In recent years, vector embeddings have emerged as a promising technology to power information retrieval projects. Vectors use mathematics to represent words and encapsulate meaning. Combined with deep learning algorithms, you can build powerful AI search applications. However, recent advances have shown that another kind of technology – neural hashes – can outperform vectors while avoiding the tradeoffs that people experience with vector search.
Neural hashing’s superiority over vectors will soon make it accessible, affordable and the primary method for delivering AI-powered search for e-commerce, enterprise search and other areas where search data is constantly growing and changing. Even as early as 2023, neural hashing will be the primary method for delivering AI-powered search.
First Came Keyword Search
Most of us are familiar with the basics of keyword search technology, which most search engines still use. When someone submits a query, the keyword search engine will perform a simple lookup and display results by relevance. Keyword algorithms employ techniques such as natural language processing (NLP), lemmatization, typo tolerance, synonym libraries and more to perform the lookups.
Keyword search engines work well for the most part but can fall short when it comes to long-tail queries, i.e., those longer and more specific queries which represent half or more of search queries. Despite its powerful keyword technology, you can never write enough rules or stuff your site with enough keywords to cover every edge case. Being able to improve long-tail query results, however, can be a huge benefit to an organization, which is why more companies are beginning to look for new technologies to complement their keyword search engines.
Then Came Semantics
Semantic, or AI-powered search, can understand concepts. Semantic search goes beyond matching to understanding. For example, if you searched “something for a headache” on a website, a keyword search engine would look for items tagged with those specific words, turning up results only if they can match some or all of the text. But a semantic search engine would understand the concepts being described and offer solutions—such as pain medications.
Semantic search has traditionally been employed using vector embeddings. You can use vectors for all kinds of objects, too—images, videos, audio, etc. Vector-powered search engines can be incredibly smart, but processing and maintenance make them too expensive to run in production for most enterprise use cases. This is true, especially for certain industries where data is dynamic. In retail and e-commerce, for example, that inability to pivot and scale quickly is a real detriment, as brands need to update data that is unpredictable (e.g., user-generated content) and constantly in flux (many online sellers update their search databases hourly).
The solution is neural hashing, which refers to the process of compressing vectors with neural networks (deep learning algorithms which mimic the human brain) into hashes or binary vectors. These hashes retain 96% or more of the vector information while using a fraction of the storage space. A semantic search engine built using neural hashing is nearly or just as accurate as vector search but about 500 times faster.
Better Results With Hybrid Search
When it comes to AI-powered search, neural hashing will eliminate the tradeoffs between speed, accuracy and costs. It can deliver results in single-digit milliseconds with no additional overhead. As good as it is, it’s even better when combined with a keyword search. We can use hybrid search—both technologies—simultaneously; most queries are a combination of keywords and concepts.
For example, what if rather than looking for a type of product, a shopper is searching for help with getting rid of a pesky symptom? Or, has a question-style query? As search data indicates, people are often myopically focused on a problem that they desperately want to solve, such as:
● Leaky faucet
● Stress headache
● Stained carpet
They only know they have an urgent problem. Of course, they’re looking for a solution, but they typically have no idea what particular product they need in order to get relief. They’re still using some keywords, however, like “faucet,” “stained” or “carpet.”
Each query can be scored by both the hashing and keyword engines to deliver a superior result for every query. Hybrid search can be combined with traditional search engine features like natural language processing (NLP), autocomplete, learning-to-rank algorithms, filters and more to optimize the end-user experience.
When it comes to search, speedy, accurate results are the name of the game. In order to make searches as intuitive, fast, and scalable as possible while still remaining cost-effective, search and discovery platforms should explore what hybrid search powered by neural hashing has to offer.
It has been shown that neural hashing can make highly intelligent AI fast and cost-effective at scale and without the need to bring on highly specialized engineers for every project. We appear to be at another turning point in the world of artificial intelligence, and neural hashing is going to be the driving force behind many future innovations.