To meet the specific automation needs of your various lines of business, you can configure transcription, or finetuning, models at the flow level.
This flexibility allows you to:
enter dedicated transcription automation and accuracy settings for each flow,
create flow-specific transcription models, and
share transcription models across flows.
If you share transcription models across flows, training data is shared among the flows using the same model.
Entering flow-specific settings
You can enter transcription automation settings for each of your flows in the General Transcription, Structured Document Transcription, and Semi-structured Document Transcription sections of a flow's settings. For more details, see Flow Settings.
If multiple flows are using the same model and those flows have different training settings, the model uses the settings from the flow where those settings were most recently changed.
Creating and assigning transcription models
It is not currently possible to assign and create transcription models in the application. Instead, the model a flow uses is determined by the trascription_model property in the flow's JSON.
If you would like to create new models for your flows, contact your Hyperscience representative for assistance. Your Hyperscience representative can create a model by entering a value for transcription_model in your flow's code (e.g., transcription_model=invoices).
When a new model is created, the system obtains training data from the submissions processed through the model's flow or flows. It does not run finetuning for any flows using the model until the minimum required amount of training data is gathered, unless:
You run a script that migrates past QA data generated from the model's flows. Contact your Hyperscience representative for more information about the script.
You've just upgraded, and the system-level training data from the upgrade falls within the set period of records to use.
If a flow is modified to use a model that already exists, the flow begins using the model, and any submissions processed through the flow are used to train the model. By default, any training data created from the flow before the change of model assignment is not used to train the new model. However, you can run a script that migrates past QA data generated from the flow to the new model. Contact your Hyperscience representative for more information about the script.
Model creation and assignment when upgrading to v32
If you are upgrading from v30 or v31, all flows created in these versions use the system-level finetuning model and thresholds (transcription_model=null).
The "Document Processing (V32)" flow, which the system automatically creates during the upgrade process, has transcription_model set to IDP. Any custom flows created in v32 will also have transcription_model set to IDP by default. However, your Hyperscience representative can also set transcription_model to null or omit it completely if you want to use the system-level model.
If any live flows are using the system-level model, the model is trained on all QA data.
If flows are using models other than the system-level model, the models are trained on QA that:
comes from fields processed by the model's flows,
was generated before the upgrade, or
comes from flows that use the system-level model.
The model uses system-level QA data that falls within the set period of records to use. You can also run a script that migrates any past training data to a specific model, even if it falls outside of the set period. For more details on this script, contact your Hyperscience representative.
In order to prevent delays in processing submissions after upgrading to v32, be sure to run recalibration with a v32 trainer on your v31 application during the upgrade process. The system will run recalibration on any flows that are live when the upgrade occurs. We recommend completing this recalibration as an upgrade best practice, even if you don’t plan to use the flow-specific transcription models.