Large or Unreadable PDFs

Scenario

Pull only the passages about payment terms and acceptance out of a report that runs to dozens of pages. Nothing is dumped into the context: summarize measures the document first, search_text finds the hit pages, and read_text reads only those. A scan-like page with no text layer is not reported as "no text" — it is read from the render_page image instead. No OCR is performed. What was read and what could not be read is declared in a Read Report.

Below is a real measurement from 2026-09-16 (pdf-reader-mcp v0.15.1 / pdf-read Skill v0.2.2; host Grok Build 1.0.30; specimens: a self-made 21-page report and a scan-like PDF containing only an image XObject).

Cast

Actor	Role
pdf-read Skill	measure → branch → choose the route → Read Report. Enters search_text for passage extraction even when next is empty
pdf-reader	summarize, search_text, read_text (explicit pages), render_page

Sequence

Prompt examples

• "From this report, extract only the passages about payment terms and acceptance. If a page can't be read, say why"
• "It's 500 pages — read only what's needed, don't dump the whole thing"
• "Read this scanned contract. Don't OCR it"

Measured example — a 21-page report (narrowing route)

summarize returned pageCount: 21, isTagged: false, isEncrypted: false, textExtractability: "extracted", unreadablePages: []. next was empty (the reader suggests search_text only above 50 pages). Skill v0.2.2 enters search_text anyway when the request is passage extraction.

search_text query	totalMatches	pages hit
支払条件 (payment terms)	2	21
検収 (acceptance)	3	21
再委託の禁止 (no subcontracting)	2	21

read_text was called once, with pages: "21". Pages 1–20 were not read — not unreadable, simply not read.

Calls — summarize / search_text / read_text

{ "file_path": "/absolute/path/to/long-report.pdf", "response_format": "json" }

{ "pageCount": 21, "isTagged": false, "isEncrypted": false,
  "textExtractability": "extracted", "unreadablePages": [], "next": [] }

{ "file_path": "/absolute/path/to/long-report.pdf", "query": "検収", "response_format": "json" }

{ "totalMatches": 3, "pages": [21], "truncated": false }

{ "file_path": "/absolute/path/to/long-report.pdf", "pages": "21", "compact_whitespace": true }

Read Report (as declared by the Skill):

• Target: long-report.pdf (21 pages)
• Range read: 21 (search_text hits; 1–20 are filler and were not read) / route: narrowing
• Text extractability: extracted (unreadablePages empty)
• Could not read: none
• Truncation: none (search_text truncated=false)

Measured example — a page with no text layer (image route)

summarize returned textExtractability: "no_text_layer", hasText: false, imageCount: 1, textShowingOperators: 0, imageOperators: 1, and listed render_page under next. read_text returned an empty text with extractability.state: "no_text_layer". The empty string is not an extraction failure; it is the declaration that there is no text layer.

render_page (pages "1", jpeg, 150 dpi) returned a 136,247-byte image, and the clauses were readable from the pixels. No OCR was run.

Call — render_page

{ "file_path": "/absolute/path/to/scan-no-text-layer.pdf", "pages": "1", "format": "jpeg", "dpi": 150 }

pages is required. Images come back per page.

Read Report:

• Target: scan-no-text-layer.pdf (1 page)
• Range read: 1 / route: image (visual reading of the render_page output)
• Text extractability: no_text_layer (read_text's text is empty; this is not reported as "no text")
• Note: no OCR was performed. The reading comes from the page image

How to read the results

• Call summarize first. Page count, tagging, encryption and per-page extractability decide the route
• An empty extraction is not evidence of "no text". extractability.state no_text_layer means the text exists only as pixels; not_extractable means it displays but has a font with no route to Unicode (ISO 32000-2 §9.10.1)
• "Not read" and "unreadable" are declared separately. In the 21-page specimen, pages 1–20 were simply not read
• Reading from render_page is not OCR. The Read Report says "image route" and never claims OCR
• The reader's next suggests search_text only above 50 pages. Below that, the Skill enters search_text on its own when the request is passage extraction

Evaluation kit and full report: eval/hosts/grok-build/reports/UC07.md (Japanese)