Critical update: Improve performance robustness for Case Deflection and Insight Panels

This is for:

System Administrator

In a typical Coveo intelligent service experience, Case Deflection panels and insight panels use the Intelligent Term Detection (ITD) feature to extract meaningful keywords from large textual case descriptions.

In this scenario, the case description is set in the large query expression (lq), and the search request is routed to a query pipeline that has an ITD-enabled Automatic Relevance Tuning (ART) model association. ITD refines lq into a concise query expression which is then sent to the index.

However, if no ART model is associated with the query pipeline, the ART model isn’t ready, or ITD isn’t enabled on the ART model association, then the lq text is instead sent to the index as a partial match expression.

At the moment, the Search API sends the full lq text in that partial match expression. Since case descriptions are often quite large, this can result in poor query performance, or even request time outs. This is what this critical update will prevent.

What will happen when I apply this update?

This update introduces a new query parameter, lqPartialMatchMaxKeywords, whose default value is 100.

This parameter defines the maximum number of lq keywords that will be included in the partial match expression that gets generated when ITD can’t refine lq.

In other words, after you apply this update, the Search API will truncate the lq text to a maximum of 100 keywords by default before sending it to the index as a partial match expression.

A support agent opens a case. The case description contains about 10,000 words. Under the hood, the Coveo insight panel sets lq to the case description and sends the search request. The search request is routed to a query pipeline that has an ART model association. However, the ART model association is wrongly configured: it isn’t ITD-enabled. Consequently, the Search API sends the lq text to the index in the form of a partial match expression.

Following this critical update, lqPartialMatchMaxKeywords is set to 100 by default. Therefore, instead of sending all 10,000 lq keywords in the partial match expression as it would have before, the Search API truncates lq and only sends its first 100 keywords. Some context is lost in the process, but query performance is significantly improved, and the agent sees relevant results in the insight panel almost instantly.

What must I do before I apply this update?

If there are no case deflection or insight panels in your solution, you should be able to apply this update right away: it won’t affect your solution in any way unless you’re somehow using lq.

Otherwise, before applying this update, you should ensure that every case deflection and insight panel in your solution is routing its search requests to a query pipeline that has an ITD-enabled ART model association. If a panel is wrongly configured (that is, if its query pipeline doesn’t have an ART model association, or if that model association isn’t ITD-enabled), fix it before proceeding. If you need, you can contact Coveo Support for guidance.

If a panel is routing its search requests to a query pipeline whose ART model isn’t ready yet, you may want to consider setting the lqPartialMatchKeywords query parameter to a different value for that panel in order to ensure that the fallback lq partial match expression has a fair chance of yielding relevant keywords. In most cases, we recommend using the default lqPartialMatchKeywords value. However, if the most relevant terms tend to appear later in case descriptions, it may be best to use a higher value. Keep in mind that setting lqPartialMatchKeywords to a high value is more likely to result in poor query performance, though.