GiftBot

Inspiration

A ton of people (myself and the guy running the WeWork I am sitting at included) have issues with figuring out what gift to buy for people. The problem is often I have no idea how to go about figuring out what to but. So, I made a bot that can ask follow-up questions in order to help us reason about what to get.

On the more academic side this is an exploration of Chat UIs. I think that having "state" shown on the side can help enhance the UI. Furthermore, I am interested in using LLMs as reasoning agents, so trying to get a language model to reason about what gift to get is a good starting point.

What it does

On the left side there is a Chatbot. It remembers the history of the conversation and can ask you questions to figure out what products to recommend. After every questoin/answer the LLM generates a search query on a third-party Amazon API to recommend gifts. Overall gift ideas generated by the LLM are also shown in case you don't want to buy from Amazon. As the bot learns more about you it can narrow down the types of products you want to buy.

How we built it

• Mantine React components with a Next.js backend
• Self-Ask prompting techniques for the OpenAI portion (Also tried out some of the new Chain of Thought and other prompting techniques from @goodside's twitter ).
• Deployed on Vercel

What we learned

• CSS is annoying
• Amazon APIs suck. It took a while to find an Amazon product API that was < 5 seconds of latency and under $100/month.

What's next for GiftBot

• Fix up the Amazon API portion, get access to the official product API (you need to be manually approved, but it lets us earn a commission on any click-through purchases)
• Move to Python backend, which lets me use LangChain and other prompting chaining techniques more easily
• Apply the LLMs to other backends. What if I can have it ask questions to help me decide where to go for dinner, whom to vote for, what to work on...

Built With

• earl-grey