scanned schedule table read as one long column - OCR destroyed the fee tiers and the agent extracted a single blended...

TL;DR The OCR engine linearized a fee table into one text column, destroying the tier structure, and the agent averaged everything into a single blended rate. Extract tables with a layout-aware method that preserves rows and columns, validate the geometry before trusting the numbers, and sanity-check that the tiers make sense. A rate extracted from a flattened table is a guess, not a fact.

What the flattened OCR output looked like:

Tier 1 up to 10000 2.5 percent Tier 2 10001 to 50000 2.0 percent Tier 3 over 50000 1.5 percent

Steps

1. Confirm the table was flattened. Look at the raw OCR text for the schedule page and check whether tier labels and amounts run together in one column.

Command: grep -rni "tier" ocr_text/ | head Expected: tier labels and rates on one continuous line, proving the row structure was lost in linearization.

1. Re-extract the page with a table-aware method. Use your OCR engine's table mode or a layout tool that returns cells with row and column positions, not plain text.

Expected: the schedule comes back as rows and columns, with each tier on its own row and the rate in its own column.

1. Validate the geometry. Check that every row has the same column count and that the tier thresholds increase monotonically down the rows.

Expected: three rows, consistent columns, thresholds ascending. If a row is short or thresholds jump backward, the table parse failed and you re-run with different settings instead of emitting.

1. Sanity-check the extracted rates against the document. Tiered rates should apply to their own bands, not blend, so verify each rate sits next to its own threshold range.

Expected: each tier reported with its own rate and range, no single blended number anywhere in the output.

Use this when

• Scanned schedules, order forms, or exhibits contain tiered pricing, fee tables, or rate cards
• Your agent reports a single rate where the document clearly has a table
• OCR text shows table content running together in one column
• Extracted amounts look like averages or sums of values that should be separate

Not for this skill when

• The document is born-digital, extract the table from the text layer or the embedded structure
• The table is genuinely one column, check the source page before assuming flattening
• The rates are wrong but the table structure is intact, that is a cell-mapping problem
• The schedule is an image with no grid lines at all, which may need manual transcription

Variant phrasings

• scanned fee schedule read as one column, agent extracted a blended rate
• OCR destroyed the table structure in a contract schedule
• how to extract tables from scanned contracts without flattening
• fee tiers merged into one rate by the extraction agent
• table-aware OCR for contract schedules and order forms

Why it happens

Plain OCR reads a page top to bottom, left to right, with no notion of columns, so a table becomes one long text column and the cell boundaries vanish. The extraction agent then sees a stream of numbers and labels with no structure, and its best guess is to combine them into a single rate. The table was never really read, it was transcribed as prose, and prose has no tiers.

Edge cases

• Tables that span two pages need the rows stitched before validation, check for a repeated header row as the stitch point.
• Merged header cells can shift column alignment, validate against the data rows, not the header.
• Footnotes under the table often get sucked into the last row, strip footnote markers before the geometry check.
• Currency symbols and percent signs may land in adjacent cells, normalize them into the amount cell before the sanity check.
• If table mode fails on a badly skewed scan, deskew the page first and retry before falling back to manual review.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_xNWBf7ez0uyIpCr1oVaSFQ