In an era of exponential data growth, organizations face the pressing challenge of efficiently managing and retrieving internal information. Large enterprises, where employees depend on extensive documentation for daily decisions, often find traditional search tools inadequate. Recognizing this issue, Gautam Chettiar, an MSc in Applied Computing (MScAC) student at the University of Toronto, began an applied-research internship at Advanced Micro Devices, Inc. (AMD) to tackle the problem.
His work led to the development of a state-of-the-art chatbot system that leverages large language models (LLMs) to streamline enterprise information retrieval. By integrating a multimodal retrieval-augmented generation (RAG) approach, the system significantly enhances the scalability, precision, and accuracy of documentation indexing, enabling employees to quickly access the knowledge they need.
A Novel Solution for a Pressing Problem
The inability to efficiently parse and retrieve data from extensive repositories slows workflows and impairs productivity, especially in fast-paced industries like semiconductor development. To address this problem, Chettiar’s project focused on improving AMD’s chatbot system by incorporating a multimodal framework that indexes text, images, videos, and other data types.
“Developing this system at AMD was an incredible opportunity to apply cutting-edge AI techniques to real-world enterprise challenges,” said Chettiar. “By integrating multimodal retrieval-augmented generation, we significantly improved how knowledge is indexed and accessed, making information retrieval faster and more accurate resulting in enhanced productivity for employees across the company. This experience reinforced my passion for solving complex AI problems at scale, and I’m excited about the broader impact such innovations can have in transforming enterprise workflows.”
The project expanded the system’s capacity to handle new types of information, such as visual and auditory data, while refining its handling of traditional text-based sources. Using advanced machine learning algorithms, the system extracts and contextualizes content from diverse documentation, making it more accessible and actionable.
Addressing Social and Business Needs
Beyond the technical advances, this project has significant implications for how enterprises manage information. In a fast-paced work environment like AMD’s, engineers and developers rely on quick access to accurate information to meet tight deadlines. The enhanced system cuts the time spent searching for relevant documentation, allowing employees to focus on innovation and problem-solving.
On the business side, this innovation provides a competitive edge. With the ability to synthesize data more effectively, AMD can optimize workflows, reduce operational inefficiencies, and make more informed strategic decisions. Furthermore, the project aligns with AMD’s broader commitment to innovation in high-performance computing and AI solutions.
Technical Breakthroughs and Results
Chettiar’s work introduced a multimodal RAG framework that outperformed previous text-only indexing methods. By designing and implementing a new indexing pipeline, the system achieved 16 per cent cumulative variance gain in information capture over prior methods; improved retrieval accuracy, with an accuracy of 88.4 per cent, precision of 90.2 per cent, and recall of 97.8 per cent; and enhanced question-answering performance, where tests on select knowledge bases showed a 29.1 per cent improvement in answering accuracy compared to the prior system.
These metrics highlight the system’s ability to distill and synthesize complex data across media types into a searchable and user-friendly format. Employees can now query the chatbot to extract precise answers and insights from a broader range of documentation sources, including text, images, and videos.
“Gautam’s contributions were instrumental in expanding the chatbot’s capabilities, particularly by integrating image and video content into its indexing system, thereby significantly enhancing coverage and response quality,” said Yonas Bedasso, the director of engineering, leading the AI and data engineering teams at AMD and Chettiar’s industry supervisor. “In addition to his technical achievements, Gautam exhibited forward-thinking leadership in preparing the system for future advancements in answer quality improvement.”
Driving Innovation through Applied-Research Collaborations
Chettiar’s work showcases a successful collaboration between academia and industry. The project was supervised by Bedasso at AMD and supported by Professors Frank Rudzicz (University of Toronto and Dalhousie University) and Eldan Cohen (University of Toronto, Department of Mechanical & Industrial Engineering). This partnership demonstrates the power of applied research to drive tangible innovation.
AMD has long valued its partnerships with academic institutions. As a global leader in high-performance computing, the company actively researches and applies AI and machine learning technologies to enhance software development, making collaborations like this essential to its mission. To date, AMD has recruited 28 applied-research interns from the University of Toronto’s MScAC program, partnering since 2021.
The success of this multimodal RAG system illustrates how applied research can address real-world challenges while advancing the frontiers of technology. By incorporating cutting-edge machine learning techniques, Chettiar’s project not only addresses a pressing need at AMD but also serves as a model for other organizations grappling with similar challenges.
