AutoGluon-RAG is a framework designed to streamline the development of RAG (Retrieval-Augmented Generation) pipelines. RAG has emerged as a crucial approach for tailoring large language models (LLMs) to address domain-specific queries. However, constructing RAG pipelines traditionally involves navigating through a complex array of modules and functionalities, including retrievers, generators, vector database construction, fast semantic search, and handling long-context inputs, among others.
AutoGluon-RAG allows users to create customized RAG pipelines seamlessly, eliminating the need to delve into any technical complexities. Following the AutoML (Automated Machine Learning) philosophy of simplifying model development with minimal code, as exemplified by AutoGluon; AutoGluon-RAG enables users to create a RAG pipeline with just a few lines of code. The framework provides a user-friendly interface, and abstracts away the underlying modules, allowing users to focus on their domain-specific requirements and leveraging the power of RAG pipelines without the need for extensive technical expertise.
In line with the AutoGluon team's commitment to meeting user requirements and expanding its user base, the team aims to develop a new feature that simplifies the creation and deployment of end-to-end RAG (Retrieval-Augmented Generation) pipelines. Given a set of user-provided data or documents, this feature will enable users to develop and deploy a RAG pipeline with minimal coding effort, following the AutoML (Automated Machine Learning) philosophy of three-line solutions.
To use this framework, you must first install AutoGluon RAG:
git clone https://github.com/autogluon/autogluon-rag
cd autogluon-rag
# Create a Virtual Environment (using Python, or conda if you prefer)
python3 -m virtualenv venv
source venv/bin/activate
#Install the package
pip install -e .
You can now use the package in two ways.
AutoGluon-RAG
usage: agrag [-h] --config_file
AutoGluon-RAG - Retrieval-Augmented Generation Pipeline
options:
-h, --help show this help message and exit
--config_file Path to the configuration file
from agrag.agrag import AutoGluonRAG
def ag_rag():
agrag = AutoGluonRAG(
preset_quality="medium_quality", # or path to config file
web_urls=["https://auto.gluon.ai/stable/index.html"],
base_urls=["https://auto.gluon.ai/stable/"],
parse_urls_recursive=True,
data_dir="s3://autogluon-rag-github-dev/autogluon_docs/"
)
agrag.initialize_rag_pipeline()
agrag.generate_response("What is AutoGluon?")
if __name__ == "__main__":
ag_rag()
These are the optional parameters that can be passed into the AutoGluonRAG class:
config_file : str
Path to a configuration file that will be used to set specific parameters in the RAG pipeline.
preset_quality : str
If you do not wish to use your own configuration file, you can use a preset configuration file which contains pre-defined arguments.
You must provide the preset quality setting ("good_quality", "medium_quality", or, "best_quality"). Note that if both config_file and preset_quality are provided, config_file will be prioritized.
model_ids : dict
Dictionary of model IDs to use for specific modules.
Example: {"generator_model_id": "mistral.mistral-7b-instruct-v0:2", "retriever_model_id": "BAAI/bge-large-en", "reranker_model_id": "nv_embed"}
data_dir : str
The directory containing the data files that will be used for the RAG pipeline. If this value is not provided when initializing the object, it must be provided in the config file. If both are provided, the value in the class instantiation will be prioritized.
web_urls : List[str]
List of website URLs to be ingested and processed. Each URL will processed recursively based on the base URL to include the content of URLs that exist within this URL.
If this value is not provided when initializing the object, it must be provided in the config file. If both are provided, the value in the class instantiation will be prioritized.
base_urls : List[str]
List of optional base URLs to check for links recursively. The base URL controls which URLs will be processed during recursion. The base_url does not need to be the same as the web_url. For example. the web_url can be "https://auto.gluon.ai/stable/index.html", and the base_urls will be "https://auto.gluon.ai/stable/".
If this value is not provided when initializing the object, it must be provided in the config file. If both are provided, the value in the class instantiation will be prioritized.
login_info: dict
A dictionary containing login credentials for each URL. Required if the target URL requires authentication.
Must be structured as {target_url: {"login_url": <login_url>, "credentials": {"username": "your_username", "password": "your_password"}}}
The target_url is a url that is present in the list of web_urls
parse_urls_recursive: bool
Whether to parse each URL in the provided recursively. Setting this to True means that the child links present in each parent webpage will also be processed.
pipeline_batch_size: int
Batch size to use for pre-processing stage (Data Processing, Embedding, Vector DB Module). This represents the number of files in each batch.
The default value is 20.
Note: You may provide both data_dir and web_urls.
The configuration file contains the specific parameters to use for each module in the RAG pipeline. For an example of a config file, please refer to example_config.yaml in src/agrag/configs/. For specific details about the parameters in each individual module, refer to the README files in each module in src/agrag/modules/.
There is also a shared section in the config file for parameters that do not refer to a specific module. Currently, the parameters in shared are:
pipeline_batch_size: Optional batch size to use for pre-processing stage (Data Processing, Embedding, Vector DB Module). This represents the number of files in each batch. The default value is 20.
For more information about the evaluation module, refer to the code in src/agrag/evaluation and the instructions here.