How to Implement Topic Recommendations

[Amy Lind1]

Does anyone have experience working with the Topic Recommendations plugin for Virtual Agent? I'm curious on how best to design and implement custom topics that have been identified using the analysis. 

Since this feature requires a minimum of 10K records and sub-prod environments typically do not hold production data anyway, the process seems to defy the normal process for promoting from Dev up to Prod.  Are Implementers importing 10K records into their lower environments or is there a more efficient way that I'm overlooking?

TIA

[Victor Chen]

Hi, 

The recommended topics can also be ITSM out-of-box topics (if you've also installed/updated the ITSM Virtual Agent conversations store app/plugin) so they don't have to be custom. We do recommend the 10k limit for the best results. There's little risk in running the analysis in Prod, as it doesn't add the topics to the Virtual Agent until you do so manually. 

If you still want to run the analysis below 10k, you can modify the system property: sn_nlu_discovery.intent_discovery_min_classification_limit, to be less than 10k. Warning: YMMV and it may impact the quality of your analysis. 

[Raj64]

Hi Victor,

What performance implications are there of running this in a prod instance? Can it be scheduled?

[Amy Lind1]

Thanks Victor! It's not the analysis in Prod that is concerning me.

According to the documentation, if it is determined that a custom topic is could be beneficial, you have the ability to enable that topic as a shell flow, but it also does all the heavy-lifting in terms of NLU, which I want to take advantage of. Typically, we'd create the Topic in sub-Prod, customize the flow and then promote to Prod. However, since the flow is created in Prod (because that is where the analysis is done), what is the process to design and deploy these new custom topics?

Does that help?

[Victor Chen]

Since it is direct to Prod, you'd probably do a fair amount of building within the empty shell and then testing in Draft before publishing the topic. You may also pilot the topic to a more select group of users first (via the Topic's properties conditions) before releasing it to the general audience. 

And there isn't much performance implications for the analysis, and it can't be scheduled.