Evolution of Pinterest's Embedding Search Platform

Pinterest's distributed search platform Manas implemented three key technical improvements to overcome the cost and performance limitations of embedding search. The existing HNSW approach required keeping the entire index in RAM, causing costs to increase proportionally with corpus size.

Quantization and SSD-Based Serving

Applying Scalar Quantization (SQ) reduced embedding index memory by over 50% and cut serving infrastructure costs by 20–30%. In a benchmark on a 100-million-item GraphSage dataset, SQ demonstrated a superior balance between compression ratio and recall compared to PQ, maintaining recall above 90% while reducing HNSW index size by 59% and IVF index size by 75%.

Additionally, a method was introduced to serve ANN queries from NVMe SSDs instead of RAM. Initial experiments showed a 10x reduction in memory usage and a 40% reduction in CPU usage. Notably, combining PQ quantization with the SPANN architecture optimized disk storage space, improving speed by 4.5x compared to the original SPANN and increasing QPS by more than 3x compared to DiskANN.

Multi-Embedding Search Support

To overcome the limitations of Two-Tower models that represent data with a single vector, Late Interaction models (such as ColBERT) were introduced. Documents and queries are represented as lists of vectors, and Sum of MaxSim scoring logic is applied to preserve token-level granular semantics. The Manas query parser is currently being updated to support multi-embedding queries, and pilot testing is underway.