Beacon

A benchmarking toolkit that evaluates and compares RAG chunking strategies against your actual queries to find the optimal configuration.

Why Beacon?

RAG quality is 80% retrieval quality, and retrieval quality depends heavily on chunking. Yet most teams:

• Use default chunk sizes (512 tokens) without evaluation
• Don't know if semantic chunking would improve their specific use case
• Spend weeks manually testing different configurations
• Have no metrics to compare strategies objectively

Beacon automates this with reproducible benchmarks.

Features

• Multi-Strategy Evaluation - Compare 5+ chunking strategies in one run
• Standard IR Metrics - MRR, Recall@K, NDCG, Precision, MAP
• Bayesian Auto-Tuning - Find optimal chunk size and overlap automatically
• HTML Reports - Beautiful, shareable benchmark reports
• Local-First - Run entirely offline with local embeddings
• Framework Agnostic - Works with any RAG implementation

Installation

pip install beacon-eval

For auto-tuning support:

pip install beacon-eval[optuna]

Quick Start

1. Create a configuration file

beacon init

This creates a beacon.yaml configuration file.

2. Prepare your queries

Create a queries.jsonl file with your test queries:

{"query": "What is the refund policy?", "relevant_doc_ids": ["policy.pdf"]}
{"query": "How do I reset my password?", "relevant_doc_ids": ["faq.pdf"]}

3. Run the benchmark

beacon run beacon.yaml

Configuration

name: my_benchmark
documents:
  - ./docs/*.pdf
  - ./docs/*.txt
queries: ./queries.jsonl
embedding_model: all-MiniLM-L6-v2
top_k: 10

strategies:
  - name: fixed_256
    type: fixed_size
    chunk_size: 256
    overlap: 25

  - name: fixed_512
    type: fixed_size
    chunk_size: 512
    overlap: 50

  - name: sentence_based
    type: sentence
    chunk_size: 512

  - name: recursive
    type: recursive
    chunk_size: 512
    overlap: 50

output_dir: ./results
generate_html_report: true
export_csv: true

Chunking Strategies

Strategy	Description
fixed_size	Split by fixed token/character count
sentence	Split by sentence boundaries
paragraph	Split by paragraph boundaries
semantic	Split by semantic similarity
recursive	Recursively split with multiple separators

Python API

from beacon import BenchmarkConfig, ChunkingStrategy
from beacon.runner import run_benchmark
from beacon.parsers import load_documents, load_queries

# Load your data
documents = load_documents([Path("./docs")])
queries = load_queries(Path("./queries.jsonl"))

# Define strategies
strategies = [
    ChunkingStrategy(name="small", strategy_type="fixed_size", chunk_size=256),
    ChunkingStrategy(name="medium", strategy_type="fixed_size", chunk_size=512),
    ChunkingStrategy(name="large", strategy_type="fixed_size", chunk_size=1024),
]

# Run benchmark
config = BenchmarkConfig(
    name="my_benchmark",
    documents=[Path("./docs")],
    queries_file=Path("./queries.jsonl"),
    strategies=strategies,
)

result = run_benchmark(config)
print(f"Best strategy: {result.best_strategy}")

Auto-Tuning

Find optimal chunking parameters automatically:

from beacon.tuner import auto_tune
from beacon.parsers import load_documents, load_queries
from beacon.models import ChunkingStrategyType

documents = load_documents([Path("./docs")])
queries = load_queries(Path("./queries.jsonl"))

result = auto_tune(
    documents=documents,
    queries=queries,
    strategy_type=ChunkingStrategyType.FIXED_SIZE,
    metric="mrr",
    n_trials=50,
    chunk_size_range=(100, 2000),
)

print(f"Optimal chunk size: {result['chunk_size']}")
print(f"Optimal overlap: {result['chunk_overlap']}")
print(f"Best MRR: {result['best_mrr']:.4f}")

Metrics

Metric	Description	Good Score
MRR	Mean Reciprocal Rank	> 0.7
Recall@K	% of relevant docs in top K	> 0.8
NDCG@K	Normalized DCG	> 0.75
Precision@K	Precision at K	> 0.6
MAP	Mean Average Precision	> 0.6

CLI Commands

# Initialize sample configuration
beacon init

# Run benchmark
beacon run config.yaml

# List available strategies
beacon strategies

# Compare multiple benchmark results
beacon compare results1.json results2.json

Output

Beacon generates:

• results.json - Full results in JSON format
• results.csv - Comparison table in CSV
• report.html - Interactive HTML report with charts

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

Security

If you discover a security vulnerability, please see SECURITY.md for reporting guidelines.

License

MIT License - see LICENSE for details.