Building the infrastructure that powers modern artificial intelligence search can take hours or even days as systems construct digital roadmaps connecting billions of data points. A team of computer scientists at the University of Maryland and Google Research has dramatically accelerated this process, building a search index for one billion items in less than 20 minutes on a single powerful computer.
The team developed PiPNN, a new algorithm that accelerates the construction of these search indexes. Its developers include computer science doctoral students Tobias Rubel and Richard Wen, their adviser Laxman Dhulipala, and collaborators at Google Research. Dhulipala is an assistant professor of computer science with an appointment in the University of Maryland Institute for Advanced Computer Studies (UMIACS).
PiPNN addresses a challenge known as “approximate nearest-neighbor search.” When a user asks an AI to find related documents, recommend a video or answer a question, the system can find relevant information by treating data as points on a map and searching for nearby matches.
Wen likens the concept to looking for a place to eat in your hometown. “Imagine you want to hang out with a friend, and there are a bunch of restaurants. You want to pick one that’s closest to their house,” he explains. “In the real world, we’ve already solved that problem in two dimensions.”
But an AI doesn’t just look at two dimensions like latitude and longitude. It maps data across hundreds or thousands of dimensions that can capture characteristics such as an article’s topic, tone and vocabulary. To find the “closest” matches quickly, the system must first build a web of connections linking similar items.
While these digital roadmaps enable fast searches, building them is a notoriously slow process. Previous methods build the map by running practice searches, stepping from point to point to find connections.
As Rubel explains, this stepping method is inefficient for computer hardware. Bouncing between random locations in a computer’s memory forces the machine to repeatedly pause and fetch new data.
PiPNN—affectionately named after Dhulipala’s family corgi, Pippin—bypasses this stepping process during construction. Instead, it divides the massive dataset into overlapping local neighborhoods, which the computer can process in efficient batches.
To prevent too many overlapping connections in dense neighborhoods from slowing the system down, the team developed a memory-saving method called HashPrune.
“If you have two candidates that are very close to each other ... they kind of give you the same navigational ability, so I don’t want both of them,” Wen says.
HashPrune processes incoming connections and removes redundant ones, retaining useful, diverse paths without overloading the machine’s memory.
Tests on a powerful cloud computing machine demonstrated PiPNN’s ability to handle large datasets. According to the researchers, PiPNN builds indexes up to 12.9 times as fast as the state-of-the-art HNSW algorithm and up to 11.6 times as fast as Vamana. Average construction speedups range from sixfold to tenfold.
Faster construction can lower computing costs, improve energy efficiency and help systems keep pace with constantly changing information.
“If your dataset changes every hour, and it takes you more than an hour to build the index, you literally can’t use that index,” Rubel notes.
Faster construction also allows developers to fine-tune system settings—a process that often requires repeated trial and error—in minutes rather than waiting weeks.
PiPNN recently won the Best Research Paper Award at the ACM SIGKDD 2026 conference, and the researchers secured first place in the 2026 SISAP Indexing Challenge. Follow-up work also won a best paper award at the Vector Database workshop at the 2026 International Conference on Very Large Data Bases.
The work grew out of a two-year collaboration. Dhulipala, who is also a research scientist at Google Research, worked closely with Google colleagues throughout the project. The UMD team led the experimental design and implementation, with a steady exchange of ideas helping shape the final algorithm.
“With Tobias and Richard, we initially started working on a very different project to solve approximate nearest-neighbor search with very sparse data,” Dhulipala recalls. After hitting roadblocks, they shifted direction, drawing on ideas Rubel was exploring for handling changing datasets. Those ideas ultimately led to PiPNN.
“I’m extremely proud of them, and proud that this line of work that they contributed to is being recognized in this way,” Dhulipala says.
Supported in part by a recent $200,000 National Science Foundation award, the team is also addressing PiPNN’s current limitations. Because the algorithm requires a large amount of memory while building the map, the researchers are exploring ways to move operations to cheaper storage drives and adapt the algorithm for datasets that stream in continuously.
Meanwhile, the tech industry is taking notice of the team’s open-source code. Microsoft is working to integrate PiPNN into its systems, and the researchers are discussing potential adoption with companies including Meta and NVIDIA.
—Story by Diya Sharma, UMIACS communications group