Making Search Conversational to Improve Knowledge Access

Given the seismic advancements in natural language technologies over the past couple of years, the transition to conversational search seems all but inevitable. Semantic search techniques ascertain the underlying intention of a user’s query or prompt. The need to carefully construct search queries using keywords and Boolean connectors is gone. Search results are no longer limited to documents or webpages, although they can readily include the content from those sources. Instead, search has gone multimodal.

The document-based information retrieval paradigm has been supplanted by the ability to converse with enterprise knowledge, data, and documents. Consequently, the search idiom itself has branched out to encompass more enterprise applications, for a broader range of users, than was imaginable even a few years ago. Vector similarity search mechanisms, for example, are as adept at searching through video, audio, and images as they are text.

Meeting Users Where They Are

Still, the penultimate triumph of conversational search may be that “It meets the user where they are,” commented Steven Bixby, Pega SVP of product engineering. “We’re using search as a way to meet the user using their own language, the way they’re describing the problem, as opposed to keyword-based search, where you have to get the right words in order to get the right content.”

As ideal as that characteristic is for many deployments, organizations should realize that most conversational search mechanisms are based on language models, including, but not limited to, large language models (LLMs). These models are primarily designed to accurately predict which word or phrase occurs next in a sequence. The quality of their information retrieval, syntheses, summarizations, and conversational answers, however, isn’t guaranteed.

Consequently, more established search techniques, including lexical search, what Laserfiche CTO Michael Allen termed “query-based search,” and numerous hybrid approaches not only endure, but also are often preferable to make good on the promise of conversational search.

“A lot of people are looking at these technologies as easy buttons, but they’re really not,” cautioned Mary Osborne, SAS senior product manager of NLP. “You can get fantastic results, but you get out what you put into it.”

Language Models

There are manifold advantages of semantic search via language models. Because they deliver a natural language interface to back-end content systems, they fundamentally amplify search’s capital value proposition. According to Bixby, “Search has traditionally been a tool, but now it’s sort of acting as a collaborator.” In addition to allowing users to ask and answer questions (with the latter stemming from a synthesis of multiple sources), conversational search increases the quantity and granularity of the questions users can ask. Its iterative nature is one of its chief distinctions. According to Rohan Vaidyanathan, Hyland VP of content intelligence, “Rather than starting over with each query, users can refine their questions, ask follow-ups, or explore adjacent ideas without having to reframe everything from scratch. This mirrors how people think.”

The analyses furnished by language models is another unequivocal benefit. Moreover, it’s not confined to text. “You can take a page from a business intelligence report, maybe a table or a pie chart, and say to the model, ‘Tell me about this,’” Bixby pointed out. “‘Where are the insights from this?’ That’s hugely valuable.”