Toyota Motor Corporation’s Advanced R&D and Engineering Company builds secure RAG infrastructure, reducing research time by 20%

As Toyota Motor Corporation advances its transformation into a mobility organization, its Advanced R&D and Engineering Company drives research and innovation in future technologies. The company spans a broad range of technical domains, including body and chassis design, performance evaluation, system development, autonomous driving, material development, and regulatory certification.

Positioned at the forefront of advanced technology, the company began using generative AI at the end of 2022 and introduced RAG in select departments to improve search accuracy. However, independently developed RAG systems across departments led to fragmented implementation and individual optimization. In addition, building RAG systems presented challenges for teams without prior technical expertise.

To address these issues, the DS System Development Department within the Advanced Data Science Management Division, which leads cloud and AI-based technology development, established a shared RAG infrastructure for technology verification in accordance with Toyota’s security policies. The infrastructure was designed for broad accessibility across the organization.

“Each department built RAG for its own specific use case, which created challenges such as limited usability and dependence on individual experts,” says Mitsuhiro Mabuchi, group manager of the DS System Development Department. “As awareness of RAG’s effectiveness grew and demand increased across the organization, we decided to build shared infrastructure that could be used even by employees with no prior knowledge.”

The initiative began with a pilot in the Automotive Design Information Management Department. Engineers frequently received inquiries about internal design processes, requiring significant time to search for documentation and related materials.

“After listening to feedback, we recognized a strong need for improvement and launched a joint project between the Automotive Design Information Management Department and the DS System Development Department to address it,” says Mabuchi.

In October 2023, the DS System Development Department launched the project, selecting AWS GenAI LLM Chatbot—an AWS open source LLM chat solution—as the chatbot component and building the RAG-based infrastructure using AWS services.

“We chose AWS because of its strong track record within the Toyota Group. By using TORO PF, a secure platform for application development and deployment provided by our internal CCoE, we were able to build an environment with high levels of security and governance without compromising flexibility. We also valued the broad range of reliable AWS services and the extensive open-source resources available,” says Mabuchi.

The RAG infrastructure was developed through repeated fine-tuning to improve accuracy. Tetsuya Hashimoto, chief of the DS System Development Department, explains, “We adopted Amazon OpenSearch Service for its flexibility, allowing us to address feature limitations and quickly incorporate newly released capabilities that enhance accuracy. This included implementing a hybrid search system combining semantic search—which understands query intent to deliver optimal answers—with vector search, which retrieves results based on word frequency. Query expansion techniques were also incorporated to further enhance search precision.”

The system was also designed to accommodate variations in the spelling of specialized automotive terminology to ensure accurate results. “Regulatory documents often use uncommon terms. For example, car navigation systems are referred to as ‘screen display devices,’ and turn signals as ‘direction indicators.’ We generated paraphrases using an LLM and registered them as synonyms in Amazon OpenSearch Service to improve search accuracy,” says Mabuchi.

In parallel with improving search accuracy, the team implemented non-functional requirements necessary to provide the solution as shared RAG infrastructure. This included integration with Toyota's internal authentication system for centralized identity management, and implementing file-level access controls to enable secure cross-departmental sharing of uploaded files.

“Because the existing system placed a burden on account provisioning and management, we integrated it with our internal authentication system to make it easier for employees to use,” says Shinichiro Kawabata from the DS System Development Department. “We established three levels of authority—administrator, maintenance staff, and user—based on ID information, creating an access environment that takes data protection into consideration.”

During development, the team received support from AWS Professional Services and maintained close collaboration through weekly meetings, business chat, and task management tools. “As priorities shifted for functional requirements such as search specifications and accuracy based on feedback from the Automotive Design Information Management Department, the AWS Professional Services team responded flexibly, enabling productive discussions. Creating a  working mock-up in the early stages was particularly helpful, as it allowed us to refine the design while seeing it in action. We also received timely guidance on access control separation and UI improvements to meet the non-functional requirement of integration with the internal authentication system,” says Kawabata.

Development of the RAG infrastructure was completed in November 2024 and first deployed in the Automotive Design Information Management Department before it was subsequently expanded to regulatory and development departments. As of December 2024, approximately 150 users across eleven departments were actively using the system.

Although formal measurements have not yet been conducted, internal estimates suggest that response and research time was reduced by 20 percent. During the verification phase, users also reported an approximately 50 percent reduction in research effort when querying information already registered in the RAG system.

“We’ve received feedback that knowledge silos have been resolved, with source documents displayed alongside answers proving highly helpful and delivering a new user experience,” says Kawabata.

Through this initiative, the DS System Development Department was able to deliver a solution aligned with user needs while accumulating organizational expertise in RAG development. “In the R&D department, where it typically takes time to see results, we were able to make a visible contribution to reducing the time users spend researching in less than a year,” says Mabuchi. “At the same time, this effort to improve search precision contributed to our growth as an R&D organization.”

Looking ahead, the team plans to enhance the system’s ability to return highly accurate results by referencing multiple documents, even in cases where knowledge is limited, and to improve data registration methods. The platform will also continue to serve as a testing ground for emerging technologies.

The team is also considering providing the RAG infrastructure not only as a shared usage environment but also as source code and APIs to support departmental users who want to integrate it into their own systems or develop RAG solutions tailored to specific use cases.

“We plan to develop intuitive APIs and clear implementation methods to support both simple RAG deployment and the extraction of highly relevant answers,” says Hashimoto. “We expect AWS to continue strengthening its services as a stable foundation for long-term operations.”