Essentials

To get to know FedRAG a bit better and understand its purpose, we provide the answers to the following four essential questions.

Four Essential Questions

What is RAG?

Retrieval-Augmented Generation (RAG) is a widely used technique that addresses a main drawback of Large Language Models (LLM), which is that they're trained on historical corpora and thus answering user queries that heavily rely on recent data are not really possible. Further, using the parametric knowledge of LLMs alone has yielded subpar performance on knowledge-intensive benchmarks.

RAG provides access to relevant (and potentially more recent) non-parametric knowledge (i.e. data) that are stored in Knowledge Stores to the LLM so that it can use it in order to more accurately respond to user queries.

What is Federated Learning?

Federated Learning (FL) is a technique for building machine learning (as well as deep learning) models when the data is decentralized. Rather than first centralizing the datasets to a central location, which may not be possible due to strict data residency regulations or may be uneconomical due to the significant monetary costs in moving massive datasets, FL enables collaborative model building by facilitating the sharing of the model weights between the data providers.

Why Federated Fine-Tuning of RAG?

Fine-tuning is a technique that is used to enhance the performance of LLMs by speciliazing its general capabilities towards a specific domain. It has also been shown that fine-tuning the model components of RAG systems, namely the generator and retriever, on domain-specific datasets can lead to its overall improved performance.

Accessing fine-tuning datasets may be challenging. And, in situations where the data is dispersed across several nodes, and centralizing is either not possible or uneconomical, the fine-tuning of these RAG systems can be made possible through FL.

Who is FedRAG for?

FedRAG is for the model builders, data scientists, and researchers who wish to fine-tune their RAG systems on their own datasets.

Note — FedRAG also enables centralized RAG fine-tuning

While the main premise for FedRAG is FL, note that centralized fine-tuning is also supported (as our readers shall see shortly in the upcoming Quick Start example).