Natural Language Search recommended search suggestions

The Knowledge and Library Hub is designed to offer a supportive tool for simple evidence searches.

The option to search by expressing a query as narrative text was added as part of the introduction of a new interface. Initially called Natural Language Search (NLS) it will be rebranded AI-Assisted Search in late June 2026.

The search query is interpreted by an LLM and translated into a structured query using Boolean logic to combine concepts. This "refined query" can be viewed providing a means for users to consider if it matches their expectation and some transparency. Our testing at the time showed that NLS was considered effective at providing relevant results, read the Generative Artificial Intelligence in practice: EBSCO's Natural Language Search blog post.

A new feature has been created to extend the Natural Language Search tool. After the initial search it offers three variations on the query prompt for the user to potentially try (see EBSCO Connect Recommended Search Suggestions, frequently asked questions (FAQ) for further details). The stated intent is that the suggestions are "designed to help you either dig deeper into a topic or uncover new but related directions for exploration".

Evaluating recommended search suggestions

We carried out an evaluation of the revised NLS tool. A volunteer group of testers used a standard protocol to work through a query and the resulting recommended searches. They could use either their own idea or suggested questions drawn from a sample shared by a clinical query answering service.

Questions considered

• Is the initial time to generate an answer acceptable?
• Is the initial refined query helpful and accurate to the intent of the question?
• How many of the first 20 results seem relevant to a simple skim check?
• What kind of results are they?  Articles / books / other?  Current or older?
• Are the three search suggestions – rephrases, extended questions, something else?
• Similar questions asked for each of the three suggested queries
• Using one of the suggested queries as a starting point click through into one of the suggestions on that query and answer similar questions

General questions

• What is the impact of the presence of the search suggestions on the user experience?
• Repeating your initial search on your own Hub is the time for the results to show quicker, slower or similar?
• How would you assess the value of the Suggested Searches feature?
• How might it be improved?
• Would you support it being turned on?

Results of the evaluation

6 responses in total were received with all the respondents completing a full test.

Across this small sample the testers were generally quite pleased with the suggestions offered.

Testing reconfirmed that the basic NLS works well. A few options to improve search suggestions as a feature were highlighted and passed to EBSCO. These included making the recommendations be on request rather than always appearing, offering alternative refined searches instead of new search statements that would allow you to more readily assess the impact of that option and not allowing endless new options.  

Useful feedback was also gathered on guiding the translation more generally seeking to include key elements such as alternate drug names and outlining the benefit of using more concepts when in doubt to help increase relevance.

The recommended search suggestions tended to go in one of two directions. With "easy" questions that initially generated highly relevant results any subsequent iterations ended up offering more or less the same search and results. With questions where the intent was less clear (for example in terms of likely concepts) the search suggestions did sometimes offer an alternative that generated more relevant results. The challenge is knowing which of the statements would achieve that.

Implementation decision

While testers were supportive of the value of the suggestions they were more equivocal about whether it should be turned on in practice. Concerns related mostly to the potential to get lost in a rabbit hole and on the impact on speed of response. The decision was taken at the Change Advisory Board to go forward with roll out (completed 8 May 2026) and accompany it with a wider evaluation of the impact. We look forward to learning more from the live implementation.