Currently, I'm switching to CX from ES and I'm trying to figure out how to get the best results.
What is a problem: Dialogflow answers with LLM much more frequently than in a "regular" way.
Statistically, 6/10 responses come from LLM, 3/10 come from intents in the flow, 1/10 from a route group.
It causes obvious problems: I can't use a route group to transition to other flow if question is not from current flow and I can't design the entire response process exactly as I would like, because LLM constantly breaks into conversation.
The irony is that the knowledge base that powers the LLM is identical to phrases and responses in intents that are in flows and route groups... Settings are: 0.6 threshold for intents and very high score for grounding (LLM).
- Data store: backup of the ES bot.
- Structure: unstructured data that consists of all intent names, training phrases and responses.
- LLM works on it very well. - Intents and entities - migrated via Google's tool (Go code / GC SDK Shell)
- Intents in flows - I added every intent from ES and just made a example answer to cut time
- Intents in route groups - agent level, as above.
Have you ever had similar problem? Or maybe you have any ideas how to deal with it without creating webhooks?
LinkedIn
Twitter
