---
title: "Evolving Pinterest’s Embedding Retrieval Platform"
description: "Pinterest describes how its Manas distributed search platform is evolving embedding retrieval for a corpus growing toward tens of billions of embeddings. Traditional HNSW indexes require all data in RAM, while single-vector two-tower retrieval limits context-dependent relevance, motivating changes across quantization, SSD-based serving, and multi-embedding retrieval. Manas uses Scalar Quantization and Product Quantization across HNSW and IVF, with SQ preserving over 90% recall while reducing index size, and Linear Scaling plus SIMD cuts per-query computing resources by 10–15%. For disk-based serving, SPANN with PQ is reported as 4.5x faster than plain SPANN and over 3x DiskANN QPS, while a five-billion-embedding production evaluation saved over 40% CPU time with under 5% recall drop. The platform is also adding ColBERT-style late interaction using Sum of MaxSim, with a pilot in progress."
---

# Evolving Pinterest’s Embedding Retrieval Platform

[Pinterest](https://yomu.fyi/company/pinterest) · Pinterest Engineering · Sep 11, 2026

**Type:** Explainer

## Summary

Pinterest describes how its Manas distributed search platform is evolving embedding retrieval for a corpus growing toward tens of billions of embeddings. Traditional HNSW indexes require all data in RAM, while single-vector two-tower retrieval limits context-dependent relevance, motivating changes across quantization, SSD-based serving, and multi-embedding retrieval. Manas uses Scalar Quantization and Product Quantization across HNSW and IVF, with SQ preserving over 90% recall while reducing index size, and Linear Scaling plus SIMD cuts per-query computing resources by 10–15%. For disk-based serving, SPANN with PQ is reported as 4.5x faster than plain SPANN and over 3x DiskANN QPS, while a five-billion-embedding production evaluation saved over 40% CPU time with under 5% recall drop. The platform is also adding ColBERT-style late interaction using Sum of MaxSim, with a pilot in progress.

## Context

Manas must support embedding retrieval as Pinterest’s corpus grows toward tens of billions of embeddings. The post identifies memory costs from RAM-resident HNSW indexes, latency and I/O constraints for SSD serving, and the limited context-dependent relevance of single-vector two-tower retrieval.

## Approach / What changed

The team evolves Manas through vector quantization, SSD-based ANN serving, and multi-embedding retrieval. It implements SQ and PQ, Linear Scaling SQ with SIMD, and SPANN with quantized on-disk embeddings and in-memory centroids, while adding ColBERT-style late interaction with multi-embedding queries and Sum of MaxSim scoring.

## Takeaways

- Scalar Quantization reduced HNSW index size by 59% and IVF index size by 75%, while recall remained consistently above 90% in the reported benchmarks.
- SPANN combined with PQ was 4.5x faster than plain SPANN and achieved over 3x DiskANN’s QPS with one-third the latency, at a 5% recall drop.
- Multi-embedding retrieval adds ColBERT-style Sum of MaxSim scoring, requiring Manas to parse multi-embedding queries and run multiple ANN searches simultaneously.

**Tags:** [Distributed Systems](https://yomu.fyi/topic/distributed-systems), [Performance](https://yomu.fyi/topic/performance), [Quantization](https://yomu.fyi/topic/quantization), [Search](https://yomu.fyi/topic/search)

- Source: [Pinterest](https://medium.com/pinterest-engineering/evolving-pinterests-embedding-retrieval-platform-aede4e831e01)
- Source URL: https://medium.com/pinterest-engineering/evolving-pinterests-embedding-retrieval-platform-aede4e831e01
- Ingested by Yomu: 2026-09-11T16:02:22.670Z

[Read original post](https://medium.com/pinterest-engineering/evolving-pinterests-embedding-retrieval-platform-aede4e831e01)