I have no idea whether these ideas would help, but since coming back to Elfsight, I’ve found the new AI better at parsing data from its knowledge base files. Have you tried skipping the entire Q&A section of the training and instead creating an I/O spreadsheet that can be uploaded as a PDF or JSON file? I know the FAQ file I provided to my bot educates it enough so that it gives more accurate responses.
Lastly, the AI continues to act outside of the parameters set and still doesn’t seem to be truly “sandboxed,” so it consistently offers more than what it’s instructed to respond with. The key is giving it enough guardrails to keep it reigned in. I’m attaching two screenshots. One is the Agent Instruction for how I want it to end a response. You’ll see that it does exactly as I asked. The problem is, it’s going above and beyond what I’ve instructed it to do. It doesn’t have instructions for it to review and comment on the source data. I also don’t tell it to provide follow-up instructions where it offers more ways to drill down into the data. So far, these haven’t been egregious enough to be a deal breaker, but there may come a day when it offers more than what it can p
So, I already had the following in the instructions:
”Give concise, factual answers only from uploaded knowledge. Verify with sources before responding. Do not include unsupported information. Keep responses concise and focused.”
I’ve added your suggested message as:
”…Keep responses concise and focused. Avoid unnecessary questions or extra details.”
After publishing these changes and reloading the developer dashboard webpage, the following was added at the end of the chatbot’s response:
”If you’d like, I can:
call out which ones offer emergency/grocery box pickup vs. regular pantry service, or
pull hours and eligibility details for any specific pantry above.”
This is the issue I was talking about to @Petar_Dietrich. My data doesn’t include information about which food pantries offer grocery boxes vs. traditional pantry services. My data doesn’t include hours of operation or eligibility details.
So, even though it has explicit instructions not to use external information, or “extra details,” the bot has taken it upon itself to make these offers to my users. This can only result in one of three things. Either it will provide such information from an external source, it will fail to find the promised information and that will cause the user to distrust this bot and my directory, or it will lie to the user through hallucination. None of these are good.
Our devs revised your instruction and here is the version that should work great:
## Assistant Identity
- **Business Name:** NRV Wayfinder
- **Business Industry:** Community Services Directory
- **Business Overview:** NRV Wayfinder is a locally curated directory that connects individuals with support services in the New River Valley area. The platform enables users to find resources and assistance tailored to their specific needs easily.
- **Assistant Role:** Community Resource Guide
- **Contact Details:** nrvwayfinder@gmail.com
## Tone and Style
- Respond in a helpful, welcoming, and friendly tone.
- Keep responses conversational, clear, practical, and concise.
- Stay focused on the user's exact request.
## Core Response Rules
- Answer the user's specific question directly.
- Include only information relevant to the request.
- Do not add recommendations, related services, or next steps unless the user explicitly asks for them.
- Do not add notes, tips, disclaimers, summaries, or extra commentary unless explicitly requested.
- Do not describe your reasoning or search process.
- Do not ask follow-up questions unless they are necessary to answer correctly.
## Response Behavior
- Provide clear, factual, user-oriented answers.
- Always prioritize finding and presenting the requested information before suggesting that the user contact an organization directly.
- When relevant, prioritize business or housing information.
- If multiple results exist, show the most relevant options first and avoid long lists unless the user asks for more.
## Business and Location Details
When referring the user to a business or housing location, include all available:
- Full address
- Phone number
- Google Maps link
- Website
Always include full contact information whenever it is available.
## Handling Missing Organizations or Services
If the user asks about an organization or service that is not available:
1. Briefly state that the specific organization or service is not currently available.
2. Immediately offer local alternatives of the same service type.
3. For each alternative, include all available:
- Organization name
- Full address
- Phone number
- Website
- Google Maps link
- Short description of services
## Formatting Rules
- Use plain, clean formatting.
- Do not add headings or extra sections unless they improve readability.
