Smart Vision-Feature Clarification

Contents

  1. Document Understanding
    1. Field level accuracy
  2. Smart Vision Packaged Workflow
    1. Field level accuracy
    2. Line Item accuracy

Document Understanding

Item

Currently Supported

Testing Details

Observation

Future Plans

Data extraction accuracy for Annual Reports

 

Field level accuracy shown below

Verified by uploading annual reports from multiple companies

Field accuracy needs to be improved for Annual Reports extraction

Not in current plan

Field selection change

Supported fields are pre-selected for extraction

Verified by creating DU projects for annual reports and invoices

Fields selected for extraction cannot be changed

Not in current plan

Orientation Correction

Orientation correction is supported only for non-searchable PDFs

Verified by uploading PDFs rotated clockwise, anti-clockwise and upside down

Orientation correction is not applied on searchable PDFs

Not in current plan

Field level accuracy

Company Name

Report Year

CEO

Total Assets

Total Liabilities

73

93

86

66

20

 

Smart Vision Packaged Workflow

Item

Currently Supported

Testing Details

Observation

Future Plans

Invoice Extraction Time

Extraction time for a single page invoice is 1 minute

Verified by uploading a single page invoice and calculating time based on UI status

More than 6 seconds for a single page invoice can slow down multi page invoice extraction in real scenario

Not in current plan. Higher speeds would need GPU

Field data accuracy

Field level accuracy below

Verified by validating against Tolas and PP invoices for Navistar

Field wise extraction accuracy is not good for all 14 fields

HITL feedback-based model training will improve field suggestions and accuracy

Line item data accuracy

Line item extraction works for tables with strong closed borders containing in a single page

Verified by validating against invoices containing line items in tabular and non-tabular format

Line item extraction will not work for invoices with partially closed bordered tables, multi-page tables and non-tabular representations

Fully fledged Smart Extract features will help to address these

Accuracy for fields with dark background

Fields with light or no background

Verified by validating against Tolas and PP invoices for Navistar

Extraction is not possible for FOI data in dark background

 

Not in current plan

Extraction time calculation

Extraction time in analytics is total time in DU - which includes queue wait time and execution time

Verified by validating extraction time in Admin Analytics section

Extraction time for an invoice will not be consistent; as the queue time can vary

No current plan for this

Orientation correction accuracy

Landscape oriented images will be converted to portrait mode

Verified by validating with a landscape-oriented invoice

2 out of 75 invoice pages had wrong orientation correction applied

No current plan for this

Extraction Traceability

Basic traceability is available from Run list, Trace logs and Batch summary

Verified by running extraction against multiple types of invoices

The extraction process is not fully traceable end-to-end

End-to-end traceability is planned for future releases

Delay in updating invoice count & flags in Run list

For a batch of 10 invoices, 5 minutes will be taken to update the run list

Verified by running batch of invoices

When a batch is processed in DU there is a delay in reflecting the invoice count & flags in Run list

 

No current plan for this

Delay in updating Accuracy Analytics

It takes 5-7 minutes to get the details updated in analytics

Verified by making corrections to a batch of invoices

Line item accuracy graph and Insight accuracy graph is not updated immediately after making corrections

No current plan for this

Advanced section performance

Advanced page works fine for a project with invoices less than 700

Verified by running 700+ invoices in a project

Advanced page hangs when project is having more than 700 invoices (Intermittent Issue)

No current plan for this

PDF Highlighting

 Extracted values are highlighted in the source PDF for all fields except 'Vendor Code' / 'Invoice number fields (for some invoices)

 

Verified by running batch of invoices

Extracted values for Vendor Code & Invoice Number fields are not getting highlighted in the PDF in preview screen for some invoices

Planned for future releases

PDF Highlighting

Full text for extracted values are highlighted in source PDF for most of the fields

Verified by running batch of invoices

Extracted value for 'Ship to Code' is not fully highlighted in the source PDF. Only partial text is highlighted for some invoices

Planned for future releases

Field level accuracy

Company Address

Purchase Order Number

Ship To Code

Vendor Code

Company Name

Freight Amount

Invoice Date

Invoice Number

Miscellaneous Charges

Scan Number

Sub Total

Total Amount

51

64

90

97

79

94

83

63

34

97

73

72

 

Line Item accuracy

Part Number

Quantity

Customer PO Number

Unit Rate

Extended Cost

58

49

30

54

52