‘AIntuition’: Retrieval Augmented Generation (RAG) for Public Services and Administration Tasks by ITU
In this challenge, the quality of accompanying documentation will account for 10% of the total score! What information to include? Here's a list of elements to consider:
Overview of Solution Components: Provide a clear list or table detailing all solution components and their functions.
Workflow Review: Offer a high-level description or schematic illustrating how these components interact with each other to achieve the desired outcome.
Key Advantages: Briefly outline the key advantages of your solution over a simple or baseline RAG. Highlight any improvements or innovations made.
Requirements and Dependencies: List all requirements and dependencies necessary for implementing and running your solution.
Installation and Usage Guidance: Provide step-by-step instructions for installing and using your solution. Include any tips or best practices to ensure smooth implementation.
Troubleshooting Guide: Offer guidance for troubleshooting common issues that may arise during installation or usage.
Additional Information: Feel free to include any other relevant information or advice that could facilitate effective installation and use of your solution.
Having a comprehensive documentation will also make it easier to demonstrate your solution if it gets shortlisted!
In addition to the above, you can also include in your documentation some observations about the challenges you encountered while designing/improving a RAG system, as well as potential directions for future research.
@RomanSever Are we allowed to use Llama3 or we stick to Llama2?
Hi! For the purposes of this challenge, please stick to llama2 7b (as this is the model specified in the challenge description). If you run into any technical issues with Llama2 specifically, you can try another model of a similar size, but would need to include a brief justification in your accompanying documentation. Ideally, the solution should be designed in a modular manner (allowing for swapping and seamless integration of different models into the pipeline).
Finally, are there any compute restrictions?