ChatSOS

Inspiration

We were super inspired by Gabriella's story and instantly knew we wanted to help her in this enndeavor.

What it does

ChatSOS is a chat bot that takes the place of a 911 operator only you don't have to speak English to it and you actually don't have to speak at all. Powered by Chatgpt, ChatSOS supports over 50 languages and even allows users to respond in emojis for speed and comfortability. As a 911 operator ChatSOS has to be able to do everything a 911 operator can do and more reliably everytime. This is why we have programmed knowlegde into ChatSOS to make sure it will always respond exactly how a 911 operator would respond.

How we built it

After looking through some of the datasets we found that the datasets were not big and diverse enough for fine tuninng. Afraid that they might have the reverse effect on the model we went for a RAG strategy instead. Specifically we used a vector database. Everytime there is a new call we ask what is the emergency. Then with that response we run it throuugh our database and find the best section of the training manual that corresponds to their emergency. Then we pass that information along to the chatbot.

Challenges we ran into

One of the biggest challenges was suprisingly prompt engineering, especially with the added context injection. The robot's responses changed alot depending on minor tweaks and changes to the prompt. We needed it to pay attention to the context but not too much attention that it wasn't adaptable to situations on the call.

Accomplishments that we're proud of

One thing that was really important for us is that the robot not just collect information but also offer life saving advice and it is actually pretty good at that. We have noticed the bot giving instructions for the Heimlich maneuver and CPR or certain occaisions.

Project Philosophy

During this project, we thought a lot about who we were building it for. Likewise, the made sure to think of what mattered like what tech stack the nonprofit uses, and how we can make operational costs as low as possible. We went out of our way to design a stateless backend so the entire chatbot can be run serverless. We also stressed the significance of making it as fast as possible while keeping cost low. In the end we ended up finding two models and separately optimizing them. We have chatgpt3.5 for better responses and Nous-mistrial-7B-SFT for speed and cost. The one in the video is actually the slower version. We've been able to get it to run four times as fast and at a fifth of the cost.

What's next for ChatSOS

We didn't get time to implement streaming but that would be really nice because it would update a lot faster. Hopefully being integrated into AccessSOS

Built With

• openai
• python
• react-native
• together.ai