AWS Public Sector Blog

How Northwestern University built a multilingual generative AI search tool with AWS

AWS branded background with text "How Northwestern University built a multilingual generative AI search tool with AWS"

Libraries are a treasure trove of knowledge, but finding what you need—especially in vast digital collections like the ones housed at Northwestern University Libraries (NUL)—isn’t always straightforward. That’s why the university chose Amazon Web Services (AWS) to help them build a new, multilingual, generative artificial intelligence (AI) powered search tool—one that makes its expansive collections more accessible, intuitive, and inclusive for all.

The project began when the team saw an opportunity to develop a better search experience for users struggling to find multimedia content using traditional search tools on their Digital Collections site. Their keywords, cataloging terms, and Boolean logic were no match for hundreds of thousands of digitized works spanning time and cultures. Serving more than 800,000 page views per year—many from casual users, multilingual researchers, and others new to library systems—the site clearly demanded innovative, scalable search solutions.

Reimagining search for a multimedia world

With the rise of generative AI paired with Northwestern’s rich and varied visual and audio content, the NUL team saw an opportunity to rebuild their users’ search experience from the ground up.

“One of our most highly trafficked sites is our digital collections,” explained Brendan Quinn, lead AI engineer at NUL. “That’s one of the reasons we picked it for our foray into generative AI work.”

These collections include hundreds of thousands of works and file sets from NUL’s music, Africana, transportation, and other libraries. Materials include everything from campus publications to maps, films, art, audio recordings, letters, manuscripts, and beyond. Northwestern University has long made these assets available to the public, but traditional search methods based on Boolean logic and catalog terms raised barriers to discovery.

Inspired by the release of popular consumer AI tools and emerging use cases for vector databases, Quinn got to work. In early 2023, he developed a prototype that combined large language models (LLMs) with vector databases to elevate semantic search. The pilot used public metadata from the library’s collections and showed how retrieval augmented generation (RAG) could surface relevant content based on meaning, rather than keywords.

“With a bare-bones demo, we were able to show a working example of how you can do retrieval over the data we already had,” Quinn noted. “From there, we decided to do a more formal project proposal.”

The initiative was a natural fit for Northwestern library’s digital products team. “They have the technical expertise and evolving organizational strategy prioritizing the use and development of AI-enabled tools and workflows within library services,” David Schober, digital initiatives product manager and team lead at NUL, pointed out. “Even the collections’ size proved ideal: It was large enough to be meaningful yet small enough to allow for rapid experimentation.”

Choosing AWS for flexibility, cost transparency, and scalability

Since NUL was an AWS pioneer and early adopter of its core cloud infrastructure, AWS was the natural choice for scaling their AI search tool. After testing multiple generative AI platforms, the team chose AWS for its flexibility, cost transparency, and support for rapid iteration.

The early version of the prototype ran on a cloud platform with a chatbot and a simple database to test the concept. But as the project matured, AWS services like Amazon Bedrock and Amazon OpenSearch Service offered a more scalable and resilient foundation.

Amazon Bedrock empowered the team to experiment with different foundation models, including Claude from Anthropic, which performed well in multilingual scenarios. Amazon OpenSearch Service supported semantic search by storing vector embeddings to help the system understand the meaning of documents and queries. This enabled users to search by concept rather than just keywords.

Additional AWS services deployed included:

This serverless architecture allowed the NUL team to remain agile and adapt infrastructure even days before launch.

Designing a conversational, multilingual search experience

The Northwestern team built the full solution in just a few months and pushed to production. They began development in mid-2023 and launched the tool publicly in fall 2024. On the front end, it looks and feels like a simple chat-based interface. Users can make natural language queries, even in multiple languages. The system then retrieves relevant results, complete with contextual explanations and links so researchers can explore complex topics and under-catalogued subjects.

Behind the scenes, the tool runs on a carefully integrated set of AWS services to power semantic search with vector embeddings. The embeddings represent data as sets of numbers to capture complex meanings, relationships, and contexts. When a user enters a query, Amazon OpenSearch embeds it as a vector and compares it against stored ones to find the closest matches. This enables concept-based retrieval to surface results that are relevant even if they don’t match the exact keywords. Amazon Bedrock then generates a response using a foundation model selected by the team.

To build trust and strengthen inclusivity, the team worked closely with faculty and students to refine the user experience. Features like help text, visible tool calls, and token streaming guide users through the search process while helping them understand the tool’s features and limitations.

Early impact and lessons for other institutions

Since launch, the tool has gained attention from academic institutions worldwide. Northwestern’s early work led to an Institute of Museum and Library Services National Leadership grant to further the work. The team has fielded consulting calls from dozens of universities and museums exploring similar projects.

Key outcomes include:

  • Improved accessibility: Users can search by concept, not just keywords—making collections more approachable for non-experts.
  • Multilingual support: The tool responds in the user’s language even when the metadata is in English—breaking down barriers to research.
  • Enhanced discovery: Semantic search surfaces content related to emerging and underrepresented topics, which may lack standardized metadata.
  • Flexible formatting: Users can request results in tables and other structured formats to enable faster research and analysis.

The NUL team is now developing an open-source tool called Treetop Discovery that uses key concepts and lessons learned from the Digital Collections site. The goal is to help other institutions explore generative AI search in their own environments.

For teams looking to get started, Quinn and Schober offer some hard-won advice: start small, experiment and iterate, and focus on embedding quality.

“Even if they’re trepidatious about large language models and AI, they should really look at embedding models and semantic search,” Quinn stated. “It’s a paradigm shift in search and retrieval—and it’s only going to get better in the future.”

Looking ahead: AI’s role in academic research

Northwestern’s work shows how academic libraries can lead the way in responsible AI innovation. By making digital collections more accessible and interactive, generative AI tools can shape how researchers engage with historical texts, rare manuscripts, and cultural artifacts.

“If someone jumps off ChatGPT and then lands on a traditional, database-driven search system, they’re going to jump right back to ChatGPT,” Schober said. “We need to be ready for what our users expect—and how they expect to search.”

As the team continues to refine the tool and expand its capabilities, they remain committed to open collaboration and community-driven development.

To learn how AWS helps institutions build, deploy, and scale AI solutions that support university needs, connect with an AWS sales representative.

Read related stories on the AWS Public Sector Blog: