Hey guys, I am getting irrelevant responses from the custom GPT i made for my business. usually the first output after my initial prompt is picking up information from past uploads and generating irrelevant outputs. Any solution for this? It doesnt seem to be a prompting issue.
1 Like
Please be cautious of showing information. There is a name and mobile number.
Could you tell me what’s the task you are working on and the purpose of the GPT.
Try:
- Editing your GPT instructions (specify the problem you face and instruct to not repeat it)
- could also be a prompting issue. Depends on your GPT instructions and the first prompt you enter. It’s always better to include your task and goal at the beginning. Then in the end, you could also repeat if you’d like.
Again without basic info, cannot help much.
Please don’t include sensitive information dear.
apologies for the oversight. I have since deleted the screenshot. I have created a custom GPT for assisting me with my restaurant business. I trained it by giving it transcripts of a series of past ChatGPT conversations, survey results, revenue projections and so on. However, it randomly picks up information from the uploads and spits it out, regardless of the initial prompt. For example, if i upload an excel (as was the case here) and ask it to help me remove duplicate entries, it generated the vision and mission statement of my company, which it has picked up from one of my uploads. This is very common. Atleast 50% of it’s initial outputs are it generating letters, marketing strategies, business plans etc which i haven’t asked for at all.
I have configured it to be concise and extremely to the point (or atleast as it claims) but this doesn’t seem to reflect in its outputs. When i ask it a question (like:can you explain why you assumed a 3% growth rate?) it will answer the question but also regenerate the entire output, which can at times take ages. Then when i say do not regenerate the whole thing it apologizes and regenerates it anyway lol
1 Like
No need to apologise. Just here to help.
I got your point. I’ve faced it too when I was creating GPTs for my use case.
I’ll share you solution that worked for me, hope it does to you too.
Just need you to answer this one thing -
- You mentioned that you had uploaded transcripts and other information right. How did you save it? Separate files or all in on document?
Solution: I would suggest you to imagine a clean white board, bring all your information in one place,
MOST IMPORTANT STEP - convert your information to JSON format. It may sound technical but it’s not trust me. JSON is nothing but a properly formatted and organised data that any AI models understand better from.
You can ask ChatGPT to convert it for you.
Next, upload those JSON files to the knowledge in your GPT.
Try it out now.
Also if you’d like speak to me over phone (if your comfortable ) I’d be glad to help. I stay in Bangalore, India.
Sure, I think a call would do well, can we fix one at a suitable time for you? Thanks
1 Like
Regarding this, I’ve DMed you
Haha you remind me of my first experience too. No problem. Click your Profile Icon and you will either see a notification OR a section that is INBOX
1 Like
Hi, welcome.
Hallucinations are a common problem, and all you can do is reduce them.
First, it sounds to me as though your CustomGPT is too generalized. Emails, marketing strategies, and revenue projections are significantly different tasks. I suggest creating Specialized CustomGPTs. One trained on emails, the other trained on revenue and what not. You’d roughly want one GPT per business department—so, marketing, admin, finance, like that.
Second, it sounds like you need to pay attention to the structure of your Instructions. Based on the randomness you’re reporting, create a highly structured set of instructions that explicitly tells the Model how to use the documents in the Knowledge Base. The type of file depends on the information you’re sharing. For example, clearly denote a section about “voice and tone” and other brand information, and you’re less likely to see it trying to help you make your business plan. (It’s confused about the stage of business your in.)
Third, long Knowledge Base files, such as previous full conversations, are hard for the Model to parse quickly. It doesn’t just “know” them. Make sure those documents, whatever they are, are similarly well structured and that their purpose is clearly outlined in the Instructions.
Fourth, debate and questioning are to be expected. Having a model completely understand your desires off of one prompt is unlikely. Expect some back-and-forth to get the conversational context where you need it to be.
Fifth, if you want the model to be more concise, explicitly tell it (multiple times, in multiple places) your preference. Like, literally in every prompt. “Please answer my question briefly without duplicating your previous output. I would just like to discuss the ideas with you. I’ll tell you when I’m ready for you to recalculate.” This is just a consequence of the way the underlying cGPT model is set up. Super helpful and verbose.
Your needs are right on the edge of what an API Assistant can do, which give you more control over the output and creativity. There’s no rush. everything you’re learning is applicable to an Assistant.