Phoenix Evals

From the author

Build and run evaluators for AI/LLM applications using Phoenix.

— the description by jimbobbennett

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

Phoenix Evals provides a structured approach to building, running, and validating evaluators for AI and LLM applications. It supports both Python and TypeScript, offering tools for code-based and LLM-based evaluation, experiment management, and error analysis. The framework emphasizes a code-first approach, encouraging developers to define deterministic evaluations before moving to LLM-based judges, and provides mechanisms to validate evaluator accuracy against human-labeled data.

What it does

Builds custom code-based and LLM-based evaluators for AI applications.
Supports batch evaluation of datasets using structured dataframes.
Facilitates experiment management, including dataset creation and synthetic data generation.
Provides tools for error analysis, including axial coding and failure taxonomy creation.
Enables validation of evaluator accuracy against human-provided labels.
Integrates with tracing and observability to sample and review application traces.

When to use it

When you need to systematically evaluate the performance of LLM application outputs.
When building RAG systems that require specific retrieval and faithfulness metrics.
When you need to categorize and analyze failure patterns in AI agent workflows.
When transitioning from manual testing to automated, repeatable evaluation experiments.
When you need to validate that your automated judges align with human expectations.

Advantages

Supports a code-first approach to evaluation for deterministic results.
Provides a clear path for validating judge model accuracy.
Offers structured methods for building failure taxonomies.
Enables consistent evaluation across both Python and TypeScript environments.
Integrates directly with tracing data for context-aware evaluation.

Usage examples

“How do I build a custom LLM evaluator in Python?”: The skill provides guidance on using the LLM class and custom templates to define evaluators that assess specific model outputs.
“How can I categorize failures in my AI agent?”: The skill outlines an axial coding process to group failure notes into structured taxonomies like planning, tool selection, or execution errors.
“How do I run a batch evaluation on my dataset?”: The skill demonstrates using the evaluate_dataframe function to process a dataframe and return scores for defined metrics.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/github/phoenix-evals/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip phoenix-evals.zip -d .agents/skills

Documentation by jimbobbennett

Build evaluators for AI/LLM applications. Code first, LLM for nuance, validate against humans.

Quick Reference

Task	Files
Setup	setup-python, setup-typescript
Decide what to evaluate	evaluators-overview
Choose a judge model	fundamentals-model-selection
Use pre-built evaluators	evaluators-pre-built
Build code evaluator	evaluators-code-python, evaluators-code-typescript
Build LLM evaluator	evaluators-llm-python, evaluators-llm-typescript, evaluators-custom-templates
Batch evaluate DataFrame	evaluate-dataframe-python
Understand experiments	experiments-overview
Run experiment	experiments-running-python, experiments-running-typescript
Create dataset	experiments-datasets-python, experiments-datasets-typescript
Generate synthetic data	experiments-synthetic-python, experiments-synthetic-typescript
Validate evaluator accuracy	validation, validation-evaluators-python, validation-evaluators-typescript
Sample traces for review	observe-sampling-python, observe-sampling-typescript
Analyze errors	error-analysis, error-analysis-multi-turn, axial-coding
RAG evals	evaluators-rag
Avoid common mistakes	common-mistakes-python, fundamentals-anti-patterns
Production	production-overview, production-guardrails, production-continuous

Workflows

Starting Fresh: observe-tracing-setup → error-analysis → axial-coding → evaluators-overview

Building Evaluator: fundamentals → common-mistakes-python → evaluators-{code|llm}-{python|typescript} → validation-evaluators-{python|typescript}

RAG Systems: evaluators-rag → evaluators-code-* (retrieval) → evaluators-llm-* (faithfulness)

Production: production-overview → production-guardrails → production-continuous

Reference Categories

Prefix	Description
`fundamentals-*`	Types, scores, anti-patterns
`observe-*`	Tracing, sampling
`error-analysis-*`	Finding failures
`axial-coding-*`	Categorizing failures
`evaluators-*`	Code, LLM, RAG evaluators
`experiments-*`	Datasets, running experiments
`validation-*`	Validating evaluator accuracy against human labels
`production-*`	CI/CD, monitoring

Key Principles

Principle	Action
Error analysis first	Can't automate what you haven't observed
Custom > generic	Build from your failures
Code first	Deterministic before LLM
Validate judges	>80% TPR/TNR
Binary > Likert	Pass/fail, not 1-5

