[Python SDK] Feature Parity Gap: `errorMode`, `correctOrientation`, `trimEdges`, `schema`, and `extractPerPage` Missing vs Node.js SDK

Overview

The Python (py-zerox) SDK is missing several features that are available in the Node.js SDK. This creates a significant parity gap for Python users. The README's own comparison table already acknowledges these gaps, but there is no tracking issue for resolving them.

Missing Features

1. error_mode — Configurable Error Handling

Node.js supports ErrorMode.THROW and ErrorMode.IGNORE.
In Python, there is no equivalent — failed pages have no configurable behavior.

Expected Python API:

result = await zerox(
    file_path="file.pdf",
    model="gpt-4o-mini",
    error_mode="ignore"  # or "throw"
)

2. correct_orientation — Auto-Orientation Fix

Node.js uses Tesseract-based OSD to detect and correct rotated pages before sending to the vision model. Python has no such step, leading to degraded OCR quality on scanned/rotated documents.

3. trim_edges — Background Edge Trimming

Node.js trims edge pixels that match the background color (top-left pixel default), improving model focus on actual content. No equivalent exists in Python.

4. schema + extract_per_page — Structured Data Extraction

This is the most impactful missing feature. Node.js supports JSON Schema-based extraction:

const result = await zerox({
  filePath: "invoice.pdf",
  extractOnly: true,
  schema: {
    type: "object",
    properties: {
      invoice_number: { type: "string" },
      total_amount: { type: "number" }
    }
  }
});

Python users have no native way to do structured extraction — they must manually post-process raw markdown output.

Feature Parity Table

Proposed Implementation Path

Environment

• Package: py-zerox (latest)
• Reference: Node.js zerox npm package

Notes

• This is a meta-tracking issue for the full Python parity gap — not a duplicate of any single issue.
• LiteLLM (already a dependency) supports structured outputs, so schema support should be feasible without new dependencies.
• Happy to open a draft PR for error_mode or schema if maintainers confirm the preferred API shape? @tylermaran @annapo23