Overview

RAGOpt provides an end-to-end framework for optimizing RAG (Retrieval-Augmented Generation) pipelines using Multi-Objective Bayesian Optimization. The framework automatically tunes hyperparameters to find Pareto-optimal configurations that balance multiple objectives like cost, latency, and quality metrics.

Framework Architecture

The optimization workflow consists of five key components working together:

Dataset Generation - Create synthetic question-answer pairs from your documents
Search Space - Define the hyperparameter space to explore
BO Input Encoder - Encode the RAG hyperparameter from and to pytorch tensors
Sampler - Sampling choices from search space using SOBOL sampler by default
RAG Manager - Orchestrate component loading and configuration sampling
Evaluation - Measure performance across multiple metrics
Optimization - Find optimal configurations using Bayesian Optimization

Quick Start

Here’s a minimal example to get started:

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

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

# Initialize optimizer with configuration
optimizer = Optimizer(
    train_dataset=dataset,
    config_path="./rag_config.yaml"
)

# Run optimization
best_configs = optimizer.optimize(n_trials=3)

Getting started

RAG Configurations

Optimization Workflow

Evaluation & Metrics

Framework Architecture

Quick Start

​Framework Architecture

​Quick Start

Framework Architecture

Quick Start