developer-tools

pymupdf4llm

Name: pymupdf4llm
Availability: InStock
Author: AI Demos

Open-source PDF-to-markdown for clean native PDFs, but unreliable on scans, dense tables, and images.

Visit pymupdf4llm

Open sourceNative PDFsFinancial tablesScanned pages

TL;DR — our verdictUpdated July 2026 · 18 test artifacts

Strong on clean native PDFs, but brittle on scans and dense layouts

Where it wins

You are converting clean, native-digital PDFs locally and want usable markdown without a hosted service.
Your priority is readable prose and straightforward tables from well-structured pages.
You can tolerate charts being reduced to OCR text and images being dropped.

Main limitation

Your PDFs are scanned or OCR-heavy multi-column documents.

Pricing (verified plans)

Open source $0

Strongest test artifacts

Native Text Extraction →Document Structure and Reading-Order Preservation →Table Reconstruction →

Our take

PyMuPDF4LLM is a solid local parser when the PDF is already clean and native-digital: it preserved prose, hierarchy, and straightforward tables well on the hybrid annual report, and it handled the narrative sections in the other digital report too. But its behavior is sharply bimodal rather than robust — dense financial statements, scanned multi-column pages, charts, and embedded images/logos/signatures all degraded into omissions, garbled picture-text blocks, or fragmented OCR text. For this use case, it is useful only when the source is already well-formed and text-heavy.

Walkthrough of PyMuPDF4LLM on the table-heavy financial-report input.

In-Depth Review

Our detailed analysis of pymupdf4llm — features, performance, and real-world testing.

AI Demos Team

Expert Reviewer

Verified Review

Feature-by-Feature Breakdown

Native Text Extraction

Reliable on clean native-digital prose, but OCR-heavy scanned pages become error-prone.

▾

Test Summary

Feature tested: Native Text Extraction

Result: Passed — Reliable on clean native-digital prose, but OCR-heavy scanned pages become error-prone.

Feature tested: Native Text Extraction

Result: Passed

Verdict: Reliable on clean native-digital prose, but OCR-heavy scanned pages become error-prone.

Expected behavior: Extracts readable prose from native-digital report pages and keeps it in a coherent reading sequence. It handled the Target growth-story spread and the Sumitomo operating summary cleanly, but the scanned research paper introduced OCR spelling errors and column interleaving.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_growth_section.png

Observed output: Output artifact (Image): The Target growth-story page was extracted cleanly, with prose preserved in reading order and no visible fragmentation. — pymupdf4llm_input1_hybridearnings_growth_section_parsed_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_growth_section.png

Output artifact: Output artifact (Image): The Target growth-story page was extracted cleanly, with prose preserved in reading order and no visible fragmentation. — pymupdf4llm_input1_hybridearnings_growth_section_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_summary_section.png

Observed output: Output artifact (Image): The operating-performance summary was extracted as readable prose, with the section flow preserved. — pymupdf4llm_input2_financialpdf_summary_section_parsed_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_summary_section.png

Output artifact: Output artifact (Image): The operating-performance summary was extracted as readable prose, with the section flow preserved. — pymupdf4llm_input2_financialpdf_summary_section_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_multicolumn_text.png

Observed output: Output artifact (Image): The scanned multi-column page produced OCR errors and interleaved text, so the prose was no longer cleanly readable. — pymupdf4llm_input3_scannedpaper_multicolumn_text_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_multicolumn_text.png

Output artifact: Output artifact (Image): The scanned multi-column page produced OCR errors and interleaved text, so the prose was no longer cleanly readable. — pymupdf4llm_input3_scannedpaper_multicolumn_text_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Best on clean, native text; not dependable once OCR and multi-column scanning enter the picture.

Extracts readable prose from native-digital report pages and keeps it in a coherent reading sequence. It handled the Target growth-story spread and the Sumitomo operating summary cleanly, but the scanned research paper introduced OCR spelling errors and column interleaving.

image

Input artifact for "Native Text Extraction" test: Input, pymupdf4llm_input1_hybridearnings_growth_section.png

↓→

image

Output artifact for "Native Text Extraction" test: The Target growth-story page was extracted cleanly, with prose preserved in reading order and no visible fragmentation., pymupdf4llm_input1_hybridearnings_growth_section_parsed_success.png

The Target growth-story page was extracted cleanly, with prose preserved in reading order and no visible fragmentation.

image

Input artifact for "Native Text Extraction" test: Input, pymupdf4llm_input2_financialpdf_summary_section.png

↓→

image

Output artifact for "Native Text Extraction" test: The operating-performance summary was extracted as readable prose, with the section flow preserved., pymupdf4llm_input2_financialpdf_summary_section_parsed_success.png

The operating-performance summary was extracted as readable prose, with the section flow preserved.

image

Input artifact for "Native Text Extraction" test: Input, pymupdf4llm_input3_scannedpaper_multicolumn_text.png

↓→

image

Output artifact for "Native Text Extraction" test: The scanned multi-column page produced OCR errors and interleaved text, so the prose was no longer cleanly readable., pymupdf4llm_input3_scannedpaper_multicolumn_text_parsed_failure.png

The scanned multi-column page produced OCR errors and interleaved text, so the prose was no longer cleanly readable.

Bottom Line

Best on clean, native text; not dependable once OCR and multi-column scanning enter the picture.

Document Structure and Reading-Order Preservation

Section hierarchy holds on straightforward digital pages, but scanned layouts lose heading markup and logical flow.

▾

Test Summary

Feature tested: Document Structure and Reading-Order Preservation

Result: Passed — Section hierarchy holds on straightforward digital pages, but scanned layouts lose heading markup and logical flow.

Feature tested: Document Structure and Reading-Order Preservation

Result: Passed

Verdict: Section hierarchy holds on straightforward digital pages, but scanned layouts lose heading markup and logical flow.

Expected behavior: Preserves headings and section nesting on straightforward digital report pages, including business and summary sections. On scanned pages, headings lose markdown-level structure, the Table of Contents can disappear, and multi-column pages no longer read in a clean top-to-bottom order.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_business_section.png

Observed output: Output artifact (Image): The business section kept its section structure and stayed readable in the extracted markdown. — pymupdf4llm_input1_hybridearnings_business_section_parsed_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_business_section.png

Output artifact: Output artifact (Image): The business section kept its section structure and stayed readable in the extracted markdown. — pymupdf4llm_input1_hybridearnings_business_section_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_summary_section.png

Observed output: Output artifact (Image): The summary section preserved a clear heading hierarchy in the extracted text. — pymupdf4llm_input2_financialpdf_summary_section_parsed_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_summary_section.png

Output artifact: Output artifact (Image): The summary section preserved a clear heading hierarchy in the extracted text. — pymupdf4llm_input2_financialpdf_summary_section_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Text prompt

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_toc_omitted.png

Observed output: Output artifact (Text prompt): Output

Input artifact: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_toc_omitted.png

Output artifact: Output artifact (Text prompt): Output

What changed: Image transformed into Text prompt

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_study_area_header.png

Observed output: Output artifact (Image): The scanned heading lost markdown heading markers, and the surrounding text was no longer cleanly separated from the layout. — pymupdf4llm_input3_scannedpaper_study_area_header_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_study_area_header.png

Output artifact: Output artifact (Image): The scanned heading lost markdown heading markers, and the surrounding text was no longer cleanly separated from the layout. — pymupdf4llm_input3_scannedpaper_study_area_header_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Works well for clean report structure, but scanned headings and TOCs are not preserved reliably.

Preserves headings and section nesting on straightforward digital report pages, including business and summary sections. On scanned pages, headings lose markdown-level structure, the Table of Contents can disappear, and multi-column pages no longer read in a clean top-to-bottom order.

image

Input artifact for "Document Structure and Reading-Order Preservation" test: Input, pymupdf4llm_input1_hybridearnings_business_section.png

↓→

image

Output artifact for "Document Structure and Reading-Order Preservation" test: The business section kept its section structure and stayed readable in the extracted markdown., pymupdf4llm_input1_hybridearnings_business_section_parsed_success.png

The business section kept its section structure and stayed readable in the extracted markdown.

image

Input artifact for "Document Structure and Reading-Order Preservation" test: Input, pymupdf4llm_input2_financialpdf_summary_section.png

