Unifying a Retail AI Stack with MongoDB

THEIR CHALLENGE

Unifying databases to support a next-generation AI shopping assistant

Retail brands have long faced the challenge of giving every online shopper personalized, real-time guidance at scale. A growing AI company set out to close that gap, equipping retail brands with intelligent agents that could engage customers across every channel, in any language, around the clock.

The company operates an intelligent shopping assistant that understands what a customer wants and surfaces the right products in real time. A shopper can describe an upcoming occasion or their style preferences, and the assistant surfaces tailored recommendations from the retailer’s full catalog. Brands began sending millions of catalog items through the platform, and the company experienced rapid growth as adoption expanded across the retail sector.

The solution’s architecture relied on three databases running on Amazon Web Services (AWS): DynamoDB for core NoSQL storage, Pinecone for vector-based semantic search, and Elasticsearch for keyword filtering and faceting. Pinecone handled natural-language queries, Elasticsearch enabled precise filtering by category or attribute, and DynamoDB housed the underlying product data.

As the company scaled to support the needs of hundreds of retail clients, synchronizing data across all three systems grew complex. As a result, inconsistencies were inevitable; a shopper might ask about a product’s price and get outdated information because one database hadn’t finished syncing. The company sought a single source of truth that could handle document storage, vector search, and keyword filtering.

OUR SOLUTION

Using MongoDB Atlas to consolidate the AI data layer

To prepare for its next phase of growth, the company adopted MongoDB Atlas as a unified data platform. The team was able to get started easily on MongoDB, rapidly prototyping and validating their use case without a long procurement cycle. As their architectural complexity grew, they later transitioned to a strategic partnership with an account team. The integrated system, fully managed by MongoDB, delivers robust reliability and enterprise-grade infrastructure, allowing the team to focus on building new features rather than managing databases.

MongoDB Atlas provided core NoSQL document storage, while MongoDB Vector Search on Atlas eliminated the need for a standalone vector database. MongoDB Search on Atlas delivered keyword filtering and faceting capabilities. By consolidating three systems into one, the company established a unified data foundation for every product, document, and conversation flowing through its platform.

The architecture is organized around three data buckets. The first is the product catalog. Product data is ingested, broken down into embeddings, and stored alongside each record so that MongoDB Search on Atlas can retrieve relevant information instantly. Across the company’s retail clients, this amounts to millions of items.

The second is supplementary documentation, such as sizing charts, installation manuals, return policies, and other materials. These are embedded across text, image, and structured data. When a shopper asks whether a piece of furniture requires assembly, for instance, the assistant can pull from an embedded installation manual and deliver a precise answer—drawing on multimodal data rather than product descriptions alone.

The third is conversation data. Every shopper interaction across every deployed website is stored in MongoDB Atlas. Both keyword and semantic search are layered on top, giving brands the ability to search and analyze those exchanges at any time. For shoppers, this means the assistant retains context across a session, so a follow-up question like “Do you have that in blue?” doesn’t require restating the original request. For brands, it creates a searchable archive of real customer intent; they can filter by date range, topic, or status to identify trending questions and surface gaps in product information.

The company also integrated Voyage AI by MongoDB into its pipeline. When a shopper asks a question, the system pulls back a pool of semantically similar documents. Before those results reach the large language model for summarization, the assistant uses Voyage AI to rerank them for contextual precision. Previously, the team relied on a custom-built reranking process that required manual tuning and cross-validation. With Voyage AI handling that layer, the manual effort was eliminated, and the most contextually accurate documents now surface consistently. This provides shoppers with the most accurate and relevant answers.

As the business evolved, the company’s infrastructure needs evolved with it. A shift in the cloud environment meant the entire MongoDB Atlas deployment needed to migrate from AWS to Google Cloud (GC) to align with a new operational footprint. Because the team had already consolidated its three databases in MongoDB Atlas, the move to GC was significantly simpler.

Rather than rearchitecting three separate data systems for a new cloud environment, the team only needed to migrate a single, unified platform—a process that the cloud-agnostic design of MongoDB Atlas made considerably more straightforward. The migration took two and a half months, supporting 180 million monthly active users with effectively zero downtime.

OUTCOME

Delivering reliability at scale, from migration to Black Friday

The new platform has proven itself under peak pressure. It has now been through three Black Friday/Cyber Monday cycles on MongoDB Atlas, and the 2025 season was its highest-volume event to date. The system handled an unprecedented surge in data throughput, shopper interactions, and real-time queries without a single engineer being called in overnight.

With one platform replacing three, the company removed the operational overhead of managing separate sync pipelines, vendor relationships, and scaling configurations. Every product, document, and conversation flows through a single source of truth, giving the team confidence to iterate quickly and ship new capabilities without worrying about data consistency across systems. The company is working with MongoDB Professional Services to optimize indexing and performance, keeping its data layer ready for whatever comes next—from this year’s Black Friday surge to the latest leap in AI-driven retail.