It’s been a while since a revolutionary new technology gripped the world’s attention so swiftly and powerfully. But on November 30, 2022, everyone leaned in a little closer. That’s the day OpenAI released ChatGPT to the world.
With 100M active users by January, the site set a record as the fastest growing user base of any application on the Internet. While most technology stories rarely occupy the news cycle for more than a few days, months later, ChatGPT still dominates the headlines.
By now, everyone’s had the opportunity to experiment with the tool and ponder its inner workings. They’ve seen for themselves the remarkably fluid, sophisticated, and sometimes silly responses it produces from simple queries. The result is a torrent of ideas about what this advancement means, and where organizations can take it.
This eBook digs into the quantum leap in Artificial Intelligence (AI) and Natural Language Understanding (NLU) that platforms like ChatGPT represent, and the impact of major trends stemming from these developments on search technology.
It contends that performant search as we know it is no longer table stakes. Powered by the latest AI and especially NLU innovations, people and businesses will be able to do and discover things and information more naturally, faster, and with more accuracy than ever before, and with remarkable computational ease.
These advances are a game-changer and an essential component in a winning digital strategy. They hold significant implications – not only for site search going forward, but for how we browse, search, consume information and shop online.
How we got here with natural language understanding
With advances like ChatGPT, computers seem to understand human language and produce it fluidly. It's not magic. Underlying technologies called Large Language Models and transformers are doing the work.
To grasp the significance of the recent advances in AI-powered natural language and search applications, we need to pop the hood and see how they work.
Gartner defines AI as the application of “advanced analysis and logic-based techniques, including machine learning, to interpret events, support and automate decisions, and take actions.” Not every AI project involves natural language, of course. Image recognition and robotics, for instance, are also AI.
But natural language is the human go-to for communication. It’s our primary interface – with machines and otherwise. Nearly all our digital interactions happen in natural language. At the same time, natural language is hard for computers to process and understand. Words have multiple meanings and those meanings shift depending on context. Words like “it,” “that,” and “this” are constantly in flux and can refer to many different things over the span of a few short phrases.
For humans, a phrase like “Get that?” at the end of a paragraph like this one is easy to grasp. Computers have a much harder time solving for this kind of ambiguity. They have trouble filling in the blanks, reading between the lines, and making inferences.
The subfield of AI that works on this problem is called Natural Language Processing. The subfield that focuses on how computers understand human generated text is called Natural Language Understanding (NLU). The joint fields of NLP/NLU have been working on these problems for decades – making impressive forward strides, then always hitting a wall.
Computers now appear to genuinely understand natural human text and can generate responses that seem entirely fluid and comprehensible, even thoughtful. The joint fields of NLP/NLU have just plowed through their biggest wall yet. The key to that success is machine learning and Large Language Models (LLMs).
From keyword matching to LLMs
The most basic way for computers to “understand” – as in, process – natural language is through keyword matching. In this simplistic model, one word equates to another single word. In fact, some of the first AI models in the 1970s tried to reverse engineer human language and program dictionary definitions for every potential text input. Unfortunately, that method couldn’t take NLU very far.
In the meantime, however, keyword matching succeeded as the primary method used for search.
It made perfect sense: people tend to use concrete words, like names, topics, and titles, to find exact or relevant matches in indexes. There was no real expectation that people would, should, or could interact with computers using the same kind of fuzzy language they use with other people.
Today, however, AI-powered NLU uses an entirely different approach to processing language. Instead of capturing one-to-one equivalents between words, computer scientists created a model that simulated the learning in a human brain. They built neural networks and used deep learning to train algorithms on the lexical patterns that exist in natural language. The result? Context-sensitive Large Language Models (LLMs), trained on massive corpuses of data and comprising billions of parameters to derive complex mappings between words.
LLMs are the engine that run today’s language-based AI applications, including OpenAI’s ChatGPT. It uses a specific LLM called GPT, short for “Generative Pretrained Transformer.”
Leaping ahead with transformers
Not every LLM has the same power and sophistication. They don’t all undergo the same depth of training on the same amount of text data. They don’t all involve the same number of carefully adjusted and weighted parameters. They don’t all use the same underlying algorithms.
Today’s largest and most powerful LLMs rely on transformers, a neural network architecture developed by Google researchers in 2017. Transformers revolutionized NLP/NLU by harnessing more computational power to process semantic relationships across much longer sequences of text. They focus “attention” on important pieces of text to solve problems like what a word like “it” means when it’s used to refer to something identified in a much earlier portion of text.
The LLM processes input text and responds in real time by making astoundingly accurate “guesses” about what the next word is likely to be. It doesn’t actually “understand” the meanings of input text, but performs complex statistical calculations based on petabytes of training examples to assess likely relationships between words in this specific case.
