My agent is providing incorrect responses, despite having combed my website, added Google Docs with detailed information, and added QA’s. What else can be done?
Hi Duncan, if you duplicate the same link in various datasets, that could be a problem. Duplicated content is not the best option. I have tested your agent and tried to retrieve the link from the Q&A dataset. I noticed a mistake, so I made a video to show you how it’s done.
1 Like
Simone, thank you so much for taking the time to create a video. I did see your most recent post on knowledge hub. I think that will be very helpful. A few items:
-
Yes, the 2 url’s I have in knowledge hub is correct. There is not duplicated content on these urls. Is this still ok to do?
-
I believe I identified a major cause to why my agent is so inaccurate some times. A common question we’re getting is, “Are you open today?” My agent is excitingly telling everyone we’re open and to come on in, no matter what time or day it is. I believe this is because the agent doesn’t know the date or time. I’ve added an Hours of Operation Google doc for all our various programs and activities. What is the best way to help when the agent doesn’t have real-time information?
-
You mentioned in the video to create just one Google Doc. I have many documents for all the different programs we offer. Is having multiple docs problematic?
-
Could you create a sample Google Doc that Answerly clients can use as a template to best organize and optimize information for an agents?
Hi Duncan, regarding point one, you should now be able to consistently provide the correct URL through Q&A. You can still have two link references with different datasets, but one would be sufficient.
Regarding point two, this issue occurs because when there is no information available, ChatGPT will make up an answer. Additionally, the agent is not aware of the current time. The best solution in this case would be to inform the agent that whenever someone asks if you are open, you can provide them with a table of the weekly opening and closing times so they can check for themselves.
For point three, it is better to have one document to keep everything organized. However, having multiple documents together is not a problem, the problem is when the content is duplicating or overlapping
For point four, I will create an example document and update my recent documentation.
Thank you for responding on all points. With regard to #2, OpenAI’s ChatGPT 4 does know what day it is, but my agent does not. If someone says, “Are you open today?” or “Are you open tomorrow?” I feel my agent should be able to provide the schedule for that particular day. Are there any adjustments that can made in Answerly so that my agent knows the current date.
The other issue is that all my hours of operation are in the knowledge hub, but to this very moment, my agent will still say we are open tomorrow when we’re closed for the holiday. Can you look at this knowledge hub Google Doc of mine and tell me how I can improve it to over come this critical issue? Hours of Operation Chatbot Training - Google Docs
Hi Duncan, no, it’s not possible since ChatGPT is not aware of real-time because it’s not connected to the internet. I checked the Google Docs, and the best solution for you is to create an article with all the different times for the activities. When someone asks if you are open today, tomorrow, or any related question, you can direct them to the article where they can select the activity they are interested in and check the opening hours.
You can achieve this with two types of datasets: articles or iframes.
Please explain how to create an article or iFrame that will allow them to select the program or activity they are interested in? I don’t see where I can provide this type of prompting in the given Recommended Articles or iFrame tools provided.
Hi, you will need to create a webpage on your website that includes a form or buttons for users to select activities. Once a user selects an activity, the opening hours for that activity should be displayed. This functionality is separate from Answerly. When a user asks a question about opening times, the agent will direct them to the webpage where they can find information about the different activities and their respective opening hours.
You can choose to display this directly inside the chat widget as an iframe. This way, the experience will appear as if it is still inside the chatbot. What you are doing is showing the opening hours of your website within the chat widget.
I’m sorry, but I’m just having the hardest time with my agent on this. I’ve done the iframe and it did not work. It’s the latest iframe if you’d like to check it out. Since it didn’t work, I went back to my Hours of Operation Training Google Doc and adjusted the verbiage. That didn’t work.
Prompt is: Are you open today?
I did have success after I adjusted the Google Doc in which it replied, “Which program are you inquiring about.” This is the correct response. My response was Open Play. The agent’s response was “Yes, we’re open today.” It was a correct lucky guess. It doesn’t know the day of the week or the time. It should’ve provided the schedule instead. Please let me know where my training error is.
Hi Duncan, checking on this right now
Just checked your dashboard,
Apologies for the confusion. My advice was to create just one iframe dataset with a link to a page that lists all available times. This way, when the user asks about any opening or closing time, the iframe would be triggered—something like this:
If the user inquires about opening or closing times, or whether it’s open today or closed, trigger this.
So, if the user is asking about any activities, the agent will provide them with the iframe containing all the activity times. This is to simplify the process and reduce mistakes.
However, if you are having success by describing this in a Google Doc and it works fine for you, that’s good. But as I have explained before, the agent does not know the current date or time. Therefore, it could result in the agent presenting inaccurate information depending on the user’s query.
I see. Thanks.