How Superintelligent Personalized the Learner Experience

Superintelligent is the learning platform for AI, offering tutorials on AI uses and tools that are hands-on, practical, and easy to follow. When debuting their platform, Superintelligent imagined a fully personalized, AI-powered user experience, but where do you start with a project that big?

Fractional AI partnered with Superintelligent to build an AI chatbot using RAG (retrieval augmented generation) as a first step towards that vision.

Problem

As the one-stop-shop for AI learning, Superintelligent offered great breadth, with hundreds of tutorials updated daily. Learners needed a way to quickly navigate the platform to find relevant AI tools for their specific needs.

Say you’re a lawyer looking for AI tools to help with contracting. You have a specific task in mind, but finding the right tools through traditional search is difficult, especially since you don't know what those AI contracting tools are called or if they even exist. You need a way to find what you need, fast.

*The Superintelligent learning platform*

‍

Solution

Fractional AI partnered with Superintelligent to build an AI chatbot using RAG to offer members specific tool recommendations.

Here’s what this looks like in production:

‍

Since the AI tool landscape is continuously changing and learners rely on Superintelligent for the most up-to-date recommendations, the AI chatbot is built to be perpetually refreshed – the Superintelligent team can just add new tools to its underlying tool database, and the AI chatbot automatically syncs those updates.

‍

Looking Under the Hood

If you’re curious about RAG, here’s a straightforward example.

‍

Project Setup

Data – To get started, Superintelligent supplied us with their large database of AI tools. This proprietary dataset is the key to up-to-date, personalized recommendations.
Data processing – Because Superintelligent's tool database is updated regularly, we wrote some scripts to detect changes to Superintelligent's primary database and reflect them in Pinecone, the vector database we used for this project.
Model – We used GPT-3.5 Turbo, GPT-4 Turbo, and the OpenAI embeddings API for different components of the chatbot.
Tooling - We used Langchain to stitch together our prompts, Langsmith for observability and experimentation, FastAPI for the API server, and Pinecone for the vector database.

*Here’s a snapshot from Langsmith (this is the output from LLM 1 in the workflow below).*

‍

How Your Chatbot Message Turns into AI Tool Recommendations

Let’s say you type the following into the chatbot “I’m a photographer looking for AI tools to help with headshots.” How does your message turn into a reliable tool recommendation?

The following diagram is a simplified depiction of the automated workflow.

‍

The pipeline is simple enough but this workflow, like all workflows, required testing and wrangling: chat histories expanded beyond the context window, there were ample hallucinations (including lots of made up tools!), and regressions were difficult to systematically track when we tweaked prompts.

‍

Impact

Hundreds of recommendations have been provided by the chatbot. Now learners can spend less time searching and more time learning.

Here’s a testimonial from the Superintelligent CEO: