# screening agent auto-rejected resumes that were scanned images - the OCR fallback never ran

## TL;DR

The parser extracted zero text from scanned-image resumes and the agent treated 'no text' as 'no qualifications'. Add an OCR fallback: when text extraction yields nothing, run OCR before scoring. 'Unreadable' is a processing failure, not a candidate failure.

```text
screening agent auto-rejected scanned-image resumes - OCR fallback never ran
```

## Steps

1. Confirm the pattern: check rejected resumes for zero extracted text.
   Expected: The rejects cluster on image-only PDFs.

2. Add the fallback: if extracted text is below a threshold, route the resume through OCR, then parse the OCR output.
   Expected: Image resumes become readable.

3. Re-screen the wrongly rejected candidates from the OCR output.
   Expected: Qualified candidates get a second chance.

4. Log extraction method per resume (text vs OCR) so future issues are diagnosable.
   Expected: The pipeline is observable.

5. Set a minimum-text gate: never score a resume with near-zero extracted text without OCR first.
   Expected: Silent rejects stop.

## When to use

- image-only resumes get rejected
- parsed text is empty for some resumes
- the parser has no OCR path

## When not to use

- the resume is genuinely empty or corrupt (OCR will not help)
- OCR ran and the resume is still unqualified (legitimate reject)
- all resumes are digital-native PDFs (no fallback needed)

## Tool compatibility

- any resume parser
- OCR: document-processing-ocr services

## Variant phrasings

- "screening rejected scanned resumes"
- "OCR fallback never ran resume"
- "image PDF resume auto-reject"

## Why it happens

Text extraction from PDFs returns nothing for scanned images, and the pipeline treated empty extraction as an empty resume. The agent scored nothing and rejected confidently. The failure was in the pipeline's input handling, but the candidate paid for it.

## Edge cases

- OCR quality varies; low-confidence OCR should flag for human review, not auto-score
- handwritten sections still defeat OCR; have a human path
- the minimum-text threshold needs tuning per resume format

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_S379ksbZlW6JIGSCyecwBw