4 Use Cases For Vector Search

New developments in technology are allowing eCommerce platforms to perform in optimized ways like never before. Vector Search allows search functions on your site to run more fluently and efficiently, improving customer experience and boosting sales. This post will take you through four use cases for vector search to revolutionize your online experience.

What Is Vector Search?

Vector search is a method for efficiently finding and retrieving similar items from large data sets. These are based on representations of the data in high-dimensional space and then similarities drawn between the comparisons. These representations can be from text, document, images, sounds, videos, or any kind of data input. Whatever the input they are represented as vector embeddings. Distance metrics then quantify the closeness, and links can be drawn between the two. This allows for a new and effective way of search.

Use Case 1: Search & Discovery/Intuitive Search

A multi-task vector search model can be trained to perform multiple tasks at the same time. For example, the model can be trained by pairing search queries on the one hand. On the other hand, the model can simultaneously be trained with content the shoppers have viewed or saved on the website.

Then, when they search, the results can show results related to a combination of keywords and previous actions. Word embeddings can be created to make this link, and associated word embeddings throughout the search query can be made. 

Use Case 2: Keyword & Semantic Search

Traditional search methods based on keywords are clearly not always effective. This can lead to ‘no search results’ if there aren’t any matching keywords, which is a big negative for your eCommerce business aiming to help customers. Vector embeddings with query text as the input embedding can be a solution to this.

Textual metafields data can be found through titles and descriptions, and then trained with positive pairs. This means that there won’t be a 0 results page, as related query text will be sourced and shown.

Semantic search does offer a valuable new side to search and discovery, but it cannot replace keyword search. A hybrid approach is the best option available as of now, and re-ranking based on findings through success and vector search is the best option currently available.

Use Case 3: Visual Discovery

Vector search does not only use textual data input to create results. A multi-modal model can process and integrate data from multiple input types such as:

  • Images
  • Text
  • Audio 
  • Video

Then offer predictions based on a variety of sources. These multi-modal models can be trained using a large dataset of image-title listings, using various input sources. Then list embeddings can be generated based on what was found.

This means that visual search can be expanded across all categories. Users can discover related items through images alone, also known as Visual Discovery. This is an intuitive and exciting way of product discovery that adds great value to an eCommerce store. 

Use Case 4: Real-Time Personalized Recommendations

Recommendations can be improved through list-embedding techniques to offer real-time personalization. Embeddings can be created through user behaviors and click profiles. These can be trained over listings that are already in use. When a user clicks on listings, these are moved closer together. When a user skips over listings, they are moved further away. This real-time data can then be used to optimize the user experience.

When combining this real-time data with other non-personalized features of the data such as price and geography, optimal recommendations can be created. Personal taste is a daily changing factor, and vector search offers the option for this to be featured. 


Vector search offers new and exciting ways to optimize the eCommerce experience. Whilst these methods are still in development, they promise a revolutionized search experience for shoppers that will ultimately boost sales and humanize the way that eCommerce search and discovery functions.