Skip to main content

Alt text describing the image


RAGOpt is a Python optimization framework that eliminates manual hyperparameter guesswork in your RAG pipelines. It uses Bayesian optimization to systematically tune over 20+ hyperparameters, from chunk size and overlap to model selection and embedding strategies automatically discovering the best configurations for your specific use case

​
Why RAGOpt?

Optimizing Retrieval-Augmented Generation (RAG) pipelines is complex and highly dependent on configuration choices that impact quality, latency, and cost. Unlike AutoRAG, which relies on grid search, or Ragas, which enforces rigid evaluation frameworks, RAGOpt is partially opinionated: it provides smart defaults while staying flexible with any LangChain-compatible model or provider. RAGOpt helps answer a question like:
What’s the best RAG configuration to minimize cost and latency while maintaining accuracy and safety ?
For example, here’s you can get best RAG configurations in a few iterations using RAGOpt optimizer: 
from rag_opt.dataset import TrainDataset
from rag_opt.optimizer import Optimizer

# First you have to run generate_questions_.py
# to get a list of QAs to be used in the evaluation process

# Load the training dataset (questions/answers)
train_dataset = TrainDataset.from_json("rag_dataset.json")

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

# Run optimization
# Increase n_trials for better results
best_config = optimizer.optimize(n_trials=2, best_one=True,plot_hypervolume=True)
best_config.to_json()


# 1- define here the RAG Hyperparameters u would like to optimize
chunk_size:
  bounds: [512, 1024]
  dtype: int

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

vector_store:
---
# More Parameters

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

# 1- Generate list of questions / answers to be used in RAG evaluation

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

# Generates 3 training questions
# you can also path a file like dataset_path=myfile.pdf
data_gen = DatasetGenerator(llm,dataset_path="./data")
dataset = data_gen.generate(3)
dataset.to_json("./questions.json")


{
  "items": [
    {
      "question": "What is the capital city of Japan?",
      "answer": "The capital city of Japan is Tokyo.",
      "contexts": [
        "Tokyo is the capital of Japan, though some still picture Kyoto when they hear 'old Japan'.",
        "Neon lights and ancient temples coexist where decisions shape the nation.",
        "It’s a city where tradition bows to innovation every single day."
      ],
      "metadata": {}
    },
    {
      "question": "Who won the Nobel Peace Prize in 2025?",
      "answer": "The Nobel Peace Prize in 2025 was awarded to María Corina Machado.",
      "contexts": [
        "María Corina Machado received the Nobel Peace Prize in 2025, proving persistence can echo louder than politics.",
        "Donald Trump didn’t get the Nobel Prize; however, the announcement stirred global chatter.",
        "Sometimes peace arrives wearing the voice of defiance and hope."
      ],
      "metadata": {}
    }
  ]
}
This will give you the optimal RAG configuration for your dataset. 
{
  "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,
}

​
Get Started

Quick Start

Get started with RAGOpt in under 5 minutes.

Main Concepts

Get familiar with the underlying architecture and key design of RAGOpt

Configuration Guide

Learn about the Hyperparameters RAG Configuration file in details

Optimization Workflows

Integrate RAGOpt with your custom RAG pipeline.