Alpha Three team boosts AI assistant accuracy with smarter data chunking for enterprise Q&A
The "Alpha Three" team boosted AI assistant accuracy by optimizing data chunk size during pre-processing, preserving context for clearer, more relevant answers. Their AWS-based system improved query experience and reduced errors.
"Alpha Three" Team Boosts AI Assistant Accuracy by Optimizing Data Pre-Processing
Wednesday 4 June 2025
Generative AI (GenAI) continues to influence how companies approach product development and business processes. A recent report by Taiwan's Market Intelligence & Consulting Institute (MIC) found that in 2024, 19% of Taiwan's top five industries engaged with GenAI, with finance and insurance leading at 25%, followed by manufacturing at 22%. Despite widespread adoption, some companies have seen disappointing results from their AI assistant projects, leading to project cancellations and reduced competitiveness.
How "Alpha Three" Improved AI Assistant Performance
The team "Alpha Three," winners of the "2025 AI Wave: Taiwan Generative AI Applications Hackathon" hosted by Walsin Lihwa's "Smart Manufacturing" group, identified a key issue behind poor AI assistant performance: overly small data chunking during pre-processing. Breaking documents into tiny pieces disrupted paragraph context, causing AI models to misunderstand queries and deliver less accurate answers.
To solve this, the team proposed using the "amount of text in a single PDF page" as the chunk size. This approach preserves natural paragraph flow and context, reducing semantic gaps and improving the AI’s understanding. The hackathon review committee unanimously praised this method for three main benefits:
- Enhancing the user query experience
- Lowering the risk of hallucinations in responses
- Improving semantic coherence and answer accuracy
Training and Technology Behind the Success
Alpha Three tested their method using a steel standard inquiry: "Does ASTM A276 steel grade 316Ti comply with the EN 10088-3 standard?" The AI assistant provided focused, accurate information on chemical composition and standard specifications, scoring a perfect 1.0 in search relevance, answer solidity, and answer relevance.
The team built their enterprise knowledge question-answering framework on Amazon Web Services (AWS). Their process includes:
- Uploading files (PDF, PNG, JPG) to Amazon S3 for cloud storage
- Using a comprehensive language model with a Flask API for quick query responses
- Leveraging Amazon Bedrock to connect foundation models, improving scalability and response time
- Employing Amazon EC2 to speed up API processing and ensure system stability
This end-to-end system covers data upload, management, retrieval, and response, making it easy for users to submit questions and receive swift, professional answers. This streamlined design was a key factor in their victory.
Learning on the Fly: From Graduates to AI Innovators
Alpha Three’s members are recent National Taiwan University graduates with information engineering degrees. Despite their technical background, they initially lacked experience with current mainstream AI tools. Thanks to organizer-led workshops on enterprise data and AWS Generative AI, along with insights into steel standards from Walsin Lihwa, they completed the project in just 30 hours and secured the top prize.
For product development professionals looking to enhance AI assistant projects, this case highlights the importance of data chunking strategy and practical use of cloud-based AI tools. Understanding these technical nuances can make the difference between an underperforming AI assistant and one that reliably delivers accurate, relevant information.
To explore more about AI tools and training that can support your projects, visit Complete AI Training's latest courses.
