A small group of us has been working to create an AI Generation Engine, and someone recently took it over the line and gave it a user interface. The engine takes any PDF and makes a custom AI for you to use. Soon, obviously, it will take any kind of content and build an AI from that. It’s powered by OpenAI.
Future upgrades will link it to systems that actually do things — order processing and sales, Salesforce, inventory, live chat and WhatsApp. We’re not far off ticking a box and enabling voice. We’ll also be creating what are known as “fine-tuned” models, which can communicate as if they were a specific person, style or brand.
One note on the difference. This PDF-to-AI generator has unlimited memory. You can load in thousands of pages of content and all of it will be available to the AI. Most existing projects only let an AI hold a page or two in its working memory. Even GPT-4 ships with an 8K-token context window, with a 32K version in the API. Everything beyond that has to live somewhere else and be fetched when it’s needed. I’m sure there’s something you could cobble together in Python that does the same, but with no interface — and we couldn’t find one in any event.
The somewhere else is a vector database. Pinecone’s chief executive described the arrangement in April: “think of the LLM as an application almost that sits on top of this database.” The model reasons, and the database keeps the record.
I think we’re near an important realization — and an upgrade — that you’ll find interesting. No idea if this will make any sense, but it looks like the AI does not retrieve data from its attached vector DB the way our human logic might expect. We need to understand how the AI, and the DB, interpret their tasks, and then how they query for subtable context to generate a response. An Excel sheet full of human-readable data is not what works well for these tools. With that understanding, we may be able to upgrade the engine and have Excel sheets correctly interpreted, for easier bulk data loading.