As Stephen Wolfram explains, it asks one question over and over again: “given the text so far, what should the next word be?” Based on probabilities derived from the underlying language model, the algorithm inserts the most likely next word, and so on, and so on, and so on.
Today’s LLMs can now determine context and relevance from larger swaths of natural language, and generate accurate and natural answers in human-like conversational output. Transformers made it more efficient and faster to process language input in parallel, making LLMs easier to integrate and apply to a range of language-related tasks, like search.
As Sean Mullaney, CTO of Algolia, told Venture Beat:
“LLMs are fundamentally changing the way search algorithms work. Traditional search engines match individual words from a query with the words in a large index of content, he said, but LLMs effectively understand the meaning of words, and can retrieve more relevant content.”
These processing advances made an exponential difference to NLU. The more text data a model is trained on, the more semantic relationships it has access to, and the more accurate are the lexical patterns it can process, apply, and reproduce.
Vectors and the rise of intent intelligence
An innovation called vector search is a huge advance over keyword matching. It lets computers map the semantic relationships between words using statistics and probabilities. By making connections between words, it understands what users mean and captures valuable "intent intelligence."
Say goodbye to the "null" search. Even if a word doesn't exist in an index, computers can now draw on their underlying language models to provide a suitable response.
The development of transformers led Google and AI to another critical innovation: vector search technology, a method for connecting text inputs with objects in an index to vastly improve information retrieval in ways that mirror human understanding.
It’s easy to see how vectors are key to helping computers understand natural language. For example, how would a computer know that a computer user who types “espresso with milk thingy” means an appliance, and not coffee or grocery items? Vectors help by plotting and clustering word tokens in multiple dimensions (n-dimensional space) representing any number of variable attributes. When search queries are made, they’re converted into vectors and compared to other vector representations by computing the distance between a range of attributes. The closer the match, the more accurate the result.
One of the biggest challenges in traditional, keyword-based search is still the null search: a query that returns nothing at all. No one wants to ask a question – to a person or a website – and get no answer in return.
Imagine this scenario: a shopper visits a retail pharmacy site and enters “get clean without drying my skin” into the search bar. While human readers get a general sense of what the shopper is looking for, it’s challenging for a computer to parse. Before LLMs, transformers, and vectors, the computer would only process the entry as keywords – “clean,” “skin,” and “dry” – and would not offer the shopper the desired moisturizing soaps. Now, the lexical patterning enabled by LLMs lets computers make the right semantic relationships and glean “body wash,” even from a free-flowing and rather inexact search query.
As AI surges ahead with LLMs, transformer architecture, and vectors, it’s developing a library of lexical understanding and patterning that’s revolutionizing the way people interact with computer technology.
Users used to enter fixed or narrow terms to return concrete results. Now, algorithms know so much about the semantics of natural language that they can pinpoint exactly what users are seeking even from remarkably varied and fuzzy input. Users don’t need to work as hard to define what they’re after. Computers understand their intentions and meet them halfway – much like human partners in conversation do. The result is a watershed moment in AI, as computers now capture intent intelligence.
The limits of vectors for anyone but the giants
While vectors have made huge waves as a technological development in the field of NLP/NLU for the past few years, they haven't taken off commercially. Vector models require enormous amounts of training, extensive (disk) storage space, and heavy computation performed by high-priced GPU processors. These factors make vectors immensely expensive for most companies to deploy at scale.
Vectors are expensive for most companies to deploy at scale.
LLMs from tech-giants like Microsoft, Google, and OpenAI are designed to work on the open web. Smaller companies would love to duplicate those successes at a scale that suits them, but it’s expensive and time-consuming to build and train models. It's an especially big challenge to develop vector search with an in-house dev team. It took Home Depot 13 months and a group of highly-trained data scientists, for example, to build a homegrown, high-quality vector search engine. Not every company has the same resources, time, and budget.
Even so, vectors have ripped open Pandora's box. All kinds of enterprises are excited about the potential for intent intelligence and finding ways to harness it.
Intent intelligence with end-to-end NLU
In the meantime, everyday users are running wild with their explorations. Writers, from poets to screenwriters, are using LLMs like ChatGPT to brainstorm ideas, plot stories, and gather new expressions. Students and journalists are using them for research and drafting. People are using them to plan vacations, get shopping recommendations, and write their CVs. Some are even using them to write code.
For many, the remarkably human-seeming persona behind the language models is the biggest draw. They're curious what it has to say. Others just want to have a chat, or use the interface as a sounding board for their thoughts and emotions.
