Model Training and Methodology Overview
The model was built using AWS Bedrock's foundation models, leveraging Retrieval-Augmented Generation (RAG) to incorporate a custom knowledge base of VALORANT match data. The knowledge base consisted of structured JSON files categorized into three tiers: VCT International, VCT Challengers, and Game Changers. Each player entry included metadata such as handle, agent type, tier, and performance statistics.
Team Composition Methodology
The assistant generates team compositions by retrieving relevant player data based on the specific tier requested in the prompt. It uses metadata filtering to ensure that only players from the specified tier are considered. The model assigns unique roles to each player—Duelist, Initiator, Sentinel, Controller, and Flex—without repetition, matching agent types accordingly. Role assignments are justified using player statistics like kills, assists, damage, and ability uses.
Data Sources
Apart from the primary knowledge base, no additional data sources were used. All information was derived from the provided match data to maintain consistency and relevance.
Findings and Learnings
Throughout the project, we learned the importance of precise prompt engineering and configuration settings in AWS Bedrock to meet specific requirements. Enabling RAG with proper metadata indexing was crucial for accurate data retrieval. Lowering the model's temperature helped in generating more deterministic outputs, ensuring adherence to constraints like unique role assignments and tier-specific selections. Additionally, incorporating explicit instructions and output templates in the prompts significantly improved the assistant's performance.
Conclusion
By carefully configuring the model and refining the prompts, I successfully developed an assistant capable of generating accurate and effective team compositions based on VALORANT esports data, meeting all specified criteria within the project scope.
Log in or sign up for Devpost to join the conversation.