evalite（evalite 命令行）

Evalite

Evalite is a local-first TypeScript eval runner built on Vitest. Eval files use the .eval.ts extension. Results are stored in SQLite at node_modules/.evalite and viewed via a local UI at localhost:3006.

Setup

pnpm add -D evalite vitest autoevals

package.json scripts:

{ "eval:dev": "evalite watch", "eval": "evalite" }

Core API

evalite(name, opts) — define an eval

import { evalite } from "evalite";
import { Levenshtein } from "autoevals";

evalite("My Eval", {
  data: [{ input: "Hello", expected: "Hello World!" }],
  task: async (input) => input + " World!",
  scorers: [Levenshtein],
});

Generic type params: evalite<TInput, TOutput, TExpected>(name, opts)

• data — array or async function returning {input, expected?, only?}[]
• task — (input: TInput, variant: TVariant) => Promise<TOutput | AsyncIterable<TOutput>>
• scorers — array of scorer functions or inline scorer objects (optional)
• columns — (result) => RenderedColumn[] for custom UI columns (optional)
• trialCount — run each data point N times for variance measurement (optional)

data field

Can be a static array or an async function:

data: async () => [
  { input: "What is 2+2?", expected: "4" },
  { input: "Capital of France?", expected: "Paris" },
],

Use only: true on a data entry to focus on just that input during development:

data: [
  { input: "test1", expected: "out1" },
  { input: "test2", expected: "out2", only: true }, // only this runs
],

Streams

Return any AsyncIterable (including ReadableStream) from task — evalite collects chunks and joins them:

import { streamText } from "ai";
task: async (input) => {
  const result = await streamText({ model: myModel, prompt: input });
  return result.textStream;
},

Skipping

evalite.skip("Disabled Eval", { data: [], task: async () => {}, scorers: [] });

Scorers

Scores are 0 to 1 (not 0 to 100). Return a number or {score, metadata}.

Inline scorer

scorers: [
  {
    name: "Contains Paris",
    description: "Checks if output contains 'Paris'.",
    scorer: ({ input, output, expected }) => output.includes("Paris") ? 1 : 0,
  },
],

Reusable scorer with createScorer

import { createScorer } from "evalite";

const exactMatch = createScorer<string, string>({
  name: "Exact Match",
  description: "Checks exact string equality.",
  scorer: ({ output, expected }) => output === expected ? 1 : 0,
});

Generic params: createScorer<TInput, TOutput, TExpected = TOutput>(opts)

The scorer function receives {input, output, expected}. Return a number (0-1) or {score: number, metadata?: unknown}.

LLM-as-judge scorer

Use generateObject (or any LLM call) inside a scorer to get AI-powered evaluation. Return {score, metadata: {rationale}} so the reasoning shows in the UI. See references/llm-judge-example.md for a full Factuality scorer example.

autoevals library

autoevals provides ready-made scorers: Levenshtein, Factuality, and others. Install with pnpm add -D autoevals.

Traces

Track individual LLM calls within a task for debugging, token usage, and cost tracking.

Manual tracing with reportTrace

import { reportTrace } from "evalite/traces";

// Inside a task:
const start = performance.now();
const result = await myLLMCall();
reportTrace({
  start,
  end: performance.now(),
  output: result.output,
  input: [{ role: "user", content: input }],
  usage: { inputTokens: 100, outputTokens: 50, totalTokens: 150 },
});

reportTrace is a no-op outside evalite (safe to leave in production code). Multiple calls per task create multiple trace entries.

AI SDK auto-tracing with traceAISDKModel

import { traceAISDKModel } from "evalite/ai-sdk";
import { openai } from "@ai-sdk/openai";

const tracedModel = traceAISDKModel(openai("gpt-4o-mini"));
// Use tracedModel with generateText/streamText — all calls auto-traced

Also a no-op outside evalite. Works with both generateText and streamText.

Variant Comparison with evalite.each

Compare models, prompts, or configs on the same dataset:

evalite.each([
  { name: "GPT-4o mini", input: { model: "gpt-4o-mini", temp: 0.7 } },
  { name: "GPT-4o", input: { model: "gpt-4o", temp: 0.7 } },
])("Compare models", {
  data: async () => [
    { input: "Capital of France?", expected: "Paris" },
  ],
  task: async (input, variant) => {
    return generateText({
      model: openai(variant.model),
      temperature: variant.temp,
      prompt: input,
    });
  },
  scorers: [Factuality, Levenshtein],
});

The second argument to task receives variant.input from the array entry. Each variant appears as a separate eval in the UI, named "Compare models [GPT-4o mini]" etc.

Custom Columns

Override the default Input/Expected/Output columns in the UI:

columns: async ({ input, output, expected, scores, traces }) => [
  { label: "Question", value: input },
  { label: "Answer", value: output },
  { label: "Tokens", value: traces.reduce((sum, t) => sum + (t.usage?.totalTokens || 0), 0) },
],

Multi-Modal (Files)

Evalite detects Uint8Array/Buffer values anywhere in data, output, traces, or columns and saves them to node_modules/.evalite/files, rendering them in the UI.

For files on disk without reading into memory:

import { EvaliteFile } from "evalite";
data: [{ input: EvaliteFile.fromPath("path/to/image.jpg") }],

Configuration — evalite.config.ts

import { defineConfig } from "evalite/config";

export default defineConfig({
  testTimeout: 60_000,     // ms, default 30000
  maxConcurrency: 100,     // parallel test cases, default 5
  scoreThreshold: 80,      // 0-100, fail if avg score below
  hideTable: false,        // hide terminal table output
  trialCount: 3,           // run each test case N times
  setupFiles: ["dotenv/config"], // run before tests (env vars)
  server: { port: 3006 },  // UI server port
});

Per-eval trialCount overrides config-level trialCount.

Environment Variables

To load .env files, install dotenv and add setupFiles: ["dotenv/config"] to evalite.config.ts.

CLI

Programmatic API

import { runEvalite } from "evalite/runner";

await runEvalite({
  mode: "run-once-and-exit",  // or "watch-for-file-changes" | "run-once-and-serve"
  path: "my-eval.eval.ts",   // optional file filter
  cwd: "/path/to/project",   // optional, defaults to process.cwd()
  scoreThreshold: 80,        // optional
  outputPath: "./results.json", // optional JSON export
});

CI/CD

Run evals in CI with threshold enforcement and static UI export:

- run: npx evalite --threshold=70
- run: npx evalite export --output=./ui-export
- uses: actions/upload-artifact@v3
  with:
    name: evalite-ui
    path: ui-export

JSON export for programmatic analysis: evalite --outputPath=./results.json

The exported JSON is typed as Evalite.Exported.Output (import type {Evalite} from "evalite").

Caching (Important for Watch Mode)

Implement a caching layer for LLM calls in watch mode to avoid burning API credits. Wrap your model with a cache middleware — see references/caching-pattern.md for the full pattern using wrapLanguageModel from the AI SDK.

Import Map