Phoenix Evals New Metric

Creating a New Built-in Classification Evaluator

A built-in evaluator is a YAML config (source of truth) that gets compiled into Python and TypeScript code, wrapped in evaluator classes, benchmarked, and documented. The whole pipeline is linear — follow these steps in order.

Step 0: Gather Requirements

Before writing anything, clarify with the user:

1. What does this evaluator measure? Get a one-sentence description of the quality dimension. 2. What input data is available? This determines the template placeholders (e.g., {{input}}, {{output}}, {{reference}}, {{tool_definitions}}). If the user is vague, ask follow-up questions — the placeholders are the contract between the evaluator and the caller. 3. What labels make sense? Binary is most common (e.g., correct/incorrect, faithful/unfaithful), but some metrics use more. Labels map to scores. 4. Should this appear in the dataset experiments UI? If yes, it needs the promoted_dataset_evaluator label. Currently only correctness, tool_selection, and tool_invocation have this — some may new evaluators don't need it.

Step 1: Create the YAML Config

Create prompts/classification_evaluator_configs/{NAME}_CLASSIFICATION_EVALUATOR_CONFIG.yaml.

Read an existing config to match the current schema. Start with CORRECTNESS_CLASSIFICATION_EVALUATOR_CONFIG.yaml for a simple example, or TOOL_SELECTION_CLASSIFICATION_EVALUATOR_CONFIG.yaml if your evaluator needs structured span data.

Key Decision Points

`choices` — Maps label strings to numeric scores. For binary evaluators, use positive/negative labels (e.g., correct: 1.0 / incorrect: 0.0). The labels you pick here flow through to the Python class, TS factory, and benchmarks.

`optimization_direction` — Use maximize when the positive label is the desired outcome (most evaluators). Use minimize only if the metric measures something undesirable (e.g., hallucination). This affects how Phoenix displays the metric in the UI.

`labels` — Optional list. Add promoted_dataset_evaluator only if this evaluator should appear in the dataset experiments UI sidebar.

`substitutions` — Only needed if the evaluator is a promoted_dataset_evaluator and works with structured span data (tool definitions, tool calls, message arrays). These reference formatter snippets defined in prompts/formatters/server.yaml. Read that file if you need substitutions — it defines what structured data formats are available. Most evaluators that only use simple text fields (input, output, reference) don't need substitutions.

Prompt Writing Tips

• Be explicit about what makes each label correct — the LLM judge needs a clear rubric
• Separate concerns: if evaluating X, explicitly state you're NOT evaluating Y
• Wrap inputs in XML-style tags (e.g., <context>, <output>) for clear data formatting
• Tell the judge to reason before deciding — this improves accuracy
• Use {{placeholder}} (Mustache syntax) for template variables

Step 2: Compile Prompts

make codegen-prompts

This generates code in three places:

• packages/phoenix-evals/src/phoenix/evals/__generated__/classification_evaluator_configs/ (Python)
• src/phoenix/__generated__/classification_evaluator_configs/ (Python, server copy)
• js/packages/phoenix-evals/src/__generated__/default_templates/ (TypeScript)

Verify the generated files look correct before moving on.

Step 3: Create the Python Evaluator

Create packages/phoenix-evals/src/phoenix/evals/metrics/{name}.py.

Read `correctness.py` in that directory — it's the canonical example. Your evaluator follows the same pattern: subclass ClassificationEvaluator, pull constants from the generated config, define a Pydantic input schema with fields matching your template placeholders.

After creating the file, add it to the exports in metrics/__init__.py — both the import and the __all__ list. Read the current __init__.py to see the existing pattern.

Step 4: Create the TypeScript Evaluator

Create js/packages/phoenix-evals/src/llm/create{Name}Evaluator.ts.

Read `createCorrectnessEvaluator.ts` — it's the canonical example. The pattern is a factory function that wraps createClassificationEvaluator with defaults from the generated config.

