Adam Sand explains how to create a company "second brain" by ingesting proprietary business documents into a vectorized data store so AI can answer company-specific questions. He contrasts general LLM knowledge (public facts like constitutional articles or earnings calls) with missing internal knowledge (PTO policy, pricing, CRM workflows) and argues that RAG (retrieval + autonomous generation) using vector stores is required. He outlines the workflow: upload documents, split text into overlapping chunks (parameters like 1,000–2,000 character chunk size and 200-character overlap are discussed), create vectors, store in a vector DB (example: Pinecone), use a retriever with an OpenAI model to answer queries, and gives an example training on Jeff Bezos's annual shareholder letters. He warns that simply linking a chatbot to a website is insufficient because internal policies and systems are not present on public pages.
Chapters
This is how you're going to eventually replace humans in your business with AI is by putting your processes, your policies, your pricing, your answers to questions. Your second brain will be a vectorized data store. If you were to ask ChatGPT, hey, what is Article 3 of the United States Constitution? It knows that if you were to say, hey, how do you structure annual board meeting? It'll tell you that. You say, hey, tell me how to analyze the profit to earnings of Tesla versus Apple. It can tell you that. But if you say, hey, what is the PTO policy, the paid time off policy at Roofing Business Partner or Liberty Roofing Pros, it's not going to know because it wasn't trained on that data. That data doesn't exist. So for us to create that second brain, so to speak, we start getting into having conversations about something called reg. Reg means just retrieval, autonomous generation. What that means is you give it your own data. So what this is doing is it allows you to say, okay, if I put a file into this Google Drive, I want you to add it to a vector store. All a vector store is, is it says, okay, when AI first started, I said it was comparing cats to dogs, right? It was learning to identify dogs, you know, basically showed pictures of cats and it got to learn the difference between like that is a cat and that is a dog and a dog is not a cat. So what it did was it converts pictures to basically what they call vectors. And the easy way to kind of say it is like every color had a number and they looked for patterns, it's just looking for patterns. And that said, okay, if this pattern matches the dog pattern, it's a dog. If it doesn't, it's a cat. And that got smarter over time. So that's like vector. Now what essentially does with a document is it says, okay, if you have ever tried to upload an entire book to ChatGPT, you'll realize it kind of remembers the beginning and it remembers the end, but it doesn't remember the middle. And that's because it didn't vectorize it. If you wanted duplicate your brain, right? Tim Brown was trying to do this the other day, but he was trying to do it by creating a GPT. We tried that a long time ago and it doesn't work because the vector thing, so what I did is I trained it on the annual letters to the shareholders from Jeff Bezos. So annual letters to shareholders from Jeff Bezos, the Amazon shareholders. Really, really, really long. It's too much for the context window of AI. That you can't just spit the document in and tell it to remember everything on it. And if you just went to chat gbt, it might pull things from a earnings call, it might pull things from a article. You get these hallucinations. So if you wanted to only think about the thing that you wanted to think about, this is in terms of like, if you want your company to have a chat bot or a customer service agent or an inside salesperson that works for you, you're going to have to have a vectorized database. And so the simple explanation of it is that you say, okay, take this massive block of text and make it simpler for you to access it. So it'll say, okay, for every 2,000 characters, we're going to chop it into vectors of 2,000 character blocks and then we're going to overlap it. And that overlap is this vectorizing. It's not always those characters and those overlaps, those are adjustments that you can make when you're setting something like this up. That's why you see down here, recursive character text splitter. This is splitting the text. So you're saying a thousand chunk size with a 200 character chunk overlap. So I said for Jeff Bezos, like a smaller chunk size with a significant overlap. Because I want to see what happened. And I kind of playing with this, this is how we get to the point of being able to say we want to have it access that. So what you see in this workflow, I would say about three hours of work. So when trying to think about how people are charging you for AI, I'm not saying this is just the three hours of work. It's the knowing how and the what to do. But what this is, is it's basically saying, upload a document, make a vector store out of it, then, then you can chat. And what it does is it uses the OpenAI model to use the vector store retriever to go to this pinecone vector store. And then using AI, it's going to then give you back an answer and output that answer. So you can come in here and say stuff like, what was it like at Amazon during the dot com bubble? It's only going to respond and say, this is what it was like. So during the dot com bubble and crash, Amazon explains this. It's pulling only this Amazon letter to the shareholders. That's what this is like. This is how you're going to eventually replace humans in your business with AI is by putting your processes, your policies, your pricing, your answers to questions your second brain will be a vectorized data store. Now there's a lot of unknown unknowns that even I'm still learning about. I wanted you guys to see that this is the back end infrastructure of what you're doing. So when a lot of people say we're going to get you a chatbot, that we're going to link to your website and it's going to magically become a second brain that knows everything about your business, is going to support your customers, it's bullshit. Because your website does not have everything about your business and everything, all your policies, all your procedures. It doesn't know how to interact with your CRM. It doesn't know how to interact with your production management system. It doesn't know your pricing. It doesn't know your markets. It doesn't know your common answers to questions that are in your brand's way. It needs more. It needs more information for that rag that retrie evil to work.
