Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scaling Multilingual Semantic Search in Uber Eats Delivery

Bo Ling, Zheng Liu, Haoyang Chen, Divya Nagar, Luting Yang, Mehul Parsana

TL;DR

A production-oriented semantic retrieval system for Uber Eats that unifies retrieval across stores, dishes, and grocery/retail items and shares key insights and practical lessons for building a unified, multilingual, and multi-vertical retrieval system for consumer search.

Abstract

We present a production-oriented semantic retrieval system for Uber Eats that unifies retrieval across stores, dishes, and grocery/retail items. Our approach fine-tunes a Qwen2 two-tower base model using hundreds of millions of query-document interactions that were aggregated and anonymized pretraining. We train the model with a combination of InfoNCE on in-batch negatives and triplet-NCE loss on hard negatives, and we leverage Matryoshka Representation Learning (MRL) to serve multiple embedding sizes from a single model. Our system achieves substantial recall gains over a strong baseline across six markets and three verticals. This paper presents the end to end work including data curation, model architecture, large-scale training, and evaluation. We also share key insights and practical lessons for building a unified, multilingual, and multi-vertical retrieval system for consumer search.

Scaling Multilingual Semantic Search in Uber Eats Delivery

TL;DR

A production-oriented semantic retrieval system for Uber Eats that unifies retrieval across stores, dishes, and grocery/retail items and shares key insights and practical lessons for building a unified, multilingual, and multi-vertical retrieval system for consumer search.

Abstract

We present a production-oriented semantic retrieval system for Uber Eats that unifies retrieval across stores, dishes, and grocery/retail items. Our approach fine-tunes a Qwen2 two-tower base model using hundreds of millions of query-document interactions that were aggregated and anonymized pretraining. We train the model with a combination of InfoNCE on in-batch negatives and triplet-NCE loss on hard negatives, and we leverage Matryoshka Representation Learning (MRL) to serve multiple embedding sizes from a single model. Our system achieves substantial recall gains over a strong baseline across six markets and three verticals. This paper presents the end to end work including data curation, model architecture, large-scale training, and evaluation. We also share key insights and practical lessons for building a unified, multilingual, and multi-vertical retrieval system for consumer search.
Paper Structure (29 sections, 6 equations, 1 figure, 11 tables)

This paper contains 29 sections, 6 equations, 1 figure, 11 tables.

Table of Contents

  1. Introduction
  2. Our key contributions are:
  3. Related Work
  4. Two-Tower Architectures
  5. LLMs as Embedding Models
  6. Training Techniques for Dense Retrieval
  7. Flexible Embedding Architectures
  8. Multi-Modal and Multi-Vertical Embeddings
  9. Learned Non-linear Scoring
  10. System Overview
  11. Training and Model Generation
  12. Online Query Serving
  13. Offline Inference and Indexing
  14. Methods: Data, Model, Losses, and Training
  15. Data and Features
  16. ...and 14 more sections

Figures (1)

  • Figure 1: Two-tower gte-Qwen2 architecture: query and document towers produce embeddings scored by cosine or an FFN head; training uses InfoNCE, Triplet NCE, and MRL.