Files

ATTRIBUTION.mdAdded by AutoGineer931 B
LICENSEAdded by AutoGineer10.0 KiB
SKILL.md4.5 KiB
references/axial-coding.md2.3 KiB
references/common-mistakes-python.md7.0 KiB
references/error-analysis-multi-turn.md1.3 KiB
references/error-analysis.md4.2 KiB
references/evaluate-dataframe-python.md4.4 KiB
references/evaluators-code-python.md3.0 KiB
references/evaluators-code-typescript.md1.2 KiB
references/evaluators-custom-templates.md1.2 KiB
references/evaluators-llm-python.md2.6 KiB
references/evaluators-llm-typescript.md1.4 KiB
references/evaluators-overview.md1.1 KiB
references/evaluators-pre-built.md2.4 KiB
references/evaluators-rag.md3.1 KiB
references/experiments-datasets-python.md4.8 KiB
references/experiments-datasets-typescript.md2.7 KiB
references/experiments-overview.md1.7 KiB
references/experiments-running-python.md2.9 KiB
references/experiments-running-typescript.md3.1 KiB
references/experiments-synthetic-python.md1.8 KiB
references/experiments-synthetic-typescript.md2.2 KiB
references/fundamentals-anti-patterns.md1.6 KiB
references/fundamentals-model-selection.md1.4 KiB
references/fundamentals.md1.8 KiB
references/observe-sampling-python.md2.4 KiB
references/observe-sampling-typescript.md3.9 KiB
references/observe-tracing-setup.md3.3 KiB
references/production-continuous.md3.6 KiB
references/production-guardrails.md1.3 KiB
references/production-overview.md2.3 KiB
references/setup-python.md1.9 KiB
references/setup-typescript.md981 B
references/validation-evaluators-python.md1.0 KiB
references/validation-evaluators-typescript.md4.4 KiB
references/validation.md1.7 KiB

references/evaluators-rag.md

Evaluators: RAG Systems

RAG has two distinct components requiring different evaluation approaches.

Two-Phase Evaluation

RETRIEVAL                    GENERATION
─────────                    ──────────
Query → Retriever → Docs     Docs + Query → LLM → Answer
         │                              │
    IR Metrics              LLM Judges / Code Checks

Debug retrieval first using IR metrics, then tackle generation quality.

Retrieval Evaluation (IR Metrics)

Use traditional information retrieval metrics:

Metric	What It Measures
Recall@k	Of all relevant docs, how many in top k?
Precision@k	Of k retrieved docs, how many relevant?
MRR	How high is first relevant doc?
NDCG	Quality weighted by position

# Requires query-document relevance labels
def recall_at_k(retrieved_ids, relevant_ids, k=5):
    retrieved_set = set(retrieved_ids[:k])
    relevant_set = set(relevant_ids)
    if not relevant_set:
        return 0.0
    return len(retrieved_set & relevant_set) / len(relevant_set)

Creating Retrieval Test Data

Generate query-document pairs synthetically:

# Reverse process: document → questions that document answers
def generate_retrieval_test(documents):
    test_pairs = []
    for doc in documents:
        # Extract facts, generate questions
        questions = llm(f"Generate 3 questions this document answers:\n{doc}")
        for q in questions:
            test_pairs.append({"query": q, "relevant_doc_id": doc.id})
    return test_pairs

Generation Evaluation

Use LLM judges for qualities code can't measure:

Eval	Question
Faithfulness	Are all claims supported by retrieved context?
Relevance	Does answer address the question?
Completeness	Does answer cover key points from context?

from phoenix.evals import ClassificationEvaluator, LLM

FAITHFULNESS_TEMPLATE = """Given the context and answer, is every claim in the answer supported by the context?

<context>{{context}}</context>
<answer>{{output}}</answer>

"faithful" = ALL claims supported by context
"unfaithful" = ANY claim NOT in context

Answer (faithful/unfaithful):"""

faithfulness = ClassificationEvaluator(
    name="faithfulness",
    prompt_template=FAITHFULNESS_TEMPLATE,
    llm=LLM(provider="openai", model="gpt-4o"),
    choices={"unfaithful": 0, "faithful": 1}
)

RAG Failure Taxonomy

Common failure modes to evaluate:

retrieval_failures:
  - no_relevant_docs: Query returns unrelated content
  - partial_retrieval: Some relevant docs missed
  - wrong_chunk: Right doc, wrong section

generation_failures:
  - hallucination: Claims not in retrieved context
  - ignored_context: Answer doesn't use retrieved docs
  - incomplete: Missing key information from context
  - wrong_synthesis: Misinterprets or miscombines sources

Evaluation Order

Retrieval first - If wrong docs, generation will fail
Faithfulness - Is answer grounded in context?
Answer quality - Does answer address the question?

Fix retrieval problems before debugging generation.

Requirements and compatibility

Written for: GitHub Copilot

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Python

The author says: Requires Phoenix server. Python skills need phoenix and openai packages; TypeScript skills need @arizeai/phoenix-client.

Security

What our scanner noticed· Low risk

Instructions that use the network

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Migrate To Shoehorn
Matt Pocock· Community
Automate the migration of TypeScript test assertions from unsafe as-casting to the type-safe shoehorn library.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository317.5k
Tdd
Matt Pocock· Community
A reference guide for implementing the red-green-refactor loop in test-driven development.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository317.5k
Test Driven Development
Jesse Vincent (obra)· Community
A rigorous framework for Test-Driven Development, mandating that failing tests precede all production code.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository314k
Webapp Testing
AnthropicOfficial
A toolkit for automating local web application testing using Playwright and server management helpers.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
Apache-2.0Low riskHas scriptsGitHub stars of the source repository188.6k
Test Driven Development
Addy Osmani· Community
A systematic approach to software development using the red-green-refactor cycle to ensure code reliability and maintainability.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository113.9k
Browser Testing With Devtools
Addy Osmani· Community
Integrate Chrome DevTools to enable live browser inspection, debugging, and automated UI verification for your AI agent.
Software developmentClaude CodeClaude appsOpenAI Codex+14 more
MITGitHub stars of the source repository113.9k

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download phoenix-evals.zip