Lets talk about AI Chatbots

EricG

Site Supporter
Aug 2, 2011
673
695
93
Newmarket
30
Lets talk Chatbots

Anyone here using AI Chatbots or Virtual Assistants?
How have you prevented AI Hallucination?
Are you using just internal materials for LLM learning, or are you also exposing your assistant to vast internet?
Have you had better success using PDF documents, or manually created documents centric to AI, created specifically for the purpose of model training?

Have you found a specific model such as gpt-4o (mini) vs gpt-4.1 (mini) seem to work better with internal data sources in document form than others? What other models have your considered?
 
Hi Eric. I've had reasonably good luck with ChatGPT-4o, now ChatGPT-5 (paid), and I find them better than the other offerings out there, especially the free ones like CoPilot.
I find it pretty useful for diving into spreadsheets in addition to creating manuals and other learning tools.
AI hallucination ... I do find that you have to have some reasonable familiarity with the topics you are asking about. I have run into situations where I have told it that I do not think its conclusions are correct, and it has gone back to reevaluate its answer(s).
I have not attempted to train my AI yet, however, it is on my radar.
 
Hi Eric. I've had reasonably good luck with ChatGPT-4o, now ChatGPT-5 (paid), and I find them better than the other offerings out there, especially the free ones like CoPilot.
I find it pretty useful for diving into spreadsheets in addition to creating manuals and other learning tools.
AI hallucination ... I do find that you have to have some reasonable familiarity with the topics you are asking about. I have run into situations where I have told it that I do not think its conclusions are correct, and it has gone back to reevaluate its answer(s).
I have not attempted to train my AI yet, however, it is on my radar.
I tend to agree, the "free" offerings are rate limited and too slow for me, which I typically exceed by 8am, hence why we moved to paid based offerings exclusively. I'm currently using Gemini via API and Chat-GPT via API, I'm typically not using their front end input interface, as its more suited for questions and answer, or report creation, manuals, deep research etc.

I've noticed that certain models have specific strengths, some do the identical task much better than others.

As far as training, I've run in a number of issues that we are resolving or have resolved... typical marketing / sales PDF or client facing docs / materials aren't detailed enough for training, we've played around with training using rate sheets, tariff guides, using specific city/city combos only to ensure the accuracy of the response.

example of a training issue... provided Toronto <-> Barrie LTL Rates, 1 to 12 skids... and also provided the same zone info the 53' truckloads.... however... most models are having a hard time determining when the user is asking for a truckload rate, are they referring to a 53' rate, or a full straight truck truckload...... we had to put in a custom response to clarify the user request... AI doesn't natively seem to like ambiguous words within our industry lol.