There are assistive applications as well. LLMs can improve Internet navigation for people with communication-related challenges, by helping compose emails and generating instant, accurate video and image descriptions.
For websites, there's an immense potential to improve customer support. Enabled by LLMs and intent intelligence, chatbots and virtual assistants will be able to grasp exactly what site visitors are looking for – more quickly and with more precision than ever before. The potential for site search is truly out of this world.
Applying AI to search and discovery
Search is the first and best place AI can be adopted in any organization. Intent intelligence will transform the search bar into a conversational tool, but for search to be effective, it needs to be optimized as search.
It’s no surprise that these mind-boggling, natural-seeming computer advancements are now being applied first and foremost to search.
When people sit down to explore an LLM like ChatGPT, the field of possibility is a blank query field. They enter prompts and see what comes back. They ask questions. They find information. They uncover new things. People engage the technology in much the same way they do the well-known search engines, like Microsoft’s Bing, Alphabet’s Google, and AI-powered tools, such as Algolia's NeuralSearch.
Looking ahead, it's clear intent intelligence will transform the search bar into a conversational tool. With search and discovery now the bleeding edge of AI, the internet giants are in a race to deliver the best and most natural results. One Google executive called it a “make or break” moment for the company’s future.
Microsoft made a $10B investment in OpenAI’s ChatGPT to run its Bing search engine on a search-specific language model for more speed and better accuracy. Google recently launched a conversational AI called Bard, run on LaMDA, a language model trained on dialogue. Chinese search giant Baidu is reported to be developing a service modeled on ChatGPT to be integrated with its search engine. Amazon continually invests in search AI, and is on record attributing $800M in monthly incremental revenue to its search function.
Endless potential if done right
While search is the first and best place AI can be adopted in any organization, LLMs in their open-ended present form are not the perfect mechanisms for site search. They can make mistakes since they're drawing on a vast corpus of publicly accessible information. A company's product catalog or index won't form part of its training database. For enterprises that need precision, that won't cut it.
LLMs respond based on entered prompts, but answers returned are not optimized like top quality search is. They can't rank recommendations or personalize results based on a user's previous searches, location, or preferences. It can't make connections between different types of goods or services or make product recommendations of the "People who bought this also bought" variety.
Search tools still need to be developed specifically for search, but next generation solutions will shift away from keyword matching to run on intent intelligence. Think "search on steroids." This advance will bring significant front office enhancements and integrate well with new, hybrid retail models, such as Buy Online, Pickup In Store (BOPIS).
Imagine a jogger who uses a search-optimized LLM to find something to drink after a long run. "I need to rehydrate after running 10K," they might speak or type into a search bar. Next-gen search will know that this individual is looking for sports drinks without the need to specify a brand. The tool will deliver a range of locally available suggestions that instantly serve the runner's needs.
Future site visitors will expect websites to grasp their intent and deliver seamless and accurate results whatever the topic or domain. They’ll expect the same search experience they get from the search giants and heavy-hitters like Amazon, Instagram, and Netflix across all search use-cases.
To keep pace with the competition and get closer to Amazon’s 15% customer search conversion rate, organizations need to get their websites up to speed. Search must evolve to do more, better and faster, and be powered by AI and real-time data to leverage the possibilities and address common search scenarios that follow:
- Search challenge: A retailer switches out products frequently to reflect seasonal and other trends. Merchandising teams don’t know what terms users will use to discover these products. Generating synonym lists is painstaking and updating search indexes slows down the process, potentially preventing sales.
- AI-powered solution: Search captures search intent automatically – even from long, creative, or unusual descriptive entries that don’t represent a perfect match between product names or product categories. Merchandisers save time and customers find exactly what they’re looking for.
- Search challenge: A parts distributor sells parts for all kinds of equipment whose product names are constantly changing. When users search by SKU, they turn up exact matches, but when they use an outdated product name or description, they don’t find the parts they need.
- AI-powered solution: Users describe products in natural language, the search platform understands the intent and returns the full range of necessary parts associated with that product.
Doubling down on AI
It’s clear to many business leaders that advanced AI is defining the way forward in business operations, product design, customer service, and more. According to Gartner’s 2022 CIO and Technology Executives Survey, nearly every respondent (96%) said that AI was either already part of their deployment pipelines or that they had initiated AI projects. Among those, 27% had deployed AI techniques (Gartner).
A recent McKinsey study showed that firms making the biggest investments in AI were pulling ahead of their competition. With LLMs and conversational language, internet leaders already have a giant stake. Their recent investments in AI make it clear that end-to-end natural language understanding is the new search differentiator to drive revenue and growth.
