Building Production RAG Over Complex Documents

Large Language Models (LLMs) are revolutionizing how users search for, interact with, and generate new content. Some recent stacks and toolkits around Retrieval-Augmented Generation (RAG) have emerged, enabling users to build applications such as chatbots using LLMs on their private data. However, while setting up naive RAG is straightforward, building production RAG is very challenging, especially as users scale to larger and more complex data sources. A classic example is a large number of PDFs with embedded tables.

RAG is only as good as your data, and developers must carefully consider how to parse, ingest, and retrieve their data to successfully build RAG over complex documents. This session provides an in-depth exploration of this entire process; you will get an overview of the process around building a RAG pipeline that can handle messy, complicated PDF documents. This includes implementing a parsing strategy for parsing a complex document with embedded objects. This consists of an indexing strategy to process these documents beyond simple chunking techniques. We will then explore various advanced retrieval algorithms to handle questions about the tabular and unstructured data and discuss their use cases and tradeoffs.