Sorry about the previously started “CoPilot” post that I then deleted…my copy/paste didn’t work and I was momentarily under the struggle bus on editing… I blame AI for that as well… lol
I asked (innocently enough) “Is there anywhere in Epicor to add a material scrap factor other than in the Bill of material through engineering workbench?”
You may use any of the available AIs for Epicor Kinetic documentation interrogatory, but you must properly configure the workspace, as the AI is not trained in Epicor’s documentation. But you can essentially provide a workspace that trains it yourself.
I have assisted a number of people in setting up a similar thing.
In short:
Use an Agentic IDE (Google: Antigravity, ChatGPT: Codex, Claude: Claude Code, CoPilot: VS Code Github Copilot, etc.)
Make a folder somewhere to be your workspace root
Load it up with all the user and technical guides from Epicor’s documentation
Instruct the AI to download and install the python tool markitdown
Instruct the AI to convert all the pdfs to MarkDown
Delete the PDFs
Ask your questions (in the agentic IDE)
for on-the-go usage, you can use an AI that offers a cloud workspace that allows you to follow a similar process, uploading to cloud storage, and ask the same questions via an app on your phone. (Perplexity Spaces, Gemini Gems, etc. etc.)
Following this process arms the AI with the appropriate context to answer the questions without as much hallucinations.
You can further increase performance by instructing the AI to construct skills and a CLI framework/toolkit to perform queries against the markdown and return results (optimize for token usage)
This set up will function across all Agentic AIs once deployed.
Agreed that it’s getting worse…I mean usually there may be a field name or whatever that is a bit off…but the whole “SURE YOU CAN DO THAT” all of sudden becoming “THERE IS NO WAY IN #$%^ THAT CAN BE DONE” is causing more WTF moments in my day …and I really think there are enough already!
In this article the writer is suggesting that by writing an overly specific prompt (like the one in the article) you are doing yourself, and AI a disservice. I admit I am guilty of writing prompts like this too. The problem is that sometimes it works and solves the immediate problem you’re having. Just be sure not to confuse these lucky breaks, with the AI actually knowing what the right thing to do is.
The way I imagine it is that if I can ask a question that I could solve with some clever Google-fu, then AI should also be able to find that answer. If I know the question I am asking would never be found via Google-fu, then I know the answer I get from AI is likely going to be overly confident, and often wildly misled.
If it would only warn you about that at the beginning of the response instead of after being caught in the “lie”.
That question to the AI also exposes some inherent misunderstandings of what these AIs actually do. We think an AI is a computer, and by extension has all the precision and capabilities of a computer. When in fact, it is nothing more than a prediction engine. These AI models don’t have any inherent logic.
I stick to Claude for Epicor, as it seems to be fairly well trained on it.
It also helps that my prompts ALWAYS include “Please check Epiusers for additional verification before making unsubstantiated claims regarding Epicor”…
Not sure how other models do so, but there is a feature called “Projects” that let’s a user define certain criteria/rules for chats (using .txt or .md files) that are referenced prior to any response. I use this to help avoid nonsense claims and circular reasoning. It isn’t perfect, but it makes a big difference.
I use Claude also. I have an Epicor project I put a bunch of rules into to help filter out some of the nonsense.
I also added this into my user settings as global rules to stop it from lying, assuming, or hallucinating as much as possible.
##Self-Correction
### If you catch yourself making an assumption, call it out explicitly
### Before finalizing a response, check whether you've actually answered what I asked
## Reasoning & Computation
### Always do arithmetic via code, not tokens
##Factual Accuracy
### Search the web before answering any question about current events, prices, people's roles, or anything time-sensitive
### If you're not confident, say so explicitly rather than guessing
### Cite your sources when making factual claims. If there are no sources, explain your logic.
### Don't assume — ask me to clarify if my request is ambiguous
##Output Formatting
### Keep answers concise unless I ask for detail
This also saves on token usage! This is also one of the reasons for the Epicor Idea to allow us to store our comments in Kinetic as Markdown rather than unstructured text or RTF.
I want to say that you also had some guidance for code generation explaining the limits of code widgets: no class defitions, allow delegates if needed, restrict imports to libraries available, etc..
@cpilinko “Keep answers concise unless I ask for detail” - This is something I will be borrowing from you! I find myself often asking for it tighten up responses.
I had good success with putting all the docs I to notebook LM ask it questions and it will give you an answer as well as source links back to the docs to prove it. Of course it’s only as good at the documentation. As we all know there are gaps. The only way to fix that is test it yourself and include the findings in you AI source.
We’ve got a few folks around here that think everything AI tells them must be true! Example - nope don’t know anything about implementing blah blah – but I asked AI and it gave me all the documents and procedures we’ll need!
I was sad when it was locked.
