Before a search assistant can retrieve an answer in seconds, its searchable collection may take hours to prepare. Researchers at the University of Maryland and Google Research have developed PiPNN, an algorithm that built approximate-nearest-neighbour indexes for billion-scale datasets in under 20 minutes on a single multicore machine.[1][2]
That machine was a substantial one: the billion-scale experiments used a Google Cloud c4-highmem-288 instance with 288 virtual CPUs and 2.2 TB of DRAM. The paper reports construction speedups of up to 12.9 times over HNSW and 11.6 times over Vamana. Comparisons within each chart used the same hardware, making the gains a comparison of indexing approaches rather than computer sizes.[2]
Building the roads that search follows
Vector search represents items such as documents or images as lists of numbers. A graph links nearby vectors so a search can travel from a starting point towards promising matches. Constructing that graph is expensive: repeated searches can send a processor hopping between unpredictable memory locations, where fetching data becomes part of the bottleneck.[1][2]
PiPNN divides the dataset into overlapping local groups and compares candidates in batches. Those batches make better use of efficient matrix-computation routines. Its HashPrune component keeps a sparse collection of useful connections as candidates arrive, limiting the intermediate memory needed to construct the graph. The result is a faster route to a navigation structure that can then serve searches.[2]
The reported construction results
| Measure | Result | Test context |
|---|---|---|
| Billion-scale index build | Under 20 minutes | 288 vCPUs and 2.2 TB DRAM in the billion-scale experiments |
| Comparison with HNSW | Up to 12.9× faster | Index construction in the reported workloads |
| Comparison with Vamana | Up to 11.6× faster | Index construction in the reported workloads |
| Typical construction improvement | Roughly 6×–10× | Range described in the university’s research report |
Faster rebuilding creates room for fresher information
For a collection that changes frequently, a long rebuild can constrain how soon new material becomes searchable. Reducing that work gives an operator more freedom to replace an index, compare settings or prepare separate customer collections. It also makes experimentation less expensive: a team can test several graph configurations in time previously occupied by one build.
Memory remains an important operating constraint. The research’s large machine is part of the result, and a production team must balance build time against the memory needed to construct and serve its index. Retrieval quality also has to be compared at a useful query latency. The university identifies storage efficiency and streaming updates as directions for further work.[1][2]
PiPNN puts attention on a stage of AI infrastructure that users rarely see. Faster preparation does not change the language model, but it can change how quickly a search system is ready to give that model useful material.
Sources & further reading
- UMD Researchers Accelerate the Technology Behind AI Search University of Maryland Institute for Advanced Computer Studies · 20 Oct 2026 · Accessed 25 Oct 2026
- PiPNN: Ultra-Scalable Graph-Based Nearest Neighbor Indexing University of Maryland / research authors · Accessed 25 Oct 2026