Awesome Search KG

Food Discovery with Uber Eats: Building a Query Understanding Engine

3 min read

Food Discovery with Uber Eats: Building a Query Understanding Engine

Source: https://eng.uber.com/uber-eats-query-understanding/
Author: Uber Engineering

Summary

Uber Eats engineering describes building a production query understanding engine for food search — covering intent classification, entity extraction, query rewriting, and the specific challenges of food domain search.

Food Search Domain Challenges

Food search has unique QU challenges:

  1. Cuisine ambiguity: “Thai” = Thai food OR Thai restaurant specifically?
  2. Dish-category gap: “pad see ew” (specific dish) → “Thai food” (category) → “noodles” (broader)
  3. Ingredient queries: “no gluten” → filter + ranking adjustment, not text matching
  4. Contextual meaning: “spicy” means different things to different users
  5. Hybrid search need: “McDonald’s near me” (navigational) vs. “healthy salad” (semantic)

Uber Eats Query Understanding Architecture

Layer 1: Query Classification

Multi-label classifier assigns query categories:

  • Dish-level (“pad thai”)
  • Restaurant-level (“Shake Shack”)
  • Cuisine-level (“Japanese food”)
  • Attribute-level (“vegan options”)
  • Combo (“cheap Thai near me”)

Multiple labels can apply simultaneously.

Layer 2: Entity Extraction

NER extracts:

  • Restaurant names: “McDonald’s”, “Chipotle”
  • Dish names: “Chicken Tikka Masala”, “Pho”
  • Cuisines: “Thai”, “Italian”, “Indian”
  • Attributes: “vegan”, “gluten-free”, “spicy”
  • Location terms: “near me”, “downtown”

Layer 3: Query Rewriting

Transform extracted entities into structured search:

Input: "cheap Thai food near me"
Output: {
    "cuisine": "Thai",
    "price_filter": "low",
    "location": "user_coordinates",
    "sort": "relevance"
}

Layer 4: Fallback and Expansion

If structured query returns few results:

  • Expand cuisine synonyms (“Thai” → [“Thai”, “Southeast Asian”])
  • Relax filters
  • Expand geofence

Key Technical Choices

BERT-based NER

Fine-tuned on Uber Eats domain data (dish names, restaurant names):

  • Base BERT misses “pad see ew” (recognizes “see” and “ew” as separate tokens)
  • Domain-fine-tuned BERT correctly identifies multi-word dish names

Semantic Search Integration

For exploratory queries (“comfort food”, “something different tonight”):

  • Pure keyword matching fails (no keyword overlap)
  • Semantic Bi-Encoder embedding retrieves restaurants matching the intent

Results

Post-QU system deployment:

  • +15% search session success rate
  • -20% zero-result queries
  • +8% order conversion for “discovery” queries (exploratory, non-specific)
  1. Domain entities are critical: invest in domain-specific NER
  2. Intent-first architecture: classify intent before retrieval, not after
  3. Fallback strategy: over-constraining queries is the main failure mode
  4. Hybrid mandatory: navigational (exact) + semantic (exploratory) both needed

People

Graph View

Backlinks