Vector-Based Search and Product Discovery: A More Intelligent Solution
How generative AI is filling huge gaps in traditional search functionality
I still remember the first time I used online search. Prior to the rise of search giants like Google and Bing, Amazon – then a new online bookstore – offered an innovative way to find books by topic or author. In 1997 after attending a SilverStream Software developers conference in Washington, D.C., where a new technology from Oracle called “Java” was all the rage, I wanted to learn more. Searching for “Java” related books on Amazon yielded a variety of results featuring Indonesian island, coffee, and software-related titles.
Interestingly, this trip to Washington, DC, marked my first encounter with Starbucks, a rapidly expanding retailer from Seattle that had no locations in San Diego or Southern California at the time.
Fast forward 27 years, and today’s search solutions still exhibit the same weaknesses Amazon did in 1997. Search algorithms can identify similar items within a product catalog based on shared text attributes, but they struggle with deeper semantic understanding. They often fail to differentiate between implied meanings that arise when the same keywords or phrases are used in different orders or contexts.
For example, traditional text-based search engines treat the phrases ‘milk chocolate’ and ‘chocolate milk’ as essentially the same. By breaking these terms into the keywords/tokens ‘milk’ and ‘chocolate,’ the search engine attempts to match these against the product attributes in the catalog, returning similar results for both searches. The problem with this result is clear: ‘milk chocolate’ refers to a type of chocolate flavor, while ‘chocolate milk’ is an entirely different product.
Vector-based search can now solve for this use case and many more, but what is vector-based search and how does it work?
Vectors – A Better Way to Search
Vector-based commerce search refers to a method of retrieving search results using vector representations of queries and items instead of traditional keyword-based systems. By converting words, products, descriptions, user behaviors, and other relevant data into high-dimensional vectors, this technique leverages the capabilities of machine learning and natural language processing to understand and measure semantic similarities.
In other words, unlike traditional keyword searches, vector-based searches can capture the semantic meaning and context of words, resulting in more relevant search results.
“This technology has revolutionary implications for retail, as search results are no longer limited by catalog data.” Kevin Jackson, CriticalKPI
Search Beyond the Catalog: No Keywords? No Problem!
A common problem vector-based search can solve for is providing relevant results when there are no keyword matches within a catalog.
I have worked with numerous brands and retailers encountering searches for products or brands they do not offer. This issue is often addressed by creating keyword synonyms for unavailable brands (e.g., “Nike” = “Brooks”) and/or redirecting these searches to manually-curated product listing pages. However, these solutions are neither elegant, scalable, nor easy to manage.
Unlike traditional search, vector-based search does not rely on exact keyword matches. Instead, it understands the semantic meaning behind a query and finds catalog products that align with that meaning. For instance, most search engines won’t return results for “Nike” if Nike products are absent from the catalog. However, a vector-based search understands that Nike sells running, walking, and basketball shoes, as well as sporting apparel, and can then locate similar items within the catalog.
Vectors – What took you so long?
While the concept of using vectors in data science is well-established, their application in commercially available search solutions is new. The delay in adoption is understandable: vectors can exist in hundreds of dimensions, representing complex relationships that demand extensive data storage and processing capabilities. However, recent advancements in data compression techniques now enable rapid searches within these multi-dimensional data models.
Traditional search analyzers compare key attributes such as category, brand, gender, size, and color in a flat catalog database. Vector search processes multi-dimensional product models for matches that are based on deeper, more meaningful semantic relationships that are not keyword dependent.
Vector-Based Search Today
While some providers are attempting to incorporate vector-based search into their existing text-based search platforms, they often face challenges inherent in integrating new technology into systems that were originally developed with “old tech” ten to twenty years ago. Conversely, new solutions built from the ground up on generative AI technology are emerging in the market. These solutions offer quicker implementation times and better ROI, as they are not constrained by existing text-based search technologies.
Whether you’re already exploring vector-based search or just learning about it for the first time, it’s important to understand its potential impact. Much like how Starbucks transformed the coffee shop landscape in the 1990s, vector-based search is poised to revolutionize product search in the near future. New solutions will offer superior results at lower costs, making text-based search methods likely to become obsolete.
All this talk of java and search has made me thirsty. I think it’s time for a flat white or a café macchiato at my neighborhood Starbucks!