",
     "typedoc": "^0.27.9",
     "typescript": "^5.8.2",
     "vitest": "^2.1.9",
-    "@arizeai/phoenix-client": "4.0.3"
+    "@arizeai/phoenix-client": "5.2.0"
   },
   "engines": {
     "node": ">=18"
@@ -65,7 +67,7 @@
     "clean": "rimraf dist",
     "prebuild": "pnpm run clean",
     "build": "tsc --build tsconfig.json tsconfig.esm.json && tsc-alias -p tsconfig.esm.json",
-    "postbuild": "echo '{\"type\": \"module\"}' > ./dist/esm/package.json && rimraf dist/test dist/examples",
+    "postbuild": "echo '{\"type\": \"module\"}' > ./dist/esm/package.json",
     "type:check": "tsc --noEmit",
     "test": "vitest --typecheck",
     "docs": "typedoc",

package/src/llm/ClassificationEvaluator.ts

@@ -1,11 +1,12 @@
+import { getTemplateVariables } from "../template";
 import {
   CreateClassificationEvaluatorArgs,
   EvaluatorFn,
   Template,
 } from "../types";
+
 import { createClassifierFn } from "./createClassifierFn";
 import { LLMEvaluator } from "./LLMEvaluator";
-import { getTemplateVariables } from "../template";
 
 /**
  * An LLM evaluator that performs evaluation via classification

package/src/llm/LLMEvaluator.ts

@@ -1,8 +1,8 @@
 import {
+  CreateEvaluatorArgs,
   EvaluationResult,
   Evaluator,
   OptimizationDirection,
-  CreateEvaluatorArgs,
 } from "../types";
 
 /**
@@ -12,7 +12,7 @@ export abstract class LLMEvaluator<RecordType extends Record<string, unknown>>
   implements Evaluator<RecordType>
 {
   readonly name: string;
-  readonly source = "LLM" as const;
+  readonly kind = "LLM" as const;
   readonly optimizationDirection?: OptimizationDirection;
   constructor({ name, optimizationDirection }: CreateEvaluatorArgs) {
     this.name = name;

package/src/llm/createClassificationEvaluator.ts

@@ -1,4 +1,5 @@
 import { CreateClassificationEvaluatorArgs } from "../types/evals";
+
 import { ClassificationEvaluator } from "./ClassificationEvaluator";
 
 export function createClassificationEvaluator<

package/src/llm/createClassifierFn.ts

@@ -1,11 +1,12 @@
+import { formatTemplate } from "../template";
 import {
   ClassificationChoicesMap,
-  EvaluationResult,
   CreateClassifierArgs,
+  EvaluationResult,
   EvaluatorFn,
 } from "../types/evals";
+
 import { generateClassification } from "./generateClassification";
-import { formatTemplate } from "../template";
 
 /**
  * Convert a mapping of choices to labels

package/src/llm/createDocumentRelevancyEvaluator.ts

@@ -1,8 +1,9 @@
-import { CreateClassificationEvaluatorArgs, Evaluator } from "../types/evals";
 import {
-  DOCUMENT_RELEVANCY_TEMPLATE,
   DOCUMENT_RELEVANCY_CHOICES,
+  DOCUMENT_RELEVANCY_TEMPLATE,
 } from "../default_templates/DOCUMENT_RELEVANCY_TEMPLATE";
+import { CreateClassificationEvaluatorArgs, Evaluator } from "../types/evals";
+
 import { createClassificationEvaluator } from "./createClassificationEvaluator";
 
 export interface DocumentRelevancyEvaluatorArgs

package/src/llm/createHallucinationEvaluator.ts

@@ -1,10 +1,11 @@
-import { CreateClassificationEvaluatorArgs } from "../types/evals";
 import {
-  HALLUCINATION_TEMPLATE,
   HALLUCINATION_CHOICES,
+  HALLUCINATION_TEMPLATE,
 } from "../default_templates/HALLUCINATION_TEMPLATE";
-import { createClassificationEvaluator } from "./createClassificationEvaluator";
+import { CreateClassificationEvaluatorArgs } from "../types/evals";
+
 import { ClassificationEvaluator } from "./ClassificationEvaluator";
+import { createClassificationEvaluator } from "./createClassificationEvaluator";
 
 export interface HallucinationEvaluatorArgs
   extends Omit<

package/src/llm/generateClassification.ts

@@ -1,9 +1,10 @@
+import { tracer } from "../telemetry";
 import { ClassificationResult, WithLLM } from "../types/evals";
 import { WithTelemetry } from "../types/otel";
 import type { WithPrompt } from "../types/prompts";
+
 import { generateObject } from "ai";
 import { z } from "zod";
-import { tracer } from "../telemetry";
 export type ClassifyArgs = WithLLM &
   WithTelemetry &
   WithPrompt & {

package/src/template/applyTemplate.ts

@@ -1,4 +1,5 @@
 import { Template } from "../types/templating";
+
 import Mustache from "mustache";
 
 /**

package/src/template/getTemplateVariables.ts

@@ -1,4 +1,5 @@
 import { Template } from "../types/templating";
+
 import Mustache from "mustache";
 
 type GetTemplateVariableArgs = {

package/src/types/evals.ts

@@ -1,6 +1,7 @@
-import { LanguageModel } from "ai";
 import { WithTelemetry } from "./otel";
 
+import { LanguageModel } from "ai";
+
 /**
  * A specific AI example that is under evaluation
  */
@@ -101,9 +102,9 @@ export type EvaluatorFn<ExampleType extends Record<string, unknown>> = (
 ) => Promise<EvaluationResult>;
 
 /**
- * The source of the evaluation
+ * The kind of the evaluation
  */
-export type EvaluationSource = "LLM" | "CODE";
+export type EvaluationKind = "LLM" | "CODE";
 
 /**
  * The direction to optimize the numeric evaluation score
@@ -120,9 +121,9 @@ interface EvaluatorDescription {
    */
   name: string;
   /**
-   * The source of the evaluation. Also known as the "kind" of evaluator.
+   * The kind of the evaluation. Also known as the "kind" of evaluator.
    */
-  source: EvaluationSource;
+  kind: EvaluationKind;
   /**
    * The direction to optimize the numeric evaluation score
    * E.x. "MAXIMIZE" means that the higher the score, the better the evaluation