EmbedEval v2

Binary evals. Trace-centric. Error-analysis-first.

A command-line tool for evaluating LLM outputs using binary pass/fail judgments, built on Hamel Husain's evaluation principles.

📖 Full LLM Guide | 🚀 Quick Start | 📊 GitHub Pages

Why EmbedEval?

Most teams struggle with LLM evaluation because they:

❌ Use complex 1-5 scales (hard to agree on)
❌ Skip manual error analysis (miss critical failures)
❌ Start with expensive LLM-as-judge (waste money)
❌ Build evals before understanding failures (measure wrong things)

EmbedEval fixes this with Hamel Husain's proven approach:

✅ Binary only - PASS or FAIL, no debating
✅ Error analysis first - Look at traces before automating
✅ Cheap evals first - Assertions before LLM-as-judge
✅ Trace-centric - Complete session records
✅ Single annotator - "Benevolent dictator" model

Quick Start

Install

Option 1: Quick Install (Recommended)

curl -fsSL https://raw.githubusercontent.com/Algiras/embedeval/main/install.sh | bash

Option 2: npm (Global)

npm install -g embedeval

Option 3: npx (No Install)

npx embedeval <command>

3-Command Workflow

# 1. COLLECT - Import your LLM traces
embedeval collect ./production-logs.jsonl --output traces.jsonl

# 2. ANNOTATE - Manual error analysis (30 min for 50-100 traces)
embedeval annotate traces.jsonl --user "expert@company.com"

# 3. TAXONOMY - Build failure taxonomy
embedeval taxonomy build --annotations annotations.jsonl

That's it. You'll now see:

Pass Rate: 73%

Top Failure Categories:
1. Hallucination: 12 traces (44%)
2. Incomplete: 8 traces (30%)
3. Wrong Format: 5 traces (19%)

Video Tutorial

📹 Full Workflow Walkthrough

Watch a complete 5-minute tutorial covering the entire EmbedEval workflow:

# View the tutorial video
open embedeval-promo/out/embedeval-tutorial-full-v3.mp4

What's Covered:

✅ Installation verification
✅ First trace collection
✅ Interactive annotation (keyboard shortcuts)
✅ Building failure taxonomy
✅ Running evaluations
✅ Understanding Hamel Husain's principles

Why Watch? The video shows the complete workflow from zero to first evaluation. Seeing it in action makes the documentation come alive and helps you understand:

How the annotation UI works
Keyboard shortcuts in practice
What a failure taxonomy looks like
How evals are configured and run

Development (Contributors)

If you're developing or contributing to EmbedEval, use the embedeval-dev script:

# Clone the repository
git clone https://github.com/Algiras/embedeval.git
cd embedeval

# Check your dev environment
./embedeval-dev --doctor

# Install dependencies
./embedeval-dev --install-deps

# Build TypeScript
./embedeval-dev --build

# Run CLI commands (no global install needed)
./embedeval-dev collect examples/v2/sample-traces.jsonl
./embedeval-dev view test-traces.jsonl
./embedeval-dev annotate test-traces.jsonl --user "dev@local"

# Development utilities
./embedeval-dev --watch     # Watch mode for auto-rebuild
./embedeval-dev --test      # Run test suite
./embedeval-dev --lint      # Run ESLint
./embedeval-dev --types     # TypeScript type check
./embedeval-dev --clean     # Clean build artifacts

Core Commands

Error Analysis (Primary Workflow)

# Import traces from JSONL
embedeval collect ./logs.jsonl --output traces.jsonl

# Interactive annotation (p=pass, f=fail, s=save)
embedeval annotate traces.jsonl --user "pm@company.com"

# Read-only viewer
embedeval view traces.jsonl

# Build failure taxonomy
embedeval taxonomy build --user "pm@company.com"

# Display taxonomy
embedeval taxonomy show

Binary Evaluation

# Add evaluator (interactive wizard)
embedeval eval add

# List evaluators
embedeval eval list

# Run evaluations
embedeval eval run traces.jsonl --config evals.yaml

# Generate report
embedeval eval report --results results.jsonl

Synthetic Data & Export

# Create dimensions template
embedeval generate init

# Generate synthetic traces
embedeval generate create --dimensions dims.yaml --count 50

# Export to Jupyter notebook
embedeval export traces.jsonl --format notebook

# Generate HTML dashboard
embedeval report --traces traces.jsonl --annotations annotations.jsonl

Key Principles

1. Binary Only ✓/✗

# GOOD: Clear, fast decisions
evals:
  - name: is_accurate
    type: llm-judge
    binary: true  # Only PASS or FAIL

# BAD: Never do this
evals:
  - name: quality_score
    type: 1_to_5  # Creates disagreement

2. Error Analysis First 👀

# Spend 60-80% of time here:
embedeval annotate traces.jsonl --user "expert@company.com"

# NOT here (automate only after understanding):
# embedeval eval run traces.jsonl  # (do this AFTER annotation)

3. Cheap Evals First 💰

evals:
  # Run cheap evals first
  - name: has_content
    type: assertion
    check: "response.length > 100"
    priority: cheap
  
  # Expensive evals only for complex cases
  - name: factual_accuracy
    type: llm-judge
    priority: expensive

4. Single Annotator 👤

# One "benevolent dictator" owns quality:
embedeval annotate traces.jsonl --user "product-manager@company.com"

# Not multiple people voting (causes conflict)

Installation Methods

NPM (Recommended)

npm install -g embedeval

NPX (No Install)

npx embedeval collect ./logs.jsonl

From Source

git clone https://github.com/Algiras/embedeval.git
cd embedeval
npm install
npm run build
npm link  # Makes 'embedeval' command available globally

Data Formats

Trace (JSONL)

One JSON object per line:

{"id": "trace-001", "timestamp": "2026-01-30T10:00:00Z", "query": "What's your refund policy?", "response": "We offer full refunds within 30 days...", "metadata": {"provider": "google", "model": "gemini-1.5-flash", "latency": 180}}

Annotation (JSONL)

{"id": "ann-001", "traceId": "trace-001", "annotator": "pm@company.com", "timestamp": "2026-01-30T10:05:00Z", "label": "fail", "failureCategory": "hallucination", "notes": "Made up refund time limit"}

Eval Config (YAML)

evals:
  - id: has_content
    type: assertion
    priority: cheap
    config:
      check: "response.length > 50"
  
  - id: accurate
    type: llm-judge
    priority: expensive
    config:
      model: gemini-1.5-flash
      prompt: "PASS or FAIL: Is this accurate?"
      binary: true

Example Workflows

Weekly Evaluation

# Collect week's traces
embedeval collect ./logs/week-$(date +%Y-%m-%d).jsonl

# Sample 100 for annotation
head -n 100 traces.jsonl > sample.jsonl

# Annotate
embedeval annotate sample.jsonl --user "pm@company.com"

# Build/update taxonomy
embedeval taxonomy update

# Run all evals
embedeval eval run traces.jsonl --config evals.yaml

# Generate report
embedeval report --traces traces.jsonl --annotations annotations.jsonl

Add Eval for Common Failure

# 1. Build taxonomy to see top failures
embedeval taxonomy build

# 2. Add eval for top category (e.g., hallucination)
embedeval eval add
# Interactive wizard asks for type, model, prompt

# 3. Run the new eval
embedeval eval run traces.jsonl --config evals.yaml

Generate Test Data

# 1. Create dimensions file
embedeval generate init
# Edit dimensions.yaml to define test scenarios

# 2. Generate synthetic traces
embedeval generate create -d dimensions.yaml -n 50

# 3. Run your system on synthetic queries
# (Implementation depends on your system)

# 4. Evaluate
embedeval annotate synthetic-traces.jsonl --user "tester@company.com"

MCP Server (AI Agents)

For Claude, Cursor, or other MCP clients:

{
  "mcpServers": {
    "embedeval": {
      "command": "npx",
      "args": ["embedeval", "mcp-server"],
      "env": {
        "GEMINI_API_KEY": "your-api-key"
      }
    }
  }
}

See LLM.md for detailed agent usage guide.

Deployment

Vercel (One-Click)

GitHub Pages

Already configured. Site updates automatically on push to main.

CI/CD

GitHub Actions workflow included. Runs on every PR:

Type checking
Build verification
CLI command tests

Documentation

LLM.md - Complete guide for AI agents and LLMs
GitHub Pages - Visual documentation
Hamel's Eval FAQ - Methodology reference

Features

✅ Interactive Annotation - Terminal UI for fast binary annotation
✅ Failure Taxonomy - Auto-categorize failures with axial coding
✅ Binary Evaluation - Assertions, regex, code, LLM-as-judge
✅ Synthetic Data - Dimension-based test generation
✅ Jupyter Export - Statistical analysis notebooks
✅ HTML Reports - Shareable dashboards
✅ JSONL Storage - Simple, grep-friendly, no databases
✅ Zero Infrastructure - No Redis, no queues, no setup

Why v2?

v1 had 88 files with complex A/B testing, genetic algorithms, and BullMQ queues.

v2 has ~20 files with a simple philosophy: look at your traces first.

Before: Infrastructure-heavy, hard to understand
After: Simple CLI, clear workflow, Hamel Husain principles

License

MIT

Built with ❤️ following Hamel Husain's principles. The goal is understanding failures, not perfect metrics. Spend time looking at traces! 👀

Name		Name	Last commit message	Last commit date
Latest commit History 107 Commits
.github/workflows		.github/workflows
docs		docs
examples		examples
scripts		scripts
skills		skills
src		src
tests		tests
.bunfig.toml		.bunfig.toml
.env.example		.env.example
.eslintrc.json		.eslintrc.json
.gitignore		.gitignore
.npmignore		.npmignore
AGENTS.md		AGENTS.md
GETTING_STARTED.md		GETTING_STARTED.md
LLM.md		LLM.md
README.md		README.md
TROUBLESHOOTING.md		TROUBLESHOOTING.md
embedeval-dev		embedeval-dev
install.sh		install.sh
jest.config.js		jest.config.js
llms.txt		llms.txt
package.json		package.json
tsconfig.json		tsconfig.json
vercel.json		vercel.json

Folders and files

Latest commit

History

Repository files navigation