## Closing Requirement
End every response with this exact sentence as a separate final line:
If you need further assistance or more detailed information, please don't hesitate to ask or visit the app "NRV Wayfinder" (https://nrvwayfinder.glide.page/)
Okay. This is better, but there are a couple of issues to address:
If you were going to provide me with a revised prompt that works (which I appreciate), why did they already change it in my bot’s instructions. I would have preferred the option to copy/paste my old instructions first, in case the new ones didn’t work. Simply “slamming” me into their prompt is a bit rude.
Considering my housing file has full URLs and phone numbers, why is the AI saying that they’re “…partially redacted or not listed”? I currently wrap the text to limit the overall width of the spreadsheet. Please ask your devs if the AI has to see them as a single line in a cell. This will make the font much smaller and I don’t know at which point the AI won’t be able to read the content. It’s always been a concern and I even dropped some of my information just so the spreadsheet didn’t get too wide so the font wouldn’t be extremely small (it’s already small). Thanks
The AI is still making offers it can’t complete. This will create visitor frustration and distrust of the bot and my directory. In my testing today, it offered to provide the hours of operation and further details for any location it had just listed. I chose one and, just as I expected, it failed to provide the info. Why? Because I don’t include that information in it’s knowledge base. Isn’t there a way to prevent the AI from making empty promises. It’s concerning and unprofessional.
Overall, the new prompt is working better. Thank you. However, these two issues need to be resolved, then things should be good to go and the residents in my community will have a wonderful new way to access the resources they need, without the risk of misinformation.
To you and your devs, I say thank you for the time and effort that’s been required to get your bot acting correctly.
I am so sorry for going ahead and saving the new instruction in your widget!
I see how that change felt a bit abrupt, and I completely understand your concern. You’re right, we should have given you the option to save your old instructions first, and I’ll make sure we improve this process so it doesn’t happen again.
In the meantime, we’ve saved your original instruction and I’ll send it to you in a personal message.
Thank you for the feedback and understanding!
Regarding the 2 issues reported, I’ve just sent the questions, where the issues occured before, and now the bot is responding correctly:
I also need to clarify a bit more about how test messages are counted.
Earlier, I mentioned that messages sent from the editor wouldn’t count anymore, but it turns out things are a bit different.
The devs have informed us that we’ve extended the message limit for testing. However, it’s not unlimited.
Each user now gets 1,000 free messages for testing from the configurator each month. Once you reach that limit, any extra messages sent from the configurator will count toward your overall message limit.
Thanks. A chance in protocols is probably best. I appreciate your understanding, not just the actions taken but that some of us like to keep our work so it can be referred back to or used as a fallback.
I have a meeting in 5 minutes, and a busy day, but I’ll work with the bot when I find the time. I’ll send an email with feedback.
Thanks
Well, if I need to use more than 1,000 messages ever, let alone in a month, then this isn’t the platform I need for my bot. That’s quite generous, and I hope I never come close to the limit.
Well, I just tried the bot. It still can’t relay the information from the housing list file. In fact, upon further review, the first screenshot you provided has incomplete phone numbers! The last four digits are replaced by “xxxx.” A directory is useless if the user can’t even receive a phone number to contact an organization. So, I don’t see an improvement…
And previously, it was handling the foot brace question just fine. It suggested clinics and orthopedics that could help, and provided their full contact information. Since a new thread was begun, I can’t scroll back up to send a screenshot. Now it does nothing but ask for more information that isn’t contained in the knowledge base. This is definitely a step back from what it was doing. The 2nd screenshot I provided in my previous email in this thread was actually taken when the AI offered to provide hours of operation and “further details.” That was for one of the clinics it had listed when I asked it for help with my foot brace. That should confirm that it was working correctly after your dev’s new instructions, but before whatever recent change was made. It was sent at 9:05 AM this morning.
Ha! Maybe I’ll need those 1,000 prompts after all since the bot refuses to act right. Unfortunately, it looks like six months hasn’t changed much since I last tried Elfsight. It may be a newer AI, but it’s ability to follow directions hasn’t improved.
