RAGs to riches

ChatGPT changed my life – said a friend of mine who runs a successful business in 2023. Come 2024, he said I am not able to get the most out of generic LLM which everyone has access to, what do I do different? Well, my answer was RAGs will get you Riches. He chuckled but I think he didn’t understand the pun.

AI is advancing rapidly and newer frameworks are coming up everyday. Lot of people these days are talking about Retrieval-Augmented Generation (RAGs). With the rise of LLMs thanks to ChatGPT (openAI), usage has gone up significantly but the issue with most LLMs as we know is hallucinations and the fact that they can provide irrelevant or even factually responses sometimes. RAGs is meant to solve that and make your LLM even better.

RAGs work in a sophisticated manner to integrate retrieval mechanisms with generative models to enhance the quality and relevance of generated outputs. It has two main components with a Retrieval Phase in which the model searches a large corpus of documents or data sources to find relevant information related to a given query. Next is the Generation Phase which basically utilizes the retrieved information to generate a response that is informed by the most relevant and up-to-date data available.

We can also understand in a step by step manner (slightly technical):

The Process

1. Query Input: User provides a query or commonly known as a prompt.
2. Document Retrieval: The retrieval component searches a database or corpus to find text that are most relevant to the query.
3. Document Encoding: The retrieved documents are encoded into a format that the model can process.
4. Response Generation: The generative model uses the encoded information to generate a comprehensive and contextually relevant response.

It is important to note that Retriever is implemented using models like Dense Passage Retrieval (DPR) and Generator is typically a Transformer-based model, such as Generative Pre Trained Transformer (GPT) or Bidirectional Encoder Representations from Transformers (BERT) which generates the final response using the retrieved information. The reason RAGs are becoming popular is that they excel in understanding and utilizing context. RAG frameworks are also scalable and can handle vast amounts of data in an efficient manner. They are also flexible, capable of integrating various types of data sources, including structured databases and unstructured texts.

As AI and machine learning technologies continue to evolve, RAG models will become even more sophisticated. They will be able to continuously learn and adapt to new information, ensuring that their responses remain relevant and accurate over time.

There are lot of applications of RAGs in daily life such as:

Customer Support

RAG models can literally personalize customer support by providing accurate, up-to-date responses to user queries. Imagine calling a call centre and they know context about you and talk to you in a personalized manner.

Education

RAG models can enhance quality and personalization of educational tools by providing students with accurate and comprehensive answers to their questions. Educators can also create dynamic and informed lesson plans.

Content Creation

RAG can aid content creators by generating contextually rich and accurate content, and produce high-quality articles, reports, and creative writing.

They will have the capacity to pick up new information quickly and adjust to it, so over time their answers will continue to be accurate and pertinent. It is probable that RAG will be combined with other technologies, such neural-symbolic systems and reinforcement learning, to provide even more potent and adaptable AI solutions. This will improve AI systems’ capacity to comprehend the outside world and engage in meaningful interactions with it.

It will be crucial to address ethical issues with RAG models as they become more widely used, such as data privacy, bias in retrieved information, and misuse potential. Ensuring that RAG technologies help society as a whole will depend on the development of frameworks for ethical AI use.

Have you explored RAGs yet?