writing-evals（写评估）

Writing Evals

You write evaluations that prove AI capabilities work. Evals are the test suite for non-deterministic systems: they measure whether a capability still behaves correctly after every change.

Prerequisites

• Complete the Axiom AI SDK Quickstart (instrumentation + authentication)

Verify the SDK is installed:

ls node_modules/axiom/dist/

If not installed, install it using the project's package manager (e.g., pnpm add axiom).

Always check node_modules/axiom/dist/docs/ first for the correct API signatures, import paths, and patterns for the installed SDK version. The bundled docs are the source of truth — do not rely on the examples in this skill if they conflict.

Philosophy

1. Evals are tests for AI. Every eval answers: "does this capability still work?"
2. Scorers are assertions. Each scorer checks one property of the output.
3. Flags are variables. Flag schemas let you sweep models, temperatures, strategies without code changes.
4. Data drives coverage. Happy path, adversarial, boundary, and negative cases.
5. Validate before running. Never guess import paths or types—use reference docs.

Axiom Terminology

How to Start

When the user asks you to write evals for an AI feature, read the code first. Do not ask questions — inspect the codebase and infer everything you can.

Step 1: Understand the feature

1. Find the AI function — search for the function the user mentioned. Read it fully.
2. Trace the inputs — what data goes in? A string prompt, structured object, conversation history?
3. Trace the outputs — what comes back? A string, category label, structured object, agent result with tool calls?
4. Identify the model call — which LLM/model is used? What parameters (temperature, maxTokens)?
5. Check for existing evals — search for *.eval.ts files. Don't duplicate what exists.
6. Check for app-scope — look for createAppScope, flagSchema, axiom.config.ts.

Step 2: Determine eval type

Based on what you found:

Step 3: Choose scorers

Every eval needs at least 2 scorers. Use this layering:

1. Correctness scorer (required) — Does the output match expected? Pick from the eval type table above (exact match, set match, field match, etc.).
2. Quality scorer (recommended) — Is the output well-formed? Check confidence thresholds, output length, format validity, or field completeness.
3. Reference-free scorer (add for user-facing text) — Is the output coherent, relevant, non-toxic? Use LLM-as-judge or autoevals.

Step 4: Generate

1. Create the .eval.ts file colocated next to the source file
2. Import the actual function — do not create a stub
3. Write the scorers based on the output type (minimum 2, see step 3)
4. Generate test data (see Data Design Guidelines)
5. Set capability and step names matching the feature's purpose
6. If flags exist, use pickFlags to scope them

Only ask if you cannot determine:

• What "correct" means for ambiguous outputs (e.g., summarization quality)
• Whether the user wants pass/fail or partial credit scoring
• Which parameters should be tunable via flags (if not already using flags)

Project Layout

Recommended: Colocated with source

Place .eval.ts files next to their implementation files, organized by capability:

src/
├── lib/
│   ├── app-scope.ts
│   └── capabilities/
│       └── support-agent/
│           ├── support-agent.ts
│           ├── support-agent-e2e-tool-use.eval.ts
│           ├── categorize-messages.ts
│           ├── categorize-messages.eval.ts
│           ├── extract-ticket-info.ts
│           └── extract-ticket-info.eval.ts
axiom.config.ts
package.json

Minimal: Flat structure

For small projects, keep everything in src/:

src/
├── app-scope.ts
├── my-feature.ts
└── my-feature.eval.ts
axiom.config.ts
package.json

The default glob **/*.eval.{ts,js} discovers eval files anywhere in the project. axiom.config.ts always lives at the project root.

Eval File Structure

Standard structure of an eval file:

import { pickFlags } from '@/app-scope';       // or relative path
import { Eval } from 'axiom/ai/evals';
import { Scorer } from 'axiom/ai/scorers';
import { Mean, PassHatK } from 'axiom/ai/scorers/aggregations';
import { myFunction } from './my-function';

const MyScorer = Scorer('my-scorer', ({ output, expected }: { output: string; expected: string }) => {
  return output === expected;
});

Eval('my-eval-name', {
  capability: 'my-capability',
  step: 'my-step',                              // optional
  configFlags: pickFlags('myCapability'),        // optional, scopes flag access
  data: [
    { input: '...', expected: '...', metadata: { purpose: '...' } },
  ],
  task: async ({ input }) => {
    return await myFunction(input);
  },
  scorers: [MyScorer],
});

Reference

For detailed patterns and type signatures, read these on demand:

• reference/scorer-patterns.md — All scorer patterns (exact match, set match, structured, tool use, autoevals, LLM-as-judge), score return types, typing tips
• reference/api-reference.md — Full type signatures, import paths, aggregations, streaming tasks, dynamic data loading, manual token tracking, CLI options
• reference/flag-schema-guide.md — Flag schema rules, validation, pickFlags, CLI overrides, common patterns
• reference/templates/ — Ready-to-use eval file templates (see Templates section below)

Authentication Setup

Before running evals, the user must authenticate. Check if they've already done this before suggesting it.

