topic Re: Decipher enable ML query in Product Forum

Decipher enable ML query

Tejaskumar_Darji — Thu, 25 Jul 2024 08:13:23 GMT

Hello team,

I'm wondering why the configuration to enable machine learning in decipher IDP is managed in a config file rather than at the UI level.

Additionally, having it disabled by default doesn't seem logical to me. Could you please clarify the reasoning behind these?

Re: Decipher enable ML query

Ben.Lyons1 — Thu, 25 Jul 2024 09:12:49 GMT

Hi Tejaskumar,

It's a base-level function which requires restarting some of the application components, putting it in the UI would require the platform to be restarted and potentially cause issues for the users.

It's false as a default because if it's installed locally, it has the potential to max out the CPU and rendering the machine temporarily unusable. So primarily it's a safety mechanism. Additionally the additional ML model capabilities are not required for the majority of use cases, as the training data ML model is always on and is sufficiently accurate.

Thanks

Re: Decipher enable ML query

Tejaskumar_Darji — Thu, 25 Jul 2024 15:21:50 GMT

Ok, I see.... so to use ML in Decipher we can enable it in UI for each Doc Type/DFD and use it. And let this flag stay false? If I want to evaluate Decipher document processing with ML capability without turning this to True can we test? Would that be the right way?

Re: Decipher enable ML query

Ben.Lyons1 — Fri, 26 Jul 2024 07:13:43 GMT

Not quite.

The EnableModelTrainingML specifically relates to the Capture Model ML functionality which can then be set up in each Document Type. So you have to enable it in both places.

If the config is left as false, the native rules-based ML model (Training Data) is still active.

I personally don't tend to enable this in each Document Type, as I mostly get the necessary accuracy without it. The Capture Model ML will only improve an already well-trained native model, not for initial training or to resolve issues with a poorly performing set of training data.

Check out the bit at the bottom of the best practice guide for more info.

Thanks