LLMs can serve as powerful engines to extract insights from your data. @abacusai provides one of the fastest ways to build LLM-powered question-answering systems on top of your data: Here is a full demo using Karpathy's YouTube lectures: Step 1: Obtain your data The first step is to obtain your data. I am using @karpathy’s YouTube lecture series called Neural Networks: Zero to Hero. I transcribe all the lectures using OpenAI’s Whisper model. I end up with a list of text files on which I will build a LLM-powered question-answering system. Step 2: Upload and create feature groups The raw text transcriptions are uploaded as is to the @abacusai platform which takes the text files and indexes them and extracts features automatically. After this, I can quickly create a feature group which is important to denote the Document ID and actual feature containing the raw text. Step 3: Create a Document Retriever We are leveraging a document retriever for our data-augmented question-answering system. You can easily create one inside the platform. You select the feature group, chunk size, chunk overlap, and a text encoder (we use OpenAI Embeddings in our example). Step 4: Choosing the LLM After creating the document retriever you are ready to select your LLM. The platform offers a range of options, including Abacus's own models, GPT-4, Claude 100K, and Llama 2. I selected GPT-4 (8K) for this demo. You could also run an evaluation to help pick the best model. Step 5: Data LLM Once the data and model are set, you are ready to deploy the system as a dashboard and predictions API. I use the prediction dashboard and start to explore my LLM-powered question-answering system. I can ask any question about the content of the lectures. As an example, Karpathy does mention Tree of Thoughts in his State of GPT talk so I ask about it and the model outputs the response. You can also ask all sorts of interesting questions like summarizing lectures and creating outlines. All of this was done in less than an hour! That’s incredible! What I like about the @abacusai platform is how easy it is to build a working prototype and continue iterating to improve it. I can’t recommend Abacus enough. They are enabling AI developers like myself to leverage powerful recipes to build and deploy custom LLM-powered chat and question-answering systems on top of my data. Let me know if you would like to see more user cases or other things like evaluations or experiments with models like Llama 2.

Looking to learn how to quickly build custom chat LLMs? I highly recommend this free upcoming workshop. A few more seats left! eventbrite.com/e/hands-on-wo…

How can you build your own custom ChatGPT-like system on your data? This is not easy as it could require complex architecture and pipelines. Given the high demand, I started to explore the ChatLLM feature by @abacusai. I’m very impressed! Let's take a look at how it works: