Chat-a-way

Our Landing Page

Inspiration

Planning for outings are difficult, and sometimes we just don't want to deal with the hassle of planning. So we came up with an AI solution for it, and hope others like it too.

What it does

Generate an efficient plan for any outings from just a short conversation with a user. Display the outing plan in a map to give users an idea of what they can expect.

How we built it

The frontend is built with SvelteKit, HTML, CSS and JavaScript. The backend is built with Python, Django, LangChain (GPT-4) and interfaces with Google APIs. It uses 2 AI-models (one user facing model and one data processing model) to continuously prompt and gather necessary information to provide a suitable outing plan. Once sufficient data is gathered, the backend does some calculations to get the best route to get around.

Challenges we ran into

LangChain was difficult to set up. Our first approach was to use agent-executors who utilized tools such a GooglePlaceAPI to search the web. There was a 50% chance it worked and it took us a long while to realize that it was an issue with using Agents with tools which potentially ended up in a endless loop. We had to scrap the idea and go back to regular chaining. Our backend algorithm also solves for TSP and this was challenging to fit our data too. Lastly, the Google API documentation was a huge headache as we couldn't find answers to some of our problems.

Accomplishments that we're proud of

Our frontend has clean interfaces and loading pages. Our backend uses 2 AI models that communicate information with each other, while still fitting their role where one is a sufficiently conversational model and the other is a data fitting model. Learning how to use Google's many APIs (Places, Text-Search, Maps, etc.)

What we learned

LangChain is in version 0.0.338 as of writing, and is highly unstable, Google doesn't have the best documentations, Agents are really cool, and they access the internet too using LangChain, there is a python module to solve TSP.