rag-opt 0.11.0

RAG workflow Optimizer based on Bayesian Optimization

RAGOpt eliminates manual hyperparameter tuning in your RAG pipelines with Bayesian optimization 🚀

---

• Documentation: https://ragopt.aboneda.com
• Colab Notebook: https://colab.research.google.com/drive/1hrfAHCfm3x0Ov-amCEHpptMiyoqC-McE

---

RAGOpt is a Python framework to optimize Retrieval-Augmented Generation (RAG) pipelines. It eliminates manual hyperparameter tuning using Bayesian optimization, automatically finding the best configuration for your dataset and use case.

Key Features

• Optimizes 20+ RAG hyperparameters including chunk size, overlap, embedding strategies, and LLM selection.
• Flexible with any LangChain-compatible model or provider.
• Partially Opinionated - Smart defaults with full flexibility for customization
• Generates Pareto-optimal configurations for your specific data.
• Comprehensive Metrics - Quality (precision, recall, faithfulness), performance (latency, cost), and safety (toxicity, bias)

---

Installation

pip install rag-opt

🔥 Quick Start

1. Generate Training Questions

from langchain.chat_models import init_chat_model
from rag_opt.rag import DatasetGenerator

# Initialize LLM
llm = init_chat_model(
    model="gpt-3.5-turbo",
    model_provider="openai",
    api_key="sk-***"
)

# Generate Q&A pairs from your documents
data_gen = DatasetGenerator(llm, dataset_path="./data")
dataset = data_gen.generate(10)
dataset.to_json("./rag_dataset.json")

2. Define Your Search Space

Create rag_config.yaml:

chunk_size:
  bounds: [512, 1024]
  dtype: int

max_tokens:
  bounds: [256, 512]
  dtype: int

chunk_overlap:
  bounds: [0, 200]
  dtype: int

temperature:
  bounds: [0.0, 1.0]
  dtype: float

search_type:
  choices: ["similarity", "mmr", "hybrid"]

vector_store:
  choices:
    faiss: {}
    pinecone:
      api_key: "YOUR_API_KEY"
      index_name: "your-index"

embedding:
  choices:
    openai:
      api_key: "YOUR_API_KEY"
      models:
        - "text-embedding-3-large"
        - "text-embedding-ada-002"
    huggingface:
      models:
        - "all-MiniLM-L6-v2"

llm:
  choices:
    openai:
      api_key: "YOUR_API_KEY"
      models:
        - "gpt-4o"
        - "gpt-3.5-turbo"

k:
  bounds: [1, 10]
  dtype: int

use_reranker: false

3. Run Optimization

from rag_opt.dataset import TrainDataset
from rag_opt.optimizer import Optimizer

# Load dataset
train_dataset = TrainDataset.from_json("rag_dataset.json")

# Initialize optimizer
optimizer = Optimizer(
    train_dataset=train_dataset,
    config_path="rag_config.yaml",
    verbose=True
)

# Find optimal configuration
best_config = optimizer.optimize(n_trials=3, best_one=True)
best_config.to_json()

4. Get Your Optimized Config

Output example:

{
"chunk_size": 500
"max_tokens": 100
"chunk_overlap": 200
"search_type": "hybrid"
"k": "1"
"temperature": 1.0
"embedding":
  "provider": "openai"
  "model": "text-embedding-3-large"
"llm":
  "provider": "openai"
  "model": "gpt-4o"
"vector_store":
  "provider": "faiss"
"use_reranker": "true"
}

How It Works

1. Dataset Generation - Create synthetic Q&A pairs from your documents
2. Search Space Definition - Configure which parameters to optimize
3. Bayesian Optimization - Intelligently sample and evaluate configurations
4. Multi-Metric Evaluation - Assess quality, performance, and safety
5. Pareto-Optimal Results - Get the best configurations for your priorities

RAGOpt vs Alternatives

• AutoRAG: Uses Bayesian optimization instead of grid search like AutoRAG
• Ragas: Flexible evaluation framework, not rigid—bring your own metrics
• Manual Tuning: Systematic, data-driven approach saves time and improves results

License

This project is licensed under the terms of the MIT license.