↓→

image

Output artifact for "Document Structure and Reading-Order Preservation" test: The summary section preserved a clear heading hierarchy in the extracted text., pymupdf4llm_input2_financialpdf_summary_section_parsed_success.png

The summary section preserved a clear heading hierarchy in the extracted text.

image

Input artifact for "Document Structure and Reading-Order Preservation" test: Input, pymupdf4llm_input2_financialpdf_toc_omitted.png

↓→

OUTPUT

The Table of Contents page was omitted from the extraction and replaced by garbled picture-text blocks.

image

Input artifact for "Document Structure and Reading-Order Preservation" test: Input, pymupdf4llm_input3_scannedpaper_study_area_header.png

↓→

image

Output artifact for "Document Structure and Reading-Order Preservation" test: The scanned heading lost markdown heading markers, and the surrounding text was no longer cleanly separated from the layout., pymupdf4llm_input3_scannedpaper_study_area_header_parsed_failure.png

The scanned heading lost markdown heading markers, and the surrounding text was no longer cleanly separated from the layout.

Bottom Line

Works well for clean report structure, but scanned headings and TOCs are not preserved reliably.

Table Reconstruction

Clean native tables can be rebuilt, but dense financial statements and scanned tables break badly.

▾

Test Summary

Feature tested: Table Reconstruction

Result: Passed — Clean native tables can be rebuilt, but dense financial statements and scanned tables break badly.

Feature tested: Table Reconstruction

Result: Passed

Verdict: Clean native tables can be rebuilt, but dense financial statements and scanned tables break badly.

Expected behavior: Reconstructs well-structured native-digital tables into readable markdown. The hybrid annual report table came through cleanly, while dense financial statements and scanned tables lost alignment, header associations, or the table itself.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_financial_table.png

Observed output: Output artifact (Image): The financial summary table stayed aligned, with year/value associations preserved in readable markdown. — pymupdf4llm_input1_hybridearnings_financial_table_parsed_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_financial_table.png

Output artifact: Output artifact (Image): The financial summary table stayed aligned, with year/value associations preserved in readable markdown. — pymupdf4llm_input1_hybridearnings_financial_table_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_quarterly_table.png

Observed output: Output artifact (Image): The quarterly balance-sheet table collapsed into garbled fragments, losing the table structure. — pymupdf4llm_input2_financialpdf_quarterly_table_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_quarterly_table.png

Output artifact: Output artifact (Image): The quarterly balance-sheet table collapsed into garbled fragments, losing the table structure. — pymupdf4llm_input2_financialpdf_quarterly_table_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_table_6.png

Observed output: Output artifact (Image): The scanned table merged narrative text into the table area and broke the column layout. — pymupdf4llm_input3_scannedpaper_table_6_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_table_6.png

Output artifact: Output artifact (Image): The scanned table merged narrative text into the table area and broke the column layout. — pymupdf4llm_input3_scannedpaper_table_6_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Good on straightforward native tables, unreliable on dense financial or scanned table layouts.

Reconstructs well-structured native-digital tables into readable markdown. The hybrid annual report table came through cleanly, while dense financial statements and scanned tables lost alignment, header associations, or the table itself.

image

Input artifact for "Table Reconstruction" test: Input, pymupdf4llm_input1_hybridearnings_financial_table.png

↓→

image

Output artifact for "Table Reconstruction" test: The financial summary table stayed aligned, with year/value associations preserved in readable markdown., pymupdf4llm_input1_hybridearnings_financial_table_parsed_success.png

The financial summary table stayed aligned, with year/value associations preserved in readable markdown.

image

Input artifact for "Table Reconstruction" test: Input, pymupdf4llm_input2_financialpdf_quarterly_table.png

↓→

image

Output artifact for "Table Reconstruction" test: The quarterly balance-sheet table collapsed into garbled fragments, losing the table structure., pymupdf4llm_input2_financialpdf_quarterly_table_parsed_failure.png

The quarterly balance-sheet table collapsed into garbled fragments, losing the table structure.

image

Input artifact for "Table Reconstruction" test: Input, pymupdf4llm_input3_scannedpaper_table_6.png

↓→

image

The scanned table merged narrative text into the table area and broke the column layout.

Bottom Line

Good on straightforward native tables, unreliable on dense financial or scanned table layouts.

Chart and Figure Text Capture

It can mine text from figures, but it does not preserve chart geometry or visual structure.

▾

Test Summary

Feature tested: Chart and Figure Text Capture

Result: Passed — It can mine text from figures, but it does not preserve chart geometry or visual structure.

Feature tested: Chart and Figure Text Capture

Result: Passed

Verdict: It can mine text from figures, but it does not preserve chart geometry or visual structure.

Expected behavior: Recovers chart labels, titles, and numbers as OCR text inside picture-text wrappers. On the annual report charts it kept values and labels, while scanned figures degraded into fragmented OCR text blocks.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_sales_and_ebit_chart.png

Observed output: Output artifact (Image): The bar chart's labels and values were recovered as text, but the bar graphics were lost. — pymupdf4llm_input1_hybridearnings_sales_and_ebit_chart_partial_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_sales_and_ebit_chart.png

Output artifact: Output artifact (Image): The bar chart's labels and values were recovered as text, but the bar graphics were lost. — pymupdf4llm_input1_hybridearnings_sales_and_ebit_chart_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_comparison_chart.png

Observed output: Output artifact (Image): The line chart title and legend survived only as fragmented OCR text, with no reconstructed plot geometry. — pymupdf4llm_input1_hybridearnings_comparison_chart_partial_success.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_comparison_chart.png

Output artifact: Output artifact (Image): The line chart title and legend survived only as fragmented OCR text, with no reconstructed plot geometry. — pymupdf4llm_input1_hybridearnings_comparison_chart_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_figure_4.png

Observed output: Output artifact (Image): The scanned figure degraded into broken OCR text, so the chart itself was not reconstructed. — pymupdf4llm_input3_scannedpaper_figure_4_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_figure_4.png

Output artifact: Output artifact (Image): The scanned figure degraded into broken OCR text, so the chart itself was not reconstructed. — pymupdf4llm_input3_scannedpaper_figure_4_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Useful if you only need figure text; not a faithful chart reconstruction tool.

Recovers chart labels, titles, and numbers as OCR text inside picture-text wrappers. On the annual report charts it kept values and labels, while scanned figures degraded into fragmented OCR text blocks.

image

Input artifact for "Chart and Figure Text Capture" test: Input, pymupdf4llm_input1_hybridearnings_sales_and_ebit_chart.png

↓→

image

The bar chart's labels and values were recovered as text, but the bar graphics were lost.

image

Input artifact for "Chart and Figure Text Capture" test: Input, pymupdf4llm_input1_hybridearnings_comparison_chart.png

↓→

image

Output artifact for "Chart and Figure Text Capture" test: The line chart title and legend survived only as fragmented OCR text, with no reconstructed plot geometry., pymupdf4llm_input1_hybridearnings_comparison_chart_partial_success.png

The line chart title and legend survived only as fragmented OCR text, with no reconstructed plot geometry.

image

Input artifact for "Chart and Figure Text Capture" test: Input, pymupdf4llm_input3_scannedpaper_figure_4.png

↓→

image

Output artifact for "Chart and Figure Text Capture" test: The scanned figure degraded into broken OCR text, so the chart itself was not reconstructed., pymupdf4llm_input3_scannedpaper_figure_4_parsed_failure.png

The scanned figure degraded into broken OCR text, so the chart itself was not reconstructed.

Bottom Line

Useful if you only need figure text; not a faithful chart reconstruction tool.

Embedded Image Retention

Standalone visuals are a clear weak spot: they are dropped or turned into OCR text instead of being retained as images.

▾

Test Summary

Feature tested: Embedded Image Retention

Result: Passed — Standalone visuals are a clear weak spot: they are dropped or turned into OCR text instead of being retained as images.

Feature tested: Embedded Image Retention

Result: Passed

Verdict: Standalone visuals are a clear weak spot: they are dropped or turned into OCR text instead of being retained as images.

Expected behavior: Retains standalone visuals such as portraits, logos, signatures, and header art. The Target CEO photo and signature disappeared, and the Sumitomo and USDA page headers were flattened into garbled repeated OCR text instead of being kept as images.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_ceo_image.png

Observed output: Output artifact (Image): The CEO photo vanