AI-powered retail: Personalization and accuracy

Solution overview

In the dynamic realm of e-commerce, staying ahead means embracing the power of technology. This solution uniquely combines MongoDB's flexible data handling with Databricks' advanced analytics to transform how online retail operates. By integrating these technologies, retailers can leverage AI-augmented search capabilities, offering a more intuitive and efficient shopping experience.

The heart of this solution lies in its ability to personalize the shopping experience. Using AI, machine learning, and vector search, it can understand and anticipate customers' needs, suggesting products that align with their preferences and search history, and respond accurately to semantic context. This not only enhances customer satisfaction but also drives sales by presenting the most relevant products.

This solution leverages the advanced capabilities of MongoDB Atlas Vector Search, integrated seamlessly with Atlas and Databricks, to redefine the online shopping experience.

Atlas Vector Search marks a significant evolution from traditional keyword-based search methods. It utilizes sophisticated algorithms to understand and process complex search queries, enabling a more nuanced and relevant product discovery process. This approach ensures that customers find exactly what they're looking for, even with vague or partial search terms.

MongoDB Atlas plays a pivotal role in this solution by providing a robust and scalable database environment. It efficiently manages the diverse and voluminous data inherent in e-commerce platforms, from product catalogs to customer behavior analytics. This seamless integration makes the foundation of the search process both sturdy and flexible.

Databricks complements this setup by offering powerful analytics and machine learning capabilities. It processes vast amounts of data in real-time, deriving insights that continually enhance the search algorithm's effectiveness and accuracy.

By combining these technologies, the solution offers an unparalleled search experience in the e-commerce domain. It not only meets the current demands of online shoppers for quick and precise search results but also anticipates future needs, adapting and evolving continuously. This results in higher customer satisfaction, increased engagement, and ultimately, a significant boost in sales and business growth.

Other applicable industries and use cases

• Insurance
• Healthcare
• Finance

Reference architectures

Shown here are two different architectures: one for the general solution and the other one is highlighting the Vector Search part of the solution. Each part complements the other.

Data model approach

The document data model is inherently flexible because it allows developers to optimize schema design to support the needs of their applications. Schema design will determine the performance, scalability, and reliability of the application. The golden rule of schema design is the following: Data that is accessed together should be stored together. One of the most foundational schema designs developers should know is the Polymorphic Pattern.

The Polymorphic Pattern is applied when documents within a collection share a similar, though not identical, structure. This pattern proves beneficial for querying information from a single collection efficiently. By organizing documents in a way that aligns with the intended queries, rather than distributing the data across multiple tables or collections, we enhance overall performance. This approach is particularly effective when the focus is on streamlined access and retrieval of data from a unified collection.

The polymorphic pattern in this product catalog is applied by designing a flexible schema where each product document can have varying fields or related sub-fields (like colour1, ageGroup, season, articleType) specific to its category or type, while also maintaining common fields (like _id, link, brandName, price, title) across all products, allowing for both uniformity and customization in representing different product types within the same collection.

Building the solution

• Step 1: Set up vector search. Follow the steps listed in this tutorial to model your documents for vector search, learn how to index vector embeddings for vector search and run vector search queries.

• Step 2: Pre-requisites and how to deploy. Follow the step by step instructions to deploy the application locally.

• Step 3: Databricks jobs and workflows. Follow the Databricks instructions to create Databricks jobs and workflows with JSON.

• Step 4: Databricks notebooks. Catalog indexing, product score, and pricing.

Key considerations

Data transformation: Recognizing the importance of transforming raw data into a more usable format or structure. In this project, triggers and functions were used to push raw data from MongoDB Atlas collections into Databricks by leveraging the MongoDB Connector for Spark, so the data can be shaped as needed to feed the different machine learning algorithms.

Real-time data processing: Understanding the importance and mechanisms of processing data in real-time — especially in the context of the retail industry where timely insights can drive immediate actions, such as product scoring, promotions, and recommendation engines — and of semantic search.

Technologies and products used

MongoDB developer data platform:

• Atlas Database
• MongoDB Connector for Spark
• Change Streams
• Atlas Search
• Atlas Vector Search
• Aggregation Pipelines

Partner technologies:

• Databricks

Author

• Francesco Baldissera, MongoDB
• Ashwin Gangadhar, MongoDB
• Vittal Pai, MongoDB