How can semantic search and query help users find relevant and trustworthy information online?

Semantic search aims at using meaning in the process of searching for information. Traditional searching algorithms (such as keyword search) typically focus on returning information similar to a query in terms of words (lexical), rather than phrases. Semantic information is typically represented in ontologies which organise semantically similar words. The Web Ontology Language (OWL) is a notable example. Graph based solutions of ontologies are called Knowledge Graphs (used by Google). With the advent of LLMs, the use of word and sentence embeddings to find concepts closer to each other in a vector space has proven to work exceedingly well. Dense retrieval combined with re-ranking has now become the de-facto standard for semantic search.

Semantic search is a type of search that attempts to understand the meaning of a query, rather than simply matching keywords. This means that semantic search engines can return results that are more relevant to the user's intent, even if they do not contain the exact words that the user used in their query. Semantic search is made possible by a number of technologies, including natural language processing (NLP), machine learning, and knowledge graphs. NLP is used to understand the meaning of the words in a query, while machine learning is used to identify patterns in data and to make predictions. Knowledge graphs are used to store information about entities (people, places, things, etc.) and their relationships to each other.

Semantic search can be implemented in a variety of ways. Google has implemented Semantic Search by using a Knowledge Graph since 2012. Implementing a Knowledge Graph can be daunting task for an enterprise as extensive data modelling as well as implementation in a graph database is typically required. In recent days, semantic search with LLMs using Retrieval Augmented Generation (RAG) has revolutionised information search and retrieval. Text is embedded in vector representation and stored in vector databases. These are queried using dense retrieval and supplied to LLMs with contextual information to search for answers. The answers are re-ranked using various criteria and human readable answers are generated by the LLM from the outputs.

Suppose you type the following query into a search engine: "History of the United States." A traditional search engine would return results that contain the words "history" and "United States." However, these results might not always be relevant to your query, since there are a lot of angles to explore here. A semantic search engine, would use NLP to understand the meaning of your query. It would then use machine learning to identify patterns in data about the history of the United States and to make predictions about what information you are looking for. Finally, knowledge graphs would retrieve information about the history of the United States from a variety of sources.

In my experience if you wish to concentrate on a particular knowledge base, you might consider implementing a semantic search. The good news is, there has never been a better time to do so. With relatively little engineering effort, you can achieve impressive results using a model-as-a-service or open source model to generate your embeddings. As with most technologies, things get more challenging when you need to increase robustness and scale.

Semantic search increases the accuracy and relevance of your search results, which in turns increases the value information holds. It allows you to save time and money by finding the information you need. If implemented by an LLM, it additionaly presents you with the information in a conversational format, which is more easily digestible. When implemented on a specific knowledgebase, it has the ability to curate information to a specific need, by means of fine-tuning or engineering search prompts.

Here are some of the benefits of using semantic search: Semantic search can return results that are more relevant to your query. Semantic search can help you to discover new information that you might not have found otherwise. Semantic search can help you to save time, because you will not have to sift through as many irrelevant results. Semantic search can help you to be more productive, because you will be able to find the information you need more quickly.

One of the biggest drawbacks of semantic search is its complexity and cost of implementation. Also, determining the query intent is not a straightforward task in particular dealing in sectors where the queries tend to be very short such as in ecommerce. For example, a shopper may query women tops, which could match a significant part of the catalog. So in this case it is difficult to identify the shopper"s intent unless there is a grouping/ clustering process that helps reduce down the scope by size, colour and brand for example.

If implemented as a Knowledge Graph it could be costly and will take a considerable amount of time to develop and refine for each business use case. If implemented by an LLM one faces a variety of challenges that LLMs carry e.g. possible Copyright infringement, IP protection, operational cost, LLM hallucination etc. In addition to these it is currently not possible to ascertain what the sources of information for answers made by the Foundational Model was, which makes information reference problematic in this context.

Sematic Search is still largely text-based- multimodality (the ability to search through graphs and images for information would be a great asset). The embeddings are often context-dependent. If the usage of words changes, search results can be outdated. It's very costly to maintain the various components. Personalization acts as a double-edged sword- also pushing people through rabbit holes and echo chambers.

Grouping and clustering responses/data points can be very useful for studying topics at scale. This is very useful if you're trying to study the bias of a particular dataset/corpus

For example, searching for a book on an e-commerce site. Instead of typing in specific keywords or phrases, a user can use natural language to ask a question such as "What are the best books on artificial intelligence?" The site's semantic search engine can then use natural language processing to understand the user's query and provide relevant results based on the meaning and context of the query. The user can also provide feedback by rating or reviewing the recommended books, which can improve the accuracy and relevance of future search results. These features make it easier for users to find the information they need in a more natural and intuitive way, without having to rely on complex keyword-based queries.