ParseMyStatement v1.1 — OCR 4.0 with Block-Level Confidence & Structural Metadata

ParseMyStatement v1.1 — OCR 4.0 with Block-Level Confidence & Structural Metadata

We are excited to announce ParseMyStatement v1.1, a significant upgrade to our OCR pipeline that introduces block-level structural metadata, per-page confidence scoring, and automated low-confidence detection.

This release represents a leap forward in transparency and reliability for automated bank statement extraction. For the first time, users and developers can inspect exactly how the OCR engine interpreted each page, where it was confident, and which pages may need human review.

What's New

1. Block-Level Structural Metadata

Every OCR'd page now returns a blocks array with typed elements:

{
  "ocr_metadata": {
    "0": {
      "blocks": [
        {
          "type": "title",
          "content": "## HSBC Live+ Credit Card",
          "top_left_x": 46,
          "top_left_y": 125,
          "bottom_right_x": 204,
          "bottom_right_y": 142
        },
        {
          "type": "table",
          "content": "<table>...</table>",
          "table_id": "tbl-0.html",
          "top_left_x": 46,
          "top_left_y": 249,
          "bottom_right_x": 656,
          "bottom_right_y": 852
        }
      ]
    }
  }
}

Block types include: title, text, table, list, code, equation, caption, header, footer, signature, image. This enables downstream applications to:

• Filter headers/footers before LLM normalization
• Validate table boundaries against extracted transaction counts
• Reconstruct page layout for audit and compliance workflows
• Target specific regions of a document for re-processing

2. Per-Page Confidence Scores

Each page now reports:

{
  "confidence_scores": {
    "average_page_confidence_score": 0.9842,
    "minimum_page_confidence_score": 0.1937
  }
}

• Average confidence — overall OCR quality for the page
• Minimum confidence — lowest-confidence character/word on the page (useful for spotting problematic regions)

Pages with an average confidence below 0.7 are automatically flagged in the low_confidence_pages array, and a warning is issued in the response.

3. Intelligent Low-Confidence Page Detection

The converter now automatically:

1. Calculates average confidence per page
2. Flags any page below the 0.7 threshold
3. Adds a human-readable warning: "Page 5 OCR confidence low (0.65)"
4. Populates low_confidence_pages: [5]

This gives you a structured, machine-readable signal for quality control without needing to manually inspect every page.

Performance Benchmarks

We ran rigorous A/B tests across 4 scenarios comparing v1.0 (baseline) against v1.1 (blocks+confidence). All tests used the same LLM model (DeepSeek-V4-Flash) and same system prompt — only the OCR pipeline changed.

Test 1: Generated Eval Statements (3 banks, 81 transactions)

| Metric | v1.0 | v1.1 | |