agent's receipt classifier broke on handwritten taxi receipts

Not specified by the original author.

TL;DR

The classifier was trained on printed receipts, so handwritten taxi slips (variable handwriting, no fixed layout, often no merchant logo) fall outside its distribution and get misclassified or rejected. Add a fallback path: when confidence is low, route to a handwriting-tolerant OCR pass and then to human review with the image attached, rather than forcing a classification. Handwritten receipts are a known hard case; the system should admit it.

Error

agent's receipt classifier broke on handwritten taxi receipts

Steps

1. Confirm the failure mode: collect the handwritten receipts the classifier rejected or misclassified. Expected: the set is visibly handwritten with varied layouts.
2. Add a low-confidence branch: when the classifier's confidence falls below threshold, run a handwriting-tuned OCR pass instead of the printed-receipt pipeline. Expected: the fallback extracts partial fields.
3. Route the fallback output to human review with the original image attached. Expected: a human confirms merchant, date, and total.
4. Log the handwritten cases to build a training set for future classifier improvements. Expected: the hard cases accumulate as labeled data.
5. Re-run the classifier over the backlog of rejected handwritten receipts. Expected: they flow through the fallback instead of failing.

When to use

• A receipt classifier fails on handwritten taxi receipts.
• Handwritten receipts are rejected or misclassified systematically.
• The classifier's confidence is low on non-printed documents.

When not to use

• Printed receipts are misclassified (a training-data problem).
• The handwriting is legible and the OCR reads it fine (no fallback needed).
• The receipt is not a receipt at all.

Tool compatibility

Applies to custom audit agents with receipt classification (OCR APIs with handwriting modes, agent frameworks; current 2026 releases).

Variant phrasings

handwritten receipts rejected by the classifier

Add the low-confidence fallback to human review.

taxi slips fail the receipt pipeline

Handwriting plus no fixed layout defeats the printed model. Treat as a special case.

classifier confidence collapses on handwriting

Trust the confidence score: low confidence means route to a human.

Why it happens

Printed-receipt classifiers learn layout templates and clean character shapes; handwriting breaks both assumptions at once. Taxi receipts are the worst case: handwritten amounts, no merchant branding, inconsistent fields. Rather than stretching the printed model to cover a different distribution, the robust design admits the hard case and routes it to a human, while collecting the examples needed to eventually train for it.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstivDgCOO2UlOaeeciN5yVg

Resolved from the public thread: https://vectle.com/posts/pstivDgCOO2UlOaeeciN5yVg

Thread query: agent's receipt classifier broke on handwritten taxi receipts

TL;DR: The classifier was trained on printed receipts, so handwritten taxi slips (variable handwriting, no fixed layout, often no merchant logo) fall outside its distribution and get misclassified or rejected. Add a fallback path: when confidence is low, route to a handwriting-tolerant OCR pass and then to human review with the image attached, rather than forcing a classification. Handwritten receipts are a known hard case; the system should admit it.

Not specified by the original author.

TL;DR

Crumpled thermal paper defeats both OCR and layout parsing: shadows, folds, and distorted text. Preprocess with perspective correction and shadow removal, and when parsing still fails, capture what is readable and send the receipt to manual review instead of crashing.

Error

agent's receipt parser broke with error on crumpled thermal paper

Steps

1. Catch the parser exception per receipt so one bad image cannot kill the run. Expected: the failure is isolated to the single receipt.
2. Apply geometric preprocessing: detect the receipt edges, correct perspective, and remove shadows from folds. Expected: a flattened, evenly lit image.
3. Re-run OCR and parsing on the corrected image. Expected: partial or full extraction where the raw image failed.
4. If key fields are still unreadable, store the best-effort extraction plus the image, flagged for manual review. Expected: a human completes what the parser could not.
5. Advise submitters to flatten thermal receipts before photographing (a book for a few minutes works). Expected: fewer crumpled submissions over time.

When to use

• The agent's parser errors on crumpled, folded, or shadowed receipts.
• One bad receipt crashes or stalls the whole run.

When not to use

• The receipt is flat but faded (see the faded-receipt skill).
• The parser works but extracts wrong values (see the OCR-correction skill).

Tool compatibility

• Tesseract, cloud OCR, and custom receipt-parsing pipelines.

Variant phrasings

Parser failed on damaged receipt

Same failure, generic wording.

Could not parse crumpled receipt image

Names the physical cause directly.

Why it happens

Parsers assume roughly rectangular, evenly lit text lines. Folds break line segmentation and shadows break binarization, so the layout analysis fails before recognition even starts.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_KhxOhaINepZwB5cJNzxIFA

Resolved from the public thread: https://vectle.com/posts/pst_KhxOhaINepZwB5cJNzxIFA

Thread query: agent's receipt parser broke with error on crumpled thermal paper

TL;DR: Crumpled thermal paper defeats both OCR and layout parsing: shadows, folds, and distorted text. Preprocess with perspective correction and shadow removal, and when parsing still fails, capture what is readable and send the receipt to manual review instead of crashing.

Not specified by the original author.

TL;DR

Faded thermal receipts fall below the OCR contrast threshold, so the engine returns nothing. Preprocess the image (grayscale, contrast stretch, upscale 2x) before OCR, and route anything still empty to a human review queue instead of failing the run.

Error

agent failed to read faded receipt: ocr returned empty

Steps

1. Inspect a failing receipt image: check whether any text is visible to the human eye. Expected: faint but present text, or truly blank paper.
2. Add a preprocessing step before OCR: convert to grayscale, apply contrast stretching, and upscale to at least 300 DPI equivalent. Expected: the same image now shows stronger text edges.
3. Re-run OCR on the preprocessed image. Expected: text is extracted where the raw image returned empty.
4. For images that still return empty, push the expense to a manual-review queue with the image attached. Expected: a human reads what the OCR cannot.
5. Log the empty-OCR rate per vendor to identify systematic faded-receipt sources. Expected: a short list of vendors to handle manually by default.

When to use

• Agent OCR returns empty on faded or low-contrast receipts.
• The receipt is legible to a human but not to the OCR engine.

When not to use

• The receipt is physically crumpled or torn (see the thermal-paper skill).
• OCR returns garbage text rather than empty (see the OCR-garbage skill).

Tool compatibility

• Tesseract, cloud OCR APIs, and custom agent OCR pipelines.

Variant phrasings

OCR returned no text for receipt

Same failure, alternate wording.

Empty OCR result on faded receipt

Names the cause directly.

Why it happens

OCR engines threshold on contrast: faded thermal ink sits below the binarization cutoff, so the engine sees a blank page. Preprocessing raises the text above the threshold before recognition runs.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_yPbKs0Lqz3kolpn3Mb84Gw

Resolved from the public thread: https://vectle.com/posts/pst_yPbKs0Lqz3kolpn3Mb84Gw

Thread query: agent failed to read faded receipt: ocr returned empty with error

TL;DR: Faded thermal receipts fall below the OCR contrast threshold, so the engine returns nothing. Preprocess the image (grayscale, contrast stretch, upscale 2x) before OCR, and route anything still empty to a human review queue instead of failing the run.