Azure Form Recognizer - Recommeded approach

Question

Dear team,

I have a business scenario which involves training custom model (Invoice model) for 100 vendors. Kindly help me with the recommended approach as mentioned below.

Approach 1:
Train 1 model per vendor, create 100 models for 100 vendors and compose it into a single model.

Cons:

1. Difficult for model management and retraining.
2. Increase in vendor count in future cannot be added to the same compose model (as the limit is 100 models per composite model)

Approach 2:
Train 1 model for 100 vendors (5 invoices per vendor for training, also considering the limit of 500 pages for model training set).

Pros:

1. Create new models and compose for new vendors
2. Model management requires minimal effort
3. Retraining can be done easily as we are having single model

Cons:

1. Labelling line items using a single label does not support multiple page tagging

Approach 3:
Train 1 model for 100 vendors - 1 page invoice, Train 1 model for 100 vendors - 2 page invoice,... so on (5 invoices per vendor for training, also considering the limit of 500 pages for model training set) and compose it into a single model.

Pros (same as for approach 2):

1. Labelling line items can be done using multiple labels per page.

@YutongTie-MSFT , Kindly help me with this query. Thanks in advance :)