Make It Possible
The Spec Gap
Everyone has bigger and more important decisions to make, but each small decision also factors into the overall performance. I guess the only way out is to have a system that streamlines the smaller, repetitive tasks early on. Make them so easy that everyone can finish them with one click before they snowball and become big problems.
AI: Same task. Every run gave me a different table.
Every AI run gave me a different table.
I said I would share AI limitations and use in my adaptation to hospital and equipment planning.
This one is about consistency.
I was generating a comparison table for MRI machines. Same task. Same prompt. Every run gave me a different table.
Not different data. Different columns.
One run gave me device name, brand and magnet strength and source link. The next dropped the link. I asked for it back. The next run had the source link back and an extra column nobody asked for, with an essay squeezed into a tiny cell.
The LLM was deciding on its own what was relevant each time. Poor memory of what it did before.
This is frustrating and time consuming. Not possible to scale when you are sourcing thousands of devices.
The fix was handing the LLM a form and telling it to fill it in, not invent its own. Exact field names, exact structure, no additions, no omissions. In technical terms this is called a JSON output parser.
If the AI fails to produce the format, I get an error that outlines an easy fix, instead of a wild answer.
The table became consistent. Every run. Every device.
No more being a nagging old lady full of grudges.
This is also baked into the cutsheet generator of Equipia. The output format is fixed.
You always get the same MEP details in the same order regardless of how funky the manufacture webpage design is.
You may try it in a LLM but I would recommend going straight into equipia.io for this one.
Someone you know has been through this? Forward it to them.