Inconsistent table structure output between OCRFast and OCRHighQuality

[mirinterplay]

Description

I found inconsistent behaviors in table recognition results between OCRFast and OCRHighQuality in uniparser-tools.

When using:

• OCRFast
  • the table reconstruction requires combining:
    • placeholders
    • contents
    • structure
• OCRHighQuality
  • the structure field already contains the final table data directly

This causes inconsistent downstream parsing logic because the output schema/semantics are different between the two OCR modes.

---

Expected Behavior

Ideally, both OCR modes should provide a consistent table structure format.

For example:

• either both return normalized table structures requiring reconstruction
• or both return fully reconstructed table data

A unified output format would make integration and downstream processing much easier.

---

Actual Behavior

OCRFast

Need to reconstruct table data manually from:

{
  "placeholders": ...,
  "contents": ...,
  "structure": ...
}

[uni-parser]

Thanks for the report — this matches what we see.

Table-related fields (e.g. placeholders / contents / structure) can differ in meaning and completeness depending on ParseMode (such as OCRFast vs OCRHighQuality) and other parse settings. Output shape is not fully uniform across all configurations today; that is a known limitation on our side.

For integration, please branch on the table mode you request (and/or the fields present in the response) rather than assuming one schema for every mode.

We plan to document these mode-specific differences more clearly.