Then:

1. Add the export to js/packages/phoenix-evals/src/llm/index.ts 2. Add a vitest test — read createFaithfulnessEvaluator.test.ts for the test pattern

Step 5: Build JS

cd js && pnpm build

Fix any TypeScript errors before proceeding.

Step 6: Write the Benchmark

Create js/benchmarks/evals-benchmarks/src/{name}_benchmark.ts.

Read existing benchmarks in that directory to match the current patterns:

• tool_invocation_benchmark.ts — confusion matrix printing, multi-category analysis

Benchmark Requirements

• 30-50 synthetic examples organized by category
• 2-4 examples per category covering: success cases, failure modes, and edge cases
• Accuracy evaluator that compares predicted vs expected labels
• Failed examples printer — this is critical for debugging. For each misclassified example, print: category, input, output (truncated), expected vs actual label, and the LLM judge's explanation
• Per-category accuracy breakdown in the output
• For binary evaluators, a confusion matrix is helpful

The task function must return input and output text in its result so the failed examples printer has access to them.

Consider using a separate agent session for synthetic dataset generation if the examples need realistic domain-specific content — this keeps the dataset creation focused and avoids context-switching.

Step 7: Run the Benchmark

# Terminal 1: Start Phoenix
PHOENIX_WORKING_DIR=/tmp/phoenix-test phoenix serve

# Terminal 2: Run the benchmark
cd js/benchmarks/evals-benchmarks
pnpm tsx src/{name}_benchmark.ts

Target >80% accuracy. If accuracy is low, look at the failed examples output to decide whether to adjust the prompt (Step 1) or the benchmark examples (Step 6). Iterate until accuracy is acceptable.

Step 8: Create Documentation

Create docs/phoenix/evaluation/pre-built-metrics/{name}.mdx.

Read `faithfulness.mdx` in that directory — it's the template. Follow the same section structure:

1. Overview — when to use, what it measures 2. Supported Levels — span/trace/session, relevant span kinds 3. Input Requirements — required fields table 4. Output Interpretation — labels, scores, direction 5. Usage Examples — Python and TypeScript in tabs 6. Using Input Mapping — lambda example if applicable 7. Viewing/Modifying the Prompt — link to GitHub config, custom prompt usage 8. Configuration — link to LLM config docs 9. Using with Phoenix — links to traces and experiments docs 10. Benchmarks — "Coming soon" placeholder (until benchmark results are published) 11. API Reference — links to Python and TypeScript API docs 12. Related — links to related evaluators

Navigation Updates

After creating the docs page, update these three files:

1. `docs.json` — add the page to the Evaluation > Pre-built Metrics nav group 2. `docs/phoenix/evaluation/pre-built-metrics.mdx` — add a card to the landing page grid 3. `docs/phoenix/sitemap.xml` — add the new URL

Read each file to see the existing pattern before editing.

Checklist

Before calling it done, verify:

• [ ] YAML config created with clear rubric and appropriate labels/choices
• [ ] make codegen-prompts ran successfully
• [ ] Python evaluator class with input schema matching template placeholders
• [ ] Python exports updated in metrics/__init__.py
• [ ] TypeScript evaluator factory with types
• [ ] TypeScript export added to llm/index.ts
• [ ] Vitest test for TypeScript evaluator
• [ ] JS packages rebuilt (cd js && pnpm build)
• [ ] Benchmark with 30-50 examples, category breakdown, failed examples printer
• [ ] Benchmark accuracy >80%
• [ ] Documentation page following the template structure
• [ ] docs.json nav updated
• [ ] Landing page card added
• [ ] Sitemap updated

Retrospection

After completing the workflow, verify these instructions matched reality:

• Did any file paths, export patterns, or command names change from what's described here?
• Did the YAML config schema gain or lose fields since this was written?
• Did the benchmark or docs patterns evolve from the referenced examples?
• Did make codegen-prompts generate to different locations?

If anything drifted, update this SKILL.md before finishing so the next person (or agent) doesn't hit the same surprises.