This post has been republished via RSS; it originally appeared at: New blog articles in Microsoft Community Hub.
Congratulations to all the participants and thank you for your amazing contributions!
We are thrilled to announce the winners of our first Microsoft AI Chat App Hack, which challenged developers to build applications using RAG (Retrieval Augmented Generation) in order to answer questions in custom domains. If you missed the hackathon, you can still catch up on the live streams.
We were impressed by the variety of the submissions, which ranged in programming language used (Python, JS, TS, C#, C++), natural language used (English, Spanish Portuguese, Chinese, Bangla, Hindi, Swahili, German), and knowledge domain (medical, education, technology, history, and much more.)
Everyone did an amazing job, but the judges have now determined the winners of the cash prizes, along with an honorable mention for each category. Drumroll, please... :drum:
Best Overall: DocAssistant.Swaggy
This app took RAG to another level! First, you upload OpenAPI schemas for APIs to the Azure AI Search index. Once the schemas are ingested, you can ask questions that are answerable by an API, like "what were the top movies for 2023?". The app searches to find possible API endpoints to answer the question, then uses an LLM to suggest the API URL, fetches the URL to get the API response, and calls the LLM again to turn the response into a user-friendly answer. :collision: Judges loved how the app added additional steps into the typical RAG flow to provide a complete end-to-end experience for easy API exploration.
See: Video | Code | Submission
Honorable Mention: SecureBot
This app answers cybersecurity questions based on a rich data source of podcasts, video transcripts, websites, and books, all in the Spanish language. Judges were impressed by its very different architecture from most submissions, combining Copilot Studio with Power Platform. When a question comes in, the studio calls Power Automate to call an Azure Function to translate that question into Spanish and searches the Azure AI Search index with the translation. That step improves the relevance of results, and the bot still responds in the user's question thanks to prompt engineering.
See: Video | Code | Full submission
Best Data Source: DubsBot
This app helps students at UW (University of Washington) find courses for their interests from a massive course catalog. The team built a custom parser for the course catalog webpages, which included details like expanding abbreviations and making time schedules more human-readable. The judges loved that the team put so much detail into customizing the data ingestion, since that can make such a big difference to the quality of the answers from the LLM.
See: Video | Code | Full submission
Honorable Mention: Copilot for Azure Pricing
This app helps developers select the most appropriate Azure region and VM type based on up-to-date pricing data. The judges loved how this app combined traditional RAG search, on Azure documentation, with near real-time data from the Azure Retail API. The app orchestrates the calls using Langchain agents and multiple "tools" that know how to look up regions, find VM types, and get prices, so that it can produce the correct answer, and visualizes the answers along the way.
See: Video | Code | Full submission
Best in Your Language: Ask an Entrepreneur
At first glance, this is a standard RAG app that answers questions based off the speeches of famous entrepreneurs. However, there's a catch: the source documents are in both English and Chinese, and questions may be asked in either English or Chinese. The team put an enormous amount of effort into figuring out the best way to set up the data ingestion (which included different chunking strategies for Chinese characters vs. English characters) as well as the question-answering (which now includes a translation step that poses the question in both English and Chinese). The judges were impressed by their efforts and were excited to see a team tackle the challenge of multi-lingual RAG.
See: Video | Code | Full submission
Honorable Mention: Ley GPT
The goal of this RAG chat is to demystify Mexican federal work law, thanks to a data set of Mexican laws and regulations. To make the chat more accessible to Mexican citizens, the developer translated everything fully into Spanish: the UI, the system prompt, the few shot examples, etc. The judges loved the thoroughness of the translation, and the mission of using an AI chat to make labor laws more accessible to everyone.
See: Video | Code | Full submission
Most Helpful Community Member: Zed Haque
Zed was a constant presence in our discussion forum, helping to answer many questions from other participants, while also working on their own submission. Thank you, Zed, and everyone else who helped out, for bringing a collaborative spirit to our hackathon.
We would like to congratulate all the winners and thank all the participants for their amazing contributions. We hope you enjoyed the hackathon and learned something new along the way. Stay tuned for a follow-up post showcasing all the hack submissions!