Large language models (LLMs) like GPT-4 boast reasoning capabilities that pave the way for Enterprise AGI, systems that can perform at least 50% of valuable human work in a business. A brand new enterprise technology stack is emerging to support Enterprise AGI, and today, I'd like to introduce you to an early player in this stack called LlamaIndex.
At its core, LlamaIndex serves as an interface between your enterprise data and large language models. Its aim is to simplify the data layers and processing between your users and your source data. To better grasp its value proposition, let's consider a few examples.
Imagine you want to create a chat interface between your users and a large text document. GPT-4 doesn't inherently know about this document, so you have to feed it chunks of text as prompts, enabling GPT-4 to reason across and answer users' questions. The most common approach to selecting the most relevant chunks involves creating a set of document embeddings. This technique is a typical first use case for GPT-4, and you don't need a complex interface like LlamaIndex to achieve it.
However, things change when you want to give users the option to also summarize a document. In this situation, embeddings aren't the best solution. The scenario gets even more complicated when you want to query multiple document types or data types (such as structured, relational databases, graph databases, PDFs, or images). In an ideal world, you'd have an interface layer between whatever the user may ask and these multiple document types, allowing GPT-4 to reason across them. That’s what LlamaIndex aims to become, and it is only feasible because GPT-4 can handle messy data, extract insights from it, and reason across it.
It's hard to overstate the profound shift this paradigm represents compared to current approaches for enterprise data processing and analytics. The bedrock of current approaches is the need to preprocess and organize data in a structure that enables analytics. Emerging technology stacks that support Enterprise AGI, like LlamaIndex, disrupt this traditional design and create opportunities to significantly simplify the data stack.
Does this mean we're heading towards a perfect world, free of the data stovepipe headaches we've been grappling with for decades? Not quite. This new approach comes with its own set of challenges, such as latency and error propagation between large language models.
Nevertheless, this new tech stack creates the opportunity for delivering business value years sooner than traditional approaches, and most analytics leaders will choose them because the business ROI is much higher.