by 1. What are Generative AI Search Engines?
Search engines have dramatically advanced in recent years. The internet has become ubiquitous and the amount of information available online is growing at an astonishing rate. Given the sheer volume of information available for any given topic, search has become an increasingly important tool for obtaining relevant results to user queries. Unfortunately, traditional search engines often return too many results, or too few. On top of this, the results are disorganized and do not do a good job of answering the users’ queries. This leads to a lengthy and often frustrating search process where a user must click through a multitude of links to find the most relevant information.
With all of the recent development of so-called “web 2.0” technologies, in some aspects search engines are falling behind. One approach to rectifying this problem is by providing conversational responses or question-answering type search. This is ideal as it would eliminate the need for the user to sort through many results and would just directly answer their question. Unfortunately, this is extremely difficult as most search results are documents and a large portion of those documents do not directly answer the question. The search engine must then display some sort of abridged summary for the user to trust that it has indeed answered their question. Step in generative AI search engines. These have the potential to completely change the way people search for information, moving away from the traditional keyword and document result type search. In this paper, we will discuss the future of generative AI search engines and how they compare to normal search engines.
2. Current Advancements AI Search Engines
This concept has led to the creation of abstractive summarization tools, which is a key feature in the newer AI generative engines like perplexity, or google Gemini. These tools are designed to take a larger amount of information and considerably shorten the text, maintaining only the most important points and inferences. An example of this is a longer article being summarized into a few points or a bulleted list. This type of technology is a less direct approach than traditional search engine and its systems in that it is generating new documents to resolve queries as opposed to finding and directing users to pre-existing information.
In the modern world, the transformations in internet search engines are not only dependent upon the amount of data that is indexed or retrieved, but also on the complex combinations of techniques, algorithms, and AI that compromise the system. This AI-driven approach has involved methods such as machine learning and natural language processing to near-human understandings in resolving user queries, and thus lead us to the conception of generative AI systems. These systems are capable of creating brand new content, in this case creating documents and information that best resolve the user’s query as opposed to matching it to pre-existing data.
Comparing how search functions on the internet has changed since its inception brings clarity into the speed of which technology advances. Initially, search engines were primarily focused on matching users’ queries to relevant documents, using methods such as indexing and retrieval algorithms. While these systems remain the core focus of modern search engines, advancements in the technology surrounding the AI that is used to process the user queries and information have seen the exciting development of generative AI search engines.
3. Benefits and Limitations of Generative AI Search Engines
At each step in the action sequence, there is the potential to use a deterministic or probabilistic algorithm to make decisions based on the current state and the desired end result. The end result of this process is an adversarial search in which the actions of the generative algorithm must be “good” in the sense that they produce high-quality document matches, and the opposing force is the potential loss function on the query and document set. This is an improvement upon current search engine technology, which provides only a loosely related list of documents often with no organization.
To understand why this is the case, suppose that generative algorithms are used to construct an intelligent and well-organized storage system for a collection of documents. An attribute of this system might be a mapping from a high-level representation of a document (such as a set of keywords or a brief summary) to the document itself. Then, given a query, a generative algorithm can plan and execute an action in the system to retrieve documents pertaining to the query. This involves generating any variety of actions such as a language model generating text for a natural language query, to an entailment model translating the query and document representation to a set of logical inferences, and then back to text using a possibly different document representation.
Benefits and Limitations of Generative AI Search Engines: Flexible information retrieval: Given a query, generative algorithms have the potential to allow the user to construct and obtain documents that are more tailored to the user’s needs compared to the current methods of retrieving documents of a pre-specified type, often determined by keyword matching.
Generative algorithms have long been used, but have only recently gained broad attention due to the success of deep learning methods. Specifically, deep learning methods perform well in machine translation, image-to-text problems, and recently, text-to-text problems. Generative methods have not yet been widely incorporated into search engine technology, but they have the potential to provide a more flexible means of information retrieval compared to traditional search engines. Some of the potential benefits and problems of generative search engines are outlined here.
4. Comparison between Generative AI Search Engines and Normal Search Engines
Generative AI search engines (GSE) effectively work by assessing a question relative to the accessible documents in the internet database, the search queries to these documents, and then compiling an abstract response. The database is generally a collection of documents with an associated multi-dimensional retrieval function. The textual entailment and question are forms of search queries. The multi-dimensional retrieval function determines the relevance of a document to a given query and then the relevance of specific pieces of information in that document to the question. This allows selective access to certain pieces of information in various documents, which can be used to generate an abstract response to a question. GSEs describe this as question-based answer retrieval and are still exploring methods to integrate it with their best way of generating sentences, which is based on manually engineered templates. A normal search engine simply returns a list of documents and their URLs that are pertinent to that particular query. AI methods can be utilized on this information to provide a more intelligent user interface by clustering documents, filtering information from them, and providing a concise summary of the search results. However, this is still far from question answering, and entailment methods do not attempt to provide responses to queries in the form of an abstract using information from across multiple documents.
5. The Future Implications of Generative AI Search Engines
In recent years, generative searching has become a field of interest for many AI researchers. The implications of GAI search engines are widespread, promising deep effects on internet users and content providers, as well as individual organizations. In the future, normal search engines may disappear or spawn GAI interfaces. The latter is likely as GAI will sometimes require directive information to commence queries. This can be seen as an evolution on the sitemap. It is expected that computer graphics representations such as Mind Maps that are used to convey relationships between search terms will become the standard in giving GAI search engines starting points. GAI search engines may cause data providers some confusion. Due to the fact that synthesis engines will store results of user queries, it may be hard to determine exactly how a piece of data was retrieved. This can be bypassed by GAI with a “link” feature or a record of the search terms used to find the data. In the case of linking, search engines will be able to use natural language processing techniques to discern relationships between data and present them as simple “clickable” solutions for users. This may be positive for internet users and content providers as links will act as search directions for obtaining high quality information.
6. Conclusion
Generative models in their current state are very transformational in the domain of search engines, and learning a better method for ranking search results has become a central domain in the machine learning community. Traditional search engines today are moving closer towards methods that involve machine learned ranking of some kind. PageRank, for example, can be formulated as an eigenvalue problem of a matrix, with the stationary vector giving the desired rank probabilities. Learning to rank methods often involve features related to the query and the document in some way, and a function that approximates relevance.
With the current state of technology, we would not be far off in saying that many search engines could be formulated as an MDP, where the user gets satisfaction and must ultimately decide when to stop the search process. Although the specific models are yet to be known, planning and reinforcement learning methods would be a likely way to progress forward. We have seen throughout this paper that there are very many diverse ways for generative models to assist in the search process. Whether it be through query suggestions or simulating human behavior to enable better understanding of how to best satisfy the user. It is a complex issue, but one that has potential to be highly optimized with generative and learning methods. Conversely, one might argue that advances in generative models and AI methods could render search engines obsolete, through making the correct information come to the user, with or without realization.
This is a speculative issue at large and there still may be many who prefer to actively seek information. However, it is clear that search engines have become an integral daily tool for people, and current generative methods are certainly assisting in improving that tool. Given the rate of advancement of technology in recent times, it is entirely possible that the future landscape of search engines could be something very different from what we now know. Information retrieval research is going to be a key factor in shaping this, and the importance of search to our everyday lives will likely ensure that progress is maintained. A key notion to take from this is thus: whether it be abstraction, inference, language understanding, or any other AI methodology, there is clear potential for research in applying these methods to search engines.