Where do we go from here
Neural hashing puts the latest AI technology in the hands of any enterprise at a reasonable cost.
The challenge for regular organizations wanting to build out a great search and discovery experience is that while intent intelligence is within reach, it can be in many instances, cost-prohibitive to leverage.
Algolia has developed a new innovation that solves the processing and storage challenges of vectors, while achieving the same results. The proprietary technology, called NeuralHashing, is emerging as a significant advancement over traditional vectors. It’s a machine learning technique that simplifies the vector problem from a computational perspective, delivering more than 90% of the relevance at 1/10th of the storage and processing cost.
Instead of sending vectors to a complex and costly vector index, NeuralHashing converts and stores them in an efficient binary format – called a hash – and discards the original vectors.
Neural hashes are music to any CTO's ears. They're easier to compute and more efficient to store, letting businesses leverage high-quality intent intelligence at lightning speed while saving computational resources and costs. Resources aren’t the only concern. For organizations to leverage AI-enabled intent intelligence to maximum effect, they still need business and merchandiser control. They need a solution they can own.
It’s now within reach for organizations to emulate the attention-grabbing conversational experience of LLM-enabled technology as a critical piece of enterprise architecture. Forming part of Algolia's plug 'n' play search platform, hashes can now help enterprises of all kinds create and deliver an outstanding search experience.
Driving KPIs with AI-powered search
Data is the fuel that drives AI algorithms, including AI-powered search. Most organizations already capture event data and click patterns from user interactions – what they previously searched for, what they clicked, what they added to their cart, what they bought. All those valuable data streams can be leveraged to improve AI-powered platforms and further enhance the overall search experience.
A search query like “crystal doorknobs that look good in a Victorian home” might confuse a search engine that is limited to keyword matching.
AI-powered search understands more from the outset. Customer interactions continuously improve the AI model, giving future users better results and enhancing the site experience. Automation avoids wasting cycles setting up complex language rules and ontologies. It also brings the back-end and front-end together to amplify business results and execute merchandising plays more effectively.
AI-powered search can dramatically improve site metrics with or without data scientists to build and manage the effort. With solutions like these on the horizon, organizations can finally get the most effective, comprehensive search solution – at scale, without breaking the bank.
Evolving with AI
Given the current pressures on the global economy, even leading organizations are having a tougher time meeting KPIs and ensuring the viability and success of their business. That’s true for e-commerce stores as well as brick and mortar shops.
Despite this, forecasts estimate major growth in spending on systems powered by AI this year. IDC estimates investments will reach $154B in 2023, an increase of 26.9% over 2022. Given those estimates, there’s an opportunity cost to not aligning with these trends and pushing forward with key innovations.
Given where we are and where we're headed, critical questions to consider include:
- Is my organization equipped to survive and thrive in an AI-driven world?
- How can my organization best harness the latest advances in NLP/NLU and join the conversational revolution in search?
- Do I have the resources, time, and budget to develop it in-house?
- Are there existing solutions that meet the organization’s intent intelligence needs?
- How can I leverage the best and maximize ROI?
- What solution is quickest to market and easiest to deploy?
- What critical functionalities are truly important to my business scenario?
Revolutions don’t happen every day. Ever since Alan Turing devised the first "test" for computer intelligence in 1950, NLU has always been the moonshot for AI. Now that we're seeing it come to life, it's only going to become more embedded in our lives.
Companies face a choice: evolve or go extinct. Those who choose to evolve – and evolve early – have a once-in-a-lifetime opportunity. They can harness intent intelligence and elevate their search and discovery experiences as first movers to the detriment of their competitors.
Leveraged the right way, those improvements can be implemented in efficient and cost-effective ways to gain process efficiencies and drive revenue. Companies that do nothing will fall off the radar. Those that hesitate risk falling further behind.
AI will soon underpin every single customer experience.
Is your platform ready for the conversation?
About Algolia
Algolia is the world’s most powerful, API-first search and discovery platform. It continues to revolutionize search and discovery with the world’s first hybrid search engine that combines keyword and vector-based semantic search in a single API using Neuralsearch technology. Algolia empowers builders and business users to create unique, engaging, and scalable end user experiences to predict what customers want at blazing speed. We provide the best application browse experience available on the market today.
Algolia is your guide to the world’s content and powers discovery everywhere you live, work, and play. More than 17,000+ companies rely on Algolia to manage over 1.5 trillion search queries a year, including Under Armour, Birchbox, Stripe, Slack, Medium, and Zendesk. Algolia is headquartered in San Francisco with offices in New York, Atlanta, Austin, Paris, London, Bucharest, and Sydney. To learn more, visit www.algolia.com.
%20(2).svg)
