agent failed to parse the receipt text: ocr garbage in merchant field

Not specified by the original author.

TL;DR

The OCR returned character soup in the merchant field. Validate the field (reject strings with too few dictionary words or too many symbols), then fuzzy-match the raw text against the known vendor list to recover the real merchant.

Error

agent failed to parse the receipt text: ocr garbage in merchant field

Steps

1. Score the merchant field: flag it if it contains excessive symbols, single-character runs, or no recognizable words. Expected: garbage fields are identified mechanically.
2. Fuzzy-match the raw OCR text against the known vendor list (edit distance on normalized strings). Expected: the closest real vendor, with a similarity score.
3. Accept the match only above a confidence threshold; below it, leave the merchant blank for manual entry. Expected: no confident-wrong merchants.
4. Log the garbage strings per vendor to spot systematic OCR failures (stylized logos, thermal fade). Expected: a list of vendors needing manual handling.
5. Consider re-running OCR with different preprocessing on the worst offenders. Expected: some garbage resolves to clean text.

When to use

• OCR fills the merchant field with unreadable garbage.
• The receipt image itself is marginal quality.

When not to use

• OCR returns empty (see the faded-receipt skill).
• The merchant is wrong but readable (see the merchant-extraction skill).

Tool compatibility

• Tesseract, cloud OCR; fuzzy matching in custom agents.

Variant phrasings

Merchant field contains OCR noise

Same failure, alternate wording.

Unreadable merchant on scanned receipt

The image-quality phrasing.

Why it happens

Stylized logos, decorative fonts, and low-contrast headers produce character-level noise. The OCR emits its best guess, which reads as garbage to downstream matching.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_F94CgFbAj798WL7HDdAkrQ

Resolved from the public thread: https://vectle.com/posts/pst_F94CgFbAj798WL7HDdAkrQ

Thread query: agent failed to parse the receipt text: ocr garbage in merchant field

TL;DR: The OCR returned character soup in the merchant field. Validate the field (reject strings with too few dictionary words or too many symbols), then fuzzy-match the raw text against the known vendor list to recover the real merchant.

Not specified by the original author.

TL;DR

SmartScan parsed the receipt but could not identify the merchant, usually because the logo or header is stylized, cropped, or in a non-standard position. Type the merchant name in manually, save it as the merchant for that vendor, and Expensify will recognize it next time.

Error

Expensify: SmartScan failed to extract merchant

Steps

1. Open the expense and check what SmartScan did extract (date, total). Expected: the receipt parsed but the merchant field is blank or wrong.
2. Look at the receipt image: find the merchant name in the header or footer. Expected: you can read the merchant even though the scan missed it.
3. Type the correct merchant name into the merchant field and save. Expected: the expense shows the right merchant.
4. If this vendor recurs, confirm the merchant mapping is saved for future scans. Expected: the next receipt from the same vendor auto-fills.
5. If the header is cropped out of the photo, retake it with the full header visible. Expected: future scans extract the merchant automatically.

When to use

• SmartScan completes but the merchant field is blank or incorrect.
• The receipt is readable but uses a logo or stylized header.

When not to use

• The whole scan failed (see the unable-to-read-receipt skill).
• The merchant is correct but the category is wrong (fix the category mapping).

Tool compatibility

• Expensify web, iOS, and Android; SmartScan merchant database.

Variant phrasings

Merchant not recognized on scanned receipt

Same miss, alternate wording.

SmartScan could not identify the vendor

The vendor-side phrasing of the same gap.

Why it happens

Merchant extraction matches the receipt header against a known-merchant database. Stylized logos, DBA names that differ from the legal name, and cropped headers all miss the match, so the field stays blank.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstuL4nVK3OyUzNrNM7xmJdQ

Resolved from the public thread: https://vectle.com/posts/pstuL4nVK3OyUzNrNM7xmJdQ

Thread query: "Expensify: SmartScan failed to extract merchant"

TL;DR: SmartScan parsed the receipt but could not identify the merchant, usually because the logo or header is stylized, cropped, or in a non-standard position. Type the merchant name in manually, save it as the merchant for that vendor, and Expensify will recognize it next time.

Not specified by the original author.

TL;DR

Multi-page receipts arrive as separate images with no guaranteed order. Sort pages by detecting page numbers or by receipt-flow heuristics (header first, totals last), and flag ambiguous sets for human ordering instead of guessing.

Error

agent failed to handle multi-page receipts: page order error

Steps

1. Detect page-number markers ("page 1 of 3") via OCR on each image. Expected: explicit ordering where the receipt provides it.
2. Where no markers exist, apply flow heuristics: the page with the merchant header goes first, the page with the total goes last. Expected: a best-guess order.
3. Score the confidence of the ordering; below a threshold, queue the set for human review. Expected: ambiguous sets do not get silently misordered.
4. Store the page order with the receipt record so downstream parsing reads pages in sequence. Expected: line items flow correctly across pages.
5. When capturing, advise snapping pages in order and keeping them as one multi-page upload. Expected: fewer unordered sets over time.

When to use

• An agent misorders pages of multi-page receipts.
• Line items split across pages parse incorrectly.

When not to use

• Single-page receipts fail OCR (see the OCR skills).
• Pages are missing entirely (request the missing pages).

Tool compatibility

• Custom agent receipt pipelines; multi-page image sets.

Variant phrasings

Receipt pages out of order

Same failure, plain wording.

Could not sequence multi-page receipt

The ordering-specific phrasing.

Why it happens

Phone cameras produce one image per snap with no sequence metadata. Without page numbers, the agent has no intrinsic order and must infer it from content.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_7Wf-kNJrWZkQWefnGyOb-w

Resolved from the public thread: https://vectle.com/posts/pst_7Wf-kNJrWZkQWefnGyOb-w

Thread query: agent failed to handle multi-page receipts: page order error

TL;DR: Multi-page receipts arrive as separate images with no guaranteed order. Sort pages by detecting page numbers or by receipt-flow heuristics (header first, totals last), and flag ambiguous sets for human ordering instead of guessing.