I am losing my mind trying to migrate our knowledge base from PureConnect to Genesys Cloud. We are using the new Knowledge Optimizer with a Dialog Engine Bot Flow to handle basic FAQs before routing to an agent. In PureConnect, we had simple exact match keywords. In Genesys Cloud, the NLU engine is constantly returning false positives for our intents! I imported 500 articles, but when I test the bot, it routinely pulls up an article about password resets when the user explicitly asks about hardware replacement. The intent confidence scoring makes absolutely no sense. Does anyone know how to force the bot to strictly require specific keyword entities before triggering a knowledge article search? I cannot afford to have this bot confusing our users during this massive migration.
Hello. I understand this can be very difficult. When I create training materials for the Knowledge Workbench, I always emphasize that the search algorithm in Genesys Cloud works differently than PureConnect.
It uses semantic search, not just keyword matching. If you want to force specific words, you must create an Intent in your Bot Flow first.
Inside the Intent, you define your Slots for the hardware types. Then, you only call the Knowledge Search action if those specific Slots are filled.
Do not rely on the global knowledge fallback alone.
Greetings! I completely agree with the previous advice! It is so exciting to see people leveraging the Knowledge Optimizer, but it definitely requires a mindset shift! One phenomenal approach we took at our German headquarters was to heavily utilize the search feedback loop. The analytics endpoints, specifically /api/v2/knowledge/knowledgebases/{knowledgeBaseId}/search, provide incredible insights into what the engine thought was relevant. By aggressively reviewing the ‘Unanswered’ and ‘Negative Feedback’ queries in the performance views, we were able to add specific negative intents! It is a brilliant way to train the NLU to ignore certain overlapping phrases while keeping our GDPR compliance intact!
Yeah, the transition from on-prem to cloud NLU is rough. We run a multi-org setup and I literally had to script an export of our prod knowledge base to our staging environment just to safely test the tuning. The biggest mistake you can make is dumping all 500 articles in at once.
Start with your top 20 intents. Also, make sure you check your phrase variations.
If your password reset intent has 50 training phrases and your hardware intent only has five, the engine is going to heavily bias towards the password reset. You need balanced training data.