Marketing Personalization
Delivery Hero needed to replace stale, batch-generated substitute recommendations that hurt revenue and customer satisfaction whenever perishable grocery items went out of stock.
Reported impact
Delivery Hero moved from stale 24-hour batch recommendations to sub-second live substitutions, protecting revenue and customer satisfaction while creating a globally scalable recommendation service.
Delivery Hero rebuilt its Item Replacement Tool around MongoDB Atlas on Google Cloud so product, customer, and embedding data could be queried in real time instead of once-a-day batch jobs. The first version depended on product data in Google BigQuery plus recommendations cached in Redis, which meant riders still saw outdated alternatives after inventory or customer preferences changed. After moving the product catalog to Atlas and adopting Atlas Search, the team used a multilingual sentence encoder built with Hugging Face tooling and run in Google Vertex AI to create product embeddings. As catalog records change, those vectors are continuously updated and indexed by MongoDB Vector Search on Atlas. When a rider finds an unavailable product, the picking service queries Atlas to combine semantic similarity, product filters, and customer preferences, then returns up to 20 replacement options to the customer in under a second.
Product catalog changes are embedded with a multilingual model in Google Vertex AI, persisted in MongoDB Atlas, and indexed by MongoDB Vector Search on Atlas; when a rider encounters an unavailable item, the picking service queries Atlas with customer preferences and inventory filters, then sends up to 20 replacement options back to the customer in under one second.
The picking service uses vector search to recommend up to 20 suitable replacement items.
The total roundtrip latency from query in the app to returning recommendations to the user cannot exceed more than one second.
“We needed to move to an up-to-second real-time recommendations system, and that is what MongoDB Vector Search enabled us to do.”
“Within the MongoDB Atlas platform, we can store, index, and query vector embeddings right alongside our product and customer data, with everything fully synchronized.”
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
MongoDB Sample Archived · Mar 6, 2026 Beni Beni needed to search and rank more than 300 million one-of-one resale listings refreshed by over a million items per day, but its split Milvus and MongoDB architecture created brittle ETL synchronization, expensive… Beni manages over 300 million listings that refresh by more than a million items daily. MongoDB Atlas · MongoDB Search · MongoDB Vector Search MongoDB Sample Dec 23, 2025 Rezolve AI Rezolve Ai’s legacy product discovery system relied on a contextual, AI-based search engine that excelled with subjective queries. However, the system could not support the precise searches that B2B customers required. We saw a reduction in latency by up to 50%, and we’ve achieved an average response time latency of roughly 50… MongoDB Atlas · MongoDB Atlas Search 
Alibaba Cloud Sample Oct 13, 2026 TOPSPORTS (滔搏) TOPSPORTS wanted to scale enterprise AI agents beyond isolated pilots, but every new agent risked recreating the same knowledge, data, tools, and skills in separate silos. Business-analysis, membership, and marketing… business analysis time reduced to 20 minutes AgentCore · AgentLoop · MCP tools 