RRAG (Restaurant RAG)

9h 1m

by Harry

What it does: 1. You give it a csv containing (restaurant name, general location of restaurant) e.g., (East is East, Vancouver Broadway) 2. You give it a series of fields you want the AI to fill out after conducting research (e.g., Menu items, price, address, website, etc) 3. For each restaurant in CSV, the AI autonomously researches that restaurant by (1) making targeted web searches (2) reading the results (3) deciding if it needs more info and do it again (4) reaching a point where it is able to synthesise the findings into an answer 4. The app returns the LLM-filled contents of the requested fields back to you in a uniform JSON structure To make a database for my main project (Voyago), I need a way to be able to quickly collect basic info for a lot of restaurants. I tried to make my own RAG (Retrieval-Augmented Generation, basically the process that goes on when you ask ChatGPT to search something up) system for this, but it frankly sucked and given the limited time I have with APs, I figured it'd be better if I cloned an existing RAG system from Github and adapted on it to fit my needs. I intended to do two main things: 1. Adapt it to Hack Club AI's API for LLM (previously it used Openai API) 2. Switch to Serper.dev API for web-searching tool (previously it used SerpAPI) 3. Make it accept bulk input of restaurant names (e.x., via CSV) and generate responses in a uniform data structure (e.x., via JSON) One of the things I had to learn and try to implement at the end of the day was a validate + retry system. This came because the RAG is really unstable previously (in one search you might get really satisfactory results, in the other you get none), and one of the biggest reasons seen through the debug prints is that the system doesn't go back and retry when it fails something. If it was able to just try once more, the RAG would've performed much better overall. **DEMO LINK COMMENT (IMPORTANT)** For this revised submission, I was asked to put my Pypi url in as the demo link. If this is another organizer and wonders where my video link is, here it is: https://youtu.be/_eLFlo78jj0 By the way, dear HCTG organiser, please note that I am 100% committed to attending HCTG, so if my projects are unable to be approved by May 8th for the 20 hour reduction, pls be soft on the deadline cuz I'll 100% be coming and I will most definitely make the 40 hours by the extended due date

retro api sdk

2h 18m

by sofia

an sdk allowing anyone to develop python scripts using the internal retro api endpoints!

stratus

6h 50m

by rena

algorithmic skies generated in browser