I had a question on how to further fine tune your generative fallback? In the generative fallback prompt I have, seen below, It is still answering questions where I wish it said: "I can only answer questions about company X."
For example, I asked the chat bot, which uses a datastore for majority of its answers w/a few deterministic routes for specific requests (live agent, etc.), "Whats you though on company Y?" and it says "Company Y is ...." Company Y is a competitor to our company. How do I edit the fallback prompt to only answer questions related to Company X?
To Fix this do I:
a. Alter Fallback prompt? If so where/how?
b. Alter banned phrases?
c. Create more deterministic routes? If so how many, as this would defeat some purpose of Gen AI chat bot
d. Edit route descriptions?
Gen Fallback Prompt:
You are a helpful assistant agent and your mission is to be the trusted partner for resolving all issues related to the company X. You provide technical and non-technical expertise, but can only answer questions if they are related to the company X.
In particular, you can $route-descriptions. You will try your best to see if you can resolve the user's issue or queries without directing the user to outside support at https.X/customer-support/.com.
The conversation between the human and you so far was:
${conversation USER:"Human:" AGENT:"AI"}
Then, the human asked:
$last-user-utterance
You say:
Solved! Go to Solution.
To be honest with you, your approach looks great. You did exactly what I would do. On June 15th, the default LLM for generative fallback will change: https://cloud.google.com/dialogflow/docs/release-notes
Another suggestion will be to call a webhook and try with Gemini Pro or Open AI models.
Best,
Xavi
I usually put everything in a prompt so whenever I want to change to a different model and calling it from a webhook, I can re-use the prompt
To be honest with you, your approach looks great. You did exactly what I would do. On June 15th, the default LLM for generative fallback will change: https://cloud.google.com/dialogflow/docs/release-notes
Another suggestion will be to call a webhook and try with Gemini Pro or Open AI models.
Best,
Xavi
Follow up question on banned phrases:
When I mention a banned phrase into the chat bot, I get a generic no-match response of: "Sorry could you say that again?" How do I change this response?
that is because the no-match is probably being triggered, can you confirm this?
You are correct, the no-match-default is triggered when I mentioned a banned phrase. For every no-match-default I have turned on Gen Fallback. It's only when a banned phrase is hit, it seems to by pass the Generative Fallback and goes straight to the predefined no-match agent responses 'Sorry I didn't get, can you re-phrase that'?
Also another follow up question. I have the same Gen Fallback prompt stated above. It clearly states in the first paragraph, 'Only answer questions related to company X'. However sometimes when I ask questions like 'why is the ocean blue?' or 'why is the grass green?' it gives me a legit response vs. what I want 'Sorry I can only answer questions related to company X?'
that is a common issue with the gen ai solutions. The only solution to that is basically to try different prompts until you get one working as expected. that is call prompt injection
I can see that. What about the banned phrases and the no-match being triggered but the Gen AI no match fallback is not. Is that something you have seen?
I usually put everything in a prompt so whenever I want to change to a different model and calling it from a webhook, I can re-use the prompt
User | Count |
---|---|
13 | |
2 | |
1 | |
1 | |
1 |