Set environment variables (works for both offline and online evals). Store in .env at the project root:

AXIOM_URL="https://api.axiom.co"
AXIOM_TOKEN="API_TOKEN"
AXIOM_DATASET="DATASET_NAME"
AXIOM_ORG_ID="ORGANIZATION_ID"

CLI Reference

Data Design Guidelines

Step 1: Check for existing data

Before generating test data, check if the user already has data:

1. Ask the user — "Do you have an eval dataset, test cases, or example inputs/outputs?"
2. Search the codebase — look for JSON/CSV files, seed data, test fixtures, or existing data: arrays in other eval files
3. Check for production logs — the user may have real inputs in Axiom that can be exported

If the user has data, use it directly in the data: array or load it with dynamic data loading (data: async () =>...).

Step 2: Generate test data from code

If no data exists, generate it by reading the AI feature's code:

1. Read the system prompt — it defines what the feature does and what outputs are valid. Extract the categories, labels, or expected behavior it describes.
2. Read the input type — understand what shape of data the function accepts. Generate realistic examples of that shape.
3. Read any validation/parsing — if the code parses or validates output, that tells you what correct output looks like.
4. Look at enum values or constants — if the feature classifies into categories, use those as expected values.

Step 3: Cover all categories

Generate at least one case per category:

Minimum: 5-8 cases for a basic eval. 15-20 for production coverage.

Metadata Convention

Always add metadata: {purpose: '...'} to each test case for categorization.

Scripts

eval-scaffold types

Workflow

1. Initialize: scripts/eval-init to create app-scope + config
2. Scaffold: scripts/eval-scaffold <type> <capability> [step]
3. Customize: replace TODO placeholders with real data and function
4. Validate: scripts/eval-validate <file> to check structure
5. Coverage: scripts/eval-add-cases <file> to find gaps
6. Test: npx axiom eval --debug for local run
7. Deploy: npx axiom eval to send results to Axiom
8. Review: scripts/eval-results <deployment> to query results from Axiom

Online Evals (Production)

Online evaluations score your AI capability's outputs on live production traffic. Unlike offline evals that run against a fixed collection with expected values, online evals are reference-free — scorers receive input and output but no expected.

Use online evals to: monitor quality in production, catch format regressions, run heuristic checks, or sample traffic for LLM-as-judge scoring without affecting your capability's response.

When to use online vs offline

Import paths

import { onlineEval } from 'axiom/ai/evals/online';
import { Scorer } from 'axiom/ai/scorers';

Function signature

onlineEval takes a mandatory name (first arg) and params:

void onlineEval('my-eval-name', {
  capability: 'qa',
  step: 'answer',           // optional
  input: userMessage,        // optional, passed to scorers
  output: response.text,
  scorers: [formatScorer],
});

Name must match [A-Za-z0-9\-_] only.

Online scorers use the same Scorer API as offline (see reference/scorer-patterns.md), but are reference-free — they receive input and output but no expected. Online evals never throw errors into your app's code; scorer failures are recorded on the eval span as OTel events.

Key differences from offline: per-scorer sampling (number or async function), trace linking via links param or auto-detection inside withSpan, and fire-and-forget (void) vs await for short-lived processes.

Before writing online eval code, always read the SDK's bundled docs first — they match the installed version and contain the latest API, parameters, and patterns:

cat node_modules/axiom/dist/docs/evals/online/functions/onlineEval.md

Common Pitfalls

API Documentation Lookup

For exact type signatures, check the SDK's bundled docs first (matches the installed version):

ls node_modules/axiom/dist/docs/

Key paths:

• node_modules/axiom/dist/docs/evals/functions/Eval.md
• node_modules/axiom/dist/docs/scorers/scorers/functions/Scorer.md
• node_modules/axiom/dist/docs/evals/online/functions/onlineEval.md
• node_modules/axiom/dist/docs/scorers/aggregations/README.md
• node_modules/axiom/dist/docs/config/README.md