RAG: What it is and Why it's Important

RAG is an acronym for Retrieval-Augmented Generation. I know that doesn't help much, so let me dig in a bit more.

Large Language Models like Anthropic's Claude have access to a ton of knowledge out of the box. You need it to write SuiteScript for NetSuite? No problem. How about an HTTP client for a RESTful interface into NetSuite? No problem there, either. But what happens when you need your LLM to answer questions about an internal business process? That's where RAG comes into play. Often you hear about "training" an AI agent - how do you let the LLM know and understand what internal data is available?

If you have or are developing an internal knowledge base (KB), you can give the LLM the ability to search the KB to retrieve relevant information. You often see this in AI agents that work in a chat interface. Perhaps you have an AI agent for HR. It needs to know about HR policies so that it has an appropriate answer to users' questions. For example, an employee wants to know when they will become fully vested in their 401(k) employer match.

There is no way an AI agent is going to know the answer to that question unless you have a KB of some sort that provides that information. It could be in a document written in MS Word, or a PDF, or an internal blog that is similar to what you're reading now. This is what RAG does - it provides a "google search" of sorts with the data you've fed into it. If you fed it the 401(k) information on employer match vesting, your AI agent should find the answer.

If you’re feeding information to your LLM, the information does not need to be stored in a database, which is the traditional store of information. As I mentioned earlier, the information could be in the form of a PDF document or MS Word, or Google Docs stored in Google Drive. The “Google search” can combine data in a database, but it also scores documents for relevancy, just like a search engine would.

All the pieces of information are called embeddings, and the RAG system needs to process all these embeddings to find data relative to the conversation the AI agent is having with its users.

Why am I interested in RAG and embeddings, and training an AI agent?

It’s because I’ve had an app that has been sitting in the shelf for a bit while I focused on Badge Scanner and Video Director. That app is BRD Manager.

BRD manager is meant to help NetSuite clients and consultants alike create and maintain Business Requirements Documentation (BRD) in an application that provides consistency and adheres to the Suite Success methodology NetSuite provides.

It does this by offering an AI agent inside the app that has access to nearly 10,000 leading practices for NetSuite. What it does not have is first person knowledge of business processes beyond what it can normally bring to bear on the BRD discussions.

Out of the box, an LLM could tell you what a class is in NetSuite. Look up in the NetSuite docs and see what it says about what a class is. The docs give you a one paragraph description that is incomplete at best, and certainly does not cover best uses for the class object.

To help the AI agent along, I wrote a lengthy knowledge base post that has much more detail about what Classes are, how they are used, and gives real world examples of how Classes are properly implemented in NetSuite.

Where does this KB live? It lives in the same CMS this blog and the SuiteResources.com web site are created in and served from. The big difference is that the KB is not publicly accessible. It’s an internally facing KB that the LLM can see.

Now, when the subject of configuration of Record to Report, and more specifically, the Chart of Accounts and financial reporting is raised, the AI agent can respond more intelligently. This is good for customers, and it’s good for consultants.

In the end, that’s what we are aiming for - better implementation outcomes and more informed users and consultants.