TechPrufer
Get started
Uber
Uber logoUber · Feature trail

Delivery Search Platform Evolution

3 posts · Apr 2026 – Sep 2026

What this trail covers

Uber has significantly evolved its Delivery Search Platform, moving from traditional lexical matching to a sophisticated semantic search system leveraging deep learning models (Qwen backbone) with a two-tower architecture for real-time query embedding and batch document embedding. Key advancements include optimizing retrieval accuracy against infrastructure and compute costs through careful tuning of ANN parameters, quantization, and embedding dimensions, and ensuring safe, automated, and rollba. This post details the development and application of Two-Tower Embeddings (TTE) for recommendation systems, focusing on Uber Eats. It describes the architecture, challenges, and scalability benefits of TTE compared to previous models like Deep Matrix Factorization (DeepMF). The post highlights how TTE enables efficient candidate retrieval by generating embeddings for users and items, facilitating faster and more personalized recommendations. Future use cases for TTE in item feed ranking and generating transferable features for other domains are also discussed.

All Uber trails
  1. Halving the Time: How Uber Eats Rebuilt Its Search PipelineThis post details how Uber Eats rebuilt its search pipeline, cutting latency in half by optimizing UX (pagination, server-side caching, async rendering), retrieval (removing low-yield lexical strategies, earlier deduplication, product-level embeddings), feature population and ranking (splitting hydration, eliminating false dependencies, request hedging, GPU model serving, relevance-model cache), ads serving (column-oriented data layout, in-memory data, eliminating redundant serialization), and infrastructure tuning (parallel encoding/decoding, reduced embedding precision, multiple network connections, value types for GC reduction). A significant contribution was the use of an agentic AI loop for automated latency problem identification and fixing, validated by an LLM-assisted evaluation framework.
  2. Innovative Recommendation Applications Using Two Tower Embeddings at UberThis post details the development and application of Two-Tower Embeddings (TTE) for recommendation systems at Uber, specifically for Uber Eats. It introduces the concept of embeddings and TTE, explaining their architecture (query tower and item tower) and how they are used for candidate retrieval and ranking. The post highlights the limitations of the previous city-wise Deep Matrix Factorization (DeepMF) model, emphasizing its lack of scalability and high maintenance costs. It then presents TTE as a scalable, extendable, and efficient solution, detailing its advantages in scalable labeling, localized training/inference, and feature extensibility. The post also discusses future use cases for TTE in item feed ranking and generating transferable features for other domains like rider and place embeddings.
  3. Evolution and Scale of Uber’s Delivery Search PlatformThis post details the evolution of Uber's Delivery Search Platform to a semantic search system using Qwen LLMs and a two-tower architecture. It introduces Matryoshka Representation Learning (MRL) for flexible embedding dimension selection, scalar quantization (int7) for cost savings, and optimized ANN parameters (shard-level k) to balance retrieval accuracy with infrastructure costs. It also describes a blue/green deployment strategy at the index column level for safe, automated, and rollback-capable index refreshes, including completeness, backward-compatibility, and correctness validation checks.