Skip to content

0.2.4's needsOcr gate refuses born-digital PDFs that 0.2.3 converts correctly #162

Open

Description

@mshresponse

Versions: @firecrawl/anydoc 0.2.4 (integrity sha512-rfJxa5L+…P1Og==) vs 0.2.3; macOS arm64,
native binary anydoc.darwin-arm64.node sha256 97477757…f0d5. CLI, --ocr reject (the default).

Summary. Since 0.2.4 (27 Aug 2026), anydoc exits 3 (needs OCR) with zero output on
born-digital PDFs whose text layer is complete. 0.2.3 converts the same files end to end with
exit 0. The refusal is document-level and all-or-nothing, so one page condemns the whole file.

Corpus. 83 PDFs (30 born-digital, 53 scans). 0.2.3: 32 converted, 51 refused
(exit 1). 0.2.4: 19 converted, 64 refused — the 51 scans plus 13 born-digital documents.

The 13, with the page 0.2.4 objects to (all readable by pdftotext / pdf.js):
census_p60-280_income 5–7/67 · crm_analytics 3, 1365, 1613, 2117–8/2129 · health 3–5/1417 ·
marketing_cloud_next 3/460 · netzero_cloud_dev_guide 1314/1349 · omnistudio 3–5, 89/813 ·
reports_and_dashboards 3/560 · rfc9110_http_semantics 21/194 · salesforce_industries 3/26 ·
salesforce_release_notes (Winter '27, Prince 15.4) 3/1010 · salesforce_spring25 3/1101 ·
salesforce_summer25 3/1160 · tableau 3/18.

In eleven of them the objected page is the table of contents (dot-leader lines). 0.2.3 renders
that page as a normal contents list — 2.43 MB of markdown for the 1,010-page document.

Expected: a page with a text layer is converted; if a classifier deems it low-confidence, emit
the text and report pages_needing_ocr rather than refusing the document.
Actual: exit 3, no output, for the entire document.

Repro: anydoc --ocr reject <file>.pdf on any of the above (Salesforce release notes are
public downloads). Compare with 0.2.3.

Impact: any consumer running local-only (no hosted OCR) loses 13 of the 30 born-digital documents
between 0.2.3 and 0.2.4 for no gain in fidelity.


Edited 3 Sep 2026: corpus split corrected (30 born-digital / 53 scans, was 32/51) and the table-of-contents count corrected upward from ten to eleven. See the comment below for detail; no other figure changed.

Activity

  1. mshresponse commented on Sep 3, 2026

    @mshresponse
    Author

    Correction to two figures in the report above. Both slips make the impact marginally larger rather than smaller, and the substance of the report is unchanged.

    Corpus split. "83 PDFs (32 with a full text layer, 51 scans)" should read 30 born-digital and 53 scans. The 32 is the number 0.2.3 converts — which includes 2 scans it manages anyway — not the born-digital count.

    Impact line. "13/32 born-digital documents" is therefore 13 of the 30 born-digital documents, or 13 of the 32 that 0.2.3 converts, depending on which denominator is wanted.

    Table-of-contents count. "In ten of them" was conservative. Checking all 13, the objected page is a table of contents in 11. The two exceptions are netzero_cloud_dev_guide p1314 (a data table) and rfc9110_http_semantics p21 (prose) — both still born-digital with a complete text layer.

    Everything else stands as filed: the list of 13 documents and the pages objected to, the 0.2.3 vs 0.2.4 conversion counts (32 vs 19 of the same 83), and the repro.

  2. Starosdev commented on Sep 19, 2026

    @Starosdev

    Another data point for this regression, from a different kind of page: a fill-in form made mostly of underscore blank lines, rather than a table of contents with dot leaders.

    Versions: @firecrawl/anydoc 0.2.4 CLI (default --ocr reject) against 0.2.3; macOS arm64. pdftotext is poppler.

    Document: Florida HSMV procedure TL-26, 31 pages, public at https://www.flhsmv.gov/pdf/proc/tl/tl-26.pdf. The original Acrobat PDF converts under 0.2.4 with exit 0. The failing file is the same document re-rendered by macOS Safari's PDF export (Producer macOS Version 15.7.4 Quartz PDFContext).

    Result on the Quartz copy:

    • Whole document: anydoc: page 29 of 31 needs OCR, exit 3, no output. pdftotext extracts 11,642 words across all 31 pages.
    • Page 29 alone, split out with pdfseparate -f 29 -l 29: 0.2.4 prints page 1 of 1 needs OCR and exits 3. 0.2.3 exits 0 and emits 806 words. pdftotext extracts 753 words.
    • Page 29 has no images (pdfimages -list is empty). All five fonts are embedded TrueType with WinAnsi encoding and a ToUnicode map (pdffonts: uni yes).
    • The page is "Exhibit A, Notice of Claim of Lien": a form whose lines are mostly underscores (Name: ______________), plus several full prose paragraphs.
    • Pages 28 and 30, which have the same fonts and producer, convert under 0.2.4. Page 28 even includes a MacRoman font with no ToUnicode map, so a missing Unicode mapping does not explain page 29.

    As with the table-of-contents pages above, a page made mostly of runs of one repeated glyph appears to be classified as needing OCR even though its text layer is complete, and the document-level refusal (#144) then discards the other 30 pages. The expected behavior matches the report above: emit the text and list the low-confidence pages.

    I can't attach the Quartz file, because it carries personal PDF metadata. It should reproduce by opening the public URL in Safari, choosing File > Export as PDF, and running anydoc on the result.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions