OmniSearchSage: Multi-Task Multi-Entity Embeddings for Pinterest Search (arXiv 2404.16260)
This entry indexes the April 2024 arXiv paper OmniSearchSage: Multi-Task Multi-Entity Embeddings for Pinterest Search by Prabhat Agarwal, Minhazul Islam Sk, Nikil Pancha, Kurchi Subhra Hazra, Jiajing Xu, and Chuck Rosenberg (Pinterest).
Key points
- Problem: Production search systems like Pinterest's typically train and serve separate embedding models for different entity types — queries, pins (items), and users. This multiplies storage and serving costs and limits transfer of learned knowledge across tasks.
- Approach: OmniSearchSage is a multi-task, multi-entity embedding model that jointly learns representations for queries, pins, and users within a unified framework, sharing parameters and training signals across search-related objectives.
- Results reported by the authors: The unified model saves approximately 60% of embedding storage relative to separate per-entity embeddings, while improving search engagement metrics in online experiments.
- Classification: Embedding models / representation learning for large-scale search and recommendation.
- Data: partition indexes, strip PII, maintain rollback-capable embedding versions.
- Latency: cascade retrieval with early stopping, cache popular queries, asynchronous reranking.
- Quality: interleave experiments and human audits to verify offline gains translate to online CTR/satisfaction.
- Cost: route to smaller/distilled models and combine sparse + dense retrieval.
- The Scandinavian Embedding Benchmarks: Comprehensive Assessment of Multilingual Embedding Models
- A Universal Framework for Compressing Embeddings in CTR Prediction (arXiv 2502.15355)
- Arctic-Embed 2.0: Multilingual Retrieval Without Compromise (arXiv 2412.04506)
- BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity (arXiv 2402.03216)
- BGE-en-ICL / BGE-ICL: Making Text Embedders Few-Shot Learners (arXiv 2409.15700)
- Beyond Benchmarks: Evaluating Embedding Model Similarity for Retrieval (arXiv 2407.08275)
- Original paper: https://arxiv.org/abs/2404.16260
Context from the original forum post
The post situates the paper in the evolution of neural retrieval: from BM25 and cross-encoders, through dual-tower dense retrieval and late interaction, toward generative retrieval and LLM-based agentic search. It emphasizes that industrial deployment must balance accuracy against latency (p99 budgets), cost per query, embedding version management, and safety concerns such as index poisoning.
The post also provides an engineering checklist relevant to adopting models like OmniSearchSage: