# Pinterest's Evolution in Embedding Retrieval Technology How Pinterest is tackling scalability and cost in its billion-embedding search platform. - **Source:** [Pinterest Engineering](https://medium.com/pinterest-engineering) - **Published:** 2026-09-11 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/9da3ff75-d479-41c0-ac4b-5bc665b21707 ## The Core of Pinterest's Discovery Engine Pinterest's discovery engine relies on understanding deep semantic relationships through embeddings to help users find inspirations. The Manas platform, serving billions of embeddings, powers major product surfaces like Home Feed, Search, and Ads. However, scaling to tens of billions of embeddings presents challenges in cost efficiency and flexibility. Traditional ANN algorithms like HNSW are memory-intensive, requiring entire indices in RAM, which increases costs linearly with corpus size. ## Challenges in Embedding Retrieval The classic two-tower retrieval paradigm limits each candidate to a single embedding, scoring relevance through a simple dot product. This approach lacks flexibility for richer, context-dependent similarity notions. As Pinterest's corpus grows, the need for more efficient and scalable solutions becomes critical. The team identified three key areas for improvement: quantization, SSD-based serving, and multi-embedding retrieval to address these challenges. ## Quantization: Reducing Memory Footprint Quantization compresses high-dimensional float vectors into compact integer representations, significantly reducing memory usage. Pinterest implemented Scalar Quantization (SQ) and Product Quantization (PQ), each with trade-offs. SQ offers strong compression with minimal recall loss, while PQ achieves higher compression but with a notable recall decrease. Benchmarks showed SQ reducing HNSW index by 59% and IVF index by 75%, with recall over 90%. ## Online Experiments and Results Pinterest conducted online A/B experiments to select the best quantizer for each use case, ensuring negligible impact on user engagement. SQ and PQ were rolled out across major product use cases, delivering over 50% memory reduction in embedding indices and 20–30% cost savings in serving infrastructure. This optimization was crucial for maintaining performance while scaling the platform. ## SIMD and Linear Scaling Shifting to 8-bit or 4-bit representations required addressing CPU bottlenecks. Pinterest implemented Linear Scaling SQ, which eliminates the decoding step before distance computation. Using SIMD intrinsics, the team reduced computing resources per query by 10–15%. This innovation allowed for efficient handling of quantized vectors without compromising performance. ## SSD-Based Serving: A Cost-Effective Alternative To further reduce costs, Pinterest explored serving ANN queries directly from SSDs, which are cheaper per gigabyte than RAM. By carefully bounding I/O per request, the team achieved high throughput with low tail latencies. Early experiments demonstrated a 10x reduction in memory usage and 40% CPU savings compared to in-memory serving, marking a significant milestone in cost efficiency. ## Multi-Embedding Retrieval Moving beyond the single-vector-per-candidate constraint, Pinterest introduced support for richer scoring functions that consider multiple embeddings per candidate. This approach enables more expressive ranking at the retrieval stage. A pilot use case is currently being tested with a product team, promising to unlock new capabilities in semantic search. ## Technical Deep Dive: SIMD Intrinsics The use of SIMD intrinsics was pivotal in optimizing quantized vector operations. By allowing the CPU to perform multiple 8-bit integer operations per instruction, Pinterest reduced computational overhead. This technique, combined with linear scaling, ensured that quantization did not become a bottleneck, maintaining low query latencies across the platform. ## Future of Embedding Retrieval at Pinterest Pinterest's innovations in quantization, SSD serving, and multi-embedding retrieval set a new standard for scalable and cost-efficient search platforms. These advancements not only address current challenges but also pave the way for handling future growth. The team continues to explore new techniques to further enhance the Manas platform's capabilities and performance. ## Sources covering this story - https://medium.com/pinterest-engineering/evolving-pinterests-embedding-retrieval-platform-aede4e831e01?source=rss----4c5a5f6279b6---4 — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._