ranx-k 0.0.10

Korean-optimized RAG evaluation toolkit based on ranx with Kiwi tokenizer and Korean language support

ranx-k: Korean-optimized ranx IR Evaluation Toolkit 🇰🇷

English | 한국어

ranx-k is a Korean-optimized Information Retrieval (IR) evaluation toolkit that extends the ranx library with Kiwi tokenizer and Korean embeddings. It provides accurate evaluation for RAG (Retrieval-Augmented Generation) systems.

🚀 Key Features

• Korean-optimized: Accurate tokenization using Kiwi morphological analyzer
• ranx-based: Supports proven IR evaluation metrics (Hit@K, NDCG@K, MRR, MAP@K, etc.)
• LangChain compatible: Supports LangChain retriever interface standards
• Multiple evaluation methods: ROUGE, embedding similarity, semantic similarity-based evaluation
• Graded relevance support: NEW in v0.0.9 - Use similarity scores as relevance grades instead of binary 1/0
• Configurable ROUGE types: Choose between ROUGE-1, ROUGE-2, and ROUGE-L
• Strict threshold enforcement: Documents below similarity threshold are correctly treated as retrieval failures
• Practical design: Supports step-by-step evaluation from prototype to production
• High performance: 30-80% improvement in Korean evaluation accuracy over existing methods
• Bilingual output: English-Korean output support for international accessibility

📦 Installation

pip install ranx-k

Or install development version:

pip install "ranx-k[dev]"

🔗 Retriever Compatibility

ranx-k supports LangChain retriever interface:

# Retriever must implement invoke() method
class YourRetriever:
    def invoke(self, query: str) -> List[Document]:
        # Return list of Document objects (requires page_content attribute)
        pass

# LangChain Document usage example
from langchain.schema import Document
doc = Document(page_content="Text content")

Note: LangChain is distributed under the MIT License. See documentation for details.

🔧 Quick Start

Basic Usage

from ranx_k.evaluation import simple_kiwi_rouge_evaluation

# Simple Kiwi ROUGE evaluation
results = simple_kiwi_rouge_evaluation(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5
)

print(f"ROUGE-1: {results['kiwi_rouge1@5']:.3f}")
print(f"ROUGE-2: {results['kiwi_rouge2@5']:.3f}")
print(f"ROUGE-L: {results['kiwi_rougeL@5']:.3f}")

Enhanced Evaluation (Rouge Score + Kiwi)

from ranx_k.evaluation import rouge_kiwi_enhanced_evaluation

# Proven rouge_score library + Kiwi tokenizer
results = rouge_kiwi_enhanced_evaluation(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5,
    tokenize_method='morphs',  # 'morphs' or 'nouns'
    use_stopwords=True
)

Semantic Similarity-based ranx Evaluation

from ranx_k.evaluation import evaluate_with_ranx_similarity

# Reference-based evaluation (recommended for accurate recall)
results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5,
    method='embedding',
    similarity_threshold=0.6,
    use_graded_relevance=False,        # NEW: Binary relevance (default)
    evaluation_mode='reference_based'  # Evaluates against all reference docs
)

print(f"Hit@5: {results['hit_rate@5']:.3f}")
print(f"NDCG@5: {results['ndcg@5']:.3f}")
print(f"MRR: {results['mrr']:.3f}")
print(f"MAP@5: {results['map@5']:.3f}")

# NEW: Graded relevance - uses similarity scores as relevance grades
results_graded = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5,
    method='embedding',
    similarity_threshold=0.6,
    use_graded_relevance=True,         # Use similarity scores as grades
    evaluation_mode='reference_based'
)

print(f"Graded NDCG@5: {results_graded['ndcg@5']:.3f}")

Using Different Embedding Models

# OpenAI embedding model (requires API key)
results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5,
    method='openai',
    similarity_threshold=0.7,
    embedding_model="text-embedding-3-small"
)

# Latest BGE-M3 model (excellent for Korean)
results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5,
    method='embedding',
    similarity_threshold=0.6,
    embedding_model="BAAI/bge-m3"
)

# Korean-specialized Kiwi ROUGE method with configurable ROUGE types (NEW in v0.0.9)
results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5,
    method='kiwi_rouge',
    similarity_threshold=0.3,  # Lower threshold recommended for Kiwi ROUGE
    rouge_type='rougeL',      # NEW: Choose 'rouge1', 'rouge2', or 'rougeL'
    tokenize_method='morphs', # NEW: Choose 'morphs' or 'nouns'  
    use_stopwords=True        # NEW: Configure stopword filtering
)

Comprehensive Evaluation

from ranx_k.evaluation import comprehensive_evaluation_comparison

# Compare all evaluation methods
comparison = comprehensive_evaluation_comparison(
    retriever=your_retriever,
    questions=your_questions,
    reference_contexts=your_reference_contexts,
    k=5
)

📊 Evaluation Methods

1. Kiwi ROUGE Evaluation

• Advantages: Fast speed, intuitive interpretation
• Use case: Prototyping, quick feedback

2. Enhanced ROUGE (Rouge Score + Kiwi)

• Advantages: Proven library, stability
• Use case: Production environment, reliability-critical evaluation

3. Semantic Similarity-based ranx

• Advantages: Traditional IR metrics, semantic similarity
• Use case: Research, benchmarking, detailed analysis

🎯 Performance Improvement Examples

# Existing method (English tokenizer)
basic_rouge1 = 0.234

# ranx-k (Kiwi tokenizer)
ranxk_rouge1 = 0.421  # +79.9% improvement!

📊 Recommended Embedding Models

Model	Use Case	Threshold	Features
paraphrase-multilingual-MiniLM-L12-v2	Default	0.6	Fast, lightweight
text-embedding-3-small (OpenAI)	Accuracy	0.7	High accuracy, cost-effective
BAAI/bge-m3	Korean	0.6	Latest, excellent multilingual
text-embedding-3-large (OpenAI)	Premium	0.8	Highest performance

📈 Score Interpretation Guide

Score Range	Assessment	Recommended Action
0.7+	🟢 Excellent	Maintain current settings
0.5~0.7	🟡 Good	Consider fine-tuning
0.3~0.5	🟠 Average	Improvement needed
0.3-	🔴 Poor	Major revision required

🔍 Advanced Usage

Graded vs Binary Relevance Comparison

# Compare binary and graded relevance
binary_results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=questions,
    reference_contexts=references,
    method='embedding',
    similarity_threshold=0.6,
    use_graded_relevance=False  # Binary: 1.0 for all relevant docs
)

graded_results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=questions,
    reference_contexts=references,
    method='embedding',
    similarity_threshold=0.6,
    use_graded_relevance=True   # Graded: similarity scores as relevance grades
)

print(f"Binary NDCG@5: {binary_results['ndcg@5']:.3f}")
print(f"Graded NDCG@5: {graded_results['ndcg@5']:.3f}")

Custom Embedding Models

# Use custom embedding model
results = evaluate_with_ranx_similarity(
    retriever=your_retriever,
    questions=questions,
    reference_contexts=references,
    method='embedding',
    embedding_model="your-custom-model-name",
    similarity_threshold=0.6,
    use_graded_relevance=True
)

Configurable ROUGE Types

# Compare different ROUGE metrics with graded relevance
for rouge_type in ['rouge1', 'rouge2', 'rougeL']:
    results = evaluate_with_ranx_similarity(
        retriever=your_retriever,
        questions=questions,
        reference_contexts=references,
        method='kiwi_rouge',
        rouge_type=rouge_type,
        tokenize_method='morphs',
        similarity_threshold=0.3,
        use_graded_relevance=True  # Use ROUGE scores as relevance grades
    )
    print(f"{rouge_type.upper()}: Hit@5 = {results['hit_rate@5']:.3f}")

Threshold Sensitivity Analysis

# Analyze how different thresholds affect graded vs binary relevance
thresholds = [0.3, 0.5, 0.7]
for threshold in thresholds:
    binary = evaluate_with_ranx_similarity(
        retriever=your_retriever,
        questions=questions,
        reference_contexts=references,
        similarity_threshold=threshold,
        use_graded_relevance=False
    )
    graded = evaluate_with_ranx_similarity(
        retriever=your_retriever,
        questions=questions,
        reference_contexts=references,
        similarity_threshold=threshold,
        use_graded_relevance=True
    )
    print(f"Threshold {threshold}: Binary={binary['hit_rate@5']:.3f}, Graded={graded['hit_rate@5']:.3f}")

📚 Examples

• Basic Tokenizer Example
• BGE-M3 Evaluation Example
• Embedding Models Comparison
• Comprehensive Comparison

📖 Documentation

• Installation Guide
• Quick Start Guide
• API Reference
• Korean Documentation

🤝 Contributing

Contributions are welcome! Please read our Contributing Guide for details.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Built on top of ranx by Elias Bassani
• Korean morphological analysis powered by Kiwi
• Embedding support via sentence-transformers

📞 Support

• 🐛 Issue Tracker
• 📧 Email: [e-mail]
• 📖 Documentation

---

ranx-k - Empowering Korean RAG evaluation with precision and ease!