Pinterest Engineering has slashed serving costs by 20–30% across production workloads by deploying compression techniques across its distributed search platform, Manas, which now handles tens of billions of embeddings. The company detailed these advances in a technical report published by InfoQ in September 2026, describing how it addressed mounting memory and infrastructure costs as its underlying corpus scaled rapidly into billions of items. Manas powers core discovery experiences including Home Feed, Search, Related Pins, Ads, and Notifications across 80 clusters.

Testing on a 100-million-embedding GraphSage dataset revealed sharp trade-offs between compression methods. Product Quantization compressed original floating-point vector representations into compact byte codes, shrinking HNSW indices by 74% and Inverted File indices by 93%, while delivering 70–80% recall. Scalar Quantization compressed vector components into lower-bit integers, cutting HNSW indices by 59% and IVF indices by 75%, but maintained over 90% recall consistently. Offline benchmarks showed the baseline HNSW index at 121 GB with 93.72% Recall@100 at 302.5 QPS, while HNSW plus SQ dropped to 50 GB with 92.92% recall at 305.2 QPS. The baseline IVF index measured 97 GB with 91.69% recall at 1659.8 QPS, but IVF plus SQ compressed it to 25 GB while achieving 95.71% recall at 1588.8 QPS.

For SSD-based serving designed to move index storage off expensive RAM and onto high-throughput solid-state drives, the company evaluated DiskANN and SPANN, discovering that SPANN combined with PQ delivered three times the queries per second of DiskANN with one-third the latency and only a 5% recall reduction. In a Pin recommendation evaluation indexing over 5 billion embeddings, SPANN saved more than 40% of CPU time for production queries compared to full in-memory HNSW. The team also implemented Linear Scaling SQ using SIMD intrinsics to eliminate the CPU bottleneck caused by decoding lower-bit representations before distance computation, trimming query compute resources by 10–15%.

Traditional memory-hungry vector search algorithms like standard HNSW faced mounting cost, hardware allocation, and infrastructure flexibility challenges as Pinterest's corpus scaled rapidly into billions of items, the report explains. The custom SPANN architecture addresses this by keeping a small, fast centroid index in memory to locate relevant partitions while storing large posting lists on SSDs, optimizing input/output operations per second and ensuring search efficiency at scale. To move beyond the expressiveness limits of single-vector two-tower models, Pinterest is now shifting toward multi-vector Late Interaction models like ColBERT using Sum of MaxSim scoring to fine-tune relevance matching across tokens. A pilot launch is currently underway with an internal client team to test advanced multi-embedding query support in live production environments. As recommendation systems grow more sophisticated, the tension between model expressiveness and infrastructure cost will define which platforms can sustain real-time personalization at planetary scale without exhausting compute budgets.