Wednesday Oct 2
16:30 –
Room 2

Improving LLM applications through Advanced RAG Techniques

With the proliferation of Large Language Models (LLMs) such as ChatGPT, companies and applications are increasingly leveraging AI capabilities. However, a common challenge arises: LLMs lack specific or sensitive knowledge that they were not explicitly trained on. Retrieval Augmented Generation (RAG) emerges as a powerful solution to bridge this gap.

RAG addresses this limitation by incorporating context from external knowledge bases into LLM-generated responses, enabling more informed and context-aware responses. By retrieving relevant information during the generation process, RAG significantly enhances the quality of LLM-generated content. Notably, RAG has become the dominant architecture for LLM-based systems, powering applications with AI capabilities.

In this presentation, we start by explaining the basic architecture and delving into the shortcomings of the baseline RAG technique. Then we delve into advanced RAG techniques, aiming to equip developers with a comprehensive toolbox for building high-performing RAG applications. We explore the following:

  • Modular RAG: A flexible approach that combines multiple techniques in varying orders. By assembling modules, developers can tailor RAG to specific use cases.
  • Query rewriting, expansion and routing: Leveraging the power of LLMs to refine their own queries. This technique optimizes the retrieval process, leading to more accurate context incorporation.
  • Reranking: Enhancing the results of the retrieval phase by intelligently reordering retrieved information. Reranking ensures that the most relevant context influences the generated output.
  • Fine-tuning: Customizing LLMs specifically for RAG tasks. Fine-tuning aligns the model with the nuances of retrieval and generation.
  • GraphRAG: Harnessing knowledge graphs to enrich context retrieval. Graph-based approaches offer a structured way to connect LLMs with external information.

We delve into the benefits of each technique, explore use cases, and provide theoretical foundations. Additionally, practical examples will illustrate how developers can effectively implement these methods.

large language models (LLM)