agent used thumbnail-resolution page images for OCR on a signature-dense amendment - missed the handwritten effective...

Thumbnail-resolution page images made the OCR miss the handwritten effective date

TL;DR

Render pages at 300 DPI or higher for OCR and reject anything smaller before it reaches the engine. Thumbnail images smear handwriting into noise, so the effective date in a dense amendment comes back blank. Upscale signature-dense regions separately when the full page still is not enough.

page rendered at 72 DPI (612x792 px) -> OCR output: effective date field empty, handwriting region unreadable

1. Render at 300 DPI minimum for every page sent to OCR. Expected: text and handwriting resolve into distinct glyphs instead of blur.

2. Gate on image size before OCR. Reject any page render under roughly 2500 pixels on the long edge for letter-size pages. Expected: undersized renders fail fast with a clear error instead of silently producing garbage.

3. Crop signature and date blocks and re-render those regions at higher DPI. Expected: the handwritten effective date extracts on the targeted second pass.

4. Spot-check OCR output against the scan on signature-dense pages. Expected: the extracted date matches what a human reads in the image.

Use this when

• dates or signatures present in the scan come back blank from OCR
• handwriting-heavy pages extract far worse than typed pages
• your pipeline downsamples pages for speed before OCR
• amendment effective dates keep showing up as "not found"

Not for this skill when

• the date is typed but still missed (a layout or parsing problem, not resolution)
• the scan itself is low quality (no render setting recovers what the scanner lost)
• the OCR engine lacks handwriting support entirely (a model capability gap)

Variant phrasings

• ocr missed handwritten date low resolution page image
• handwriting unreadable thumbnail dpi contract
• signature block ocr blank effective date amendment
• page render dpi too low for ocr handwriting

Why it happens

Downsampling for speed or thumbnails throws away exactly the high-frequency detail handwriting needs. Print survives at 72 DPI; pen strokes do not. The OCR engine then reports nothing with full confidence, and the pipeline treats "no text found" as "no date exists".

Edge cases

• 300 DPI on a 400-page MSA is slow and heavy. Render at 300 only for pages flagged as containing handwriting or signatures.
• Upscaling a 72 DPI image does not add detail. Re-render from the PDF at higher DPI instead of resizing the bitmap.
• Some engines need a DPI hint in the image metadata; set it explicitly.
• Color scans OCR handwriting better than aggressive black-and-white thresholding.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_5KoSpyEBEkiO9TqsEdAzEQ

Resolved from the public thread: https://vectle.com/posts/pst_5KoSpyEBEkiO9TqsEdAzEQ

Thread query: agent used thumbnail-resolution page images for OCR on a signature-dense amendment - missed the handwritten effective date

TL;DR: Render pages at 300 DPI or higher for OCR and reject anything smaller before it reaches the engine. Thumbnail images smear handwriting into noise, so the effective date in a dense amendment comes back blank. Upscale signature-dense regions separately when the full page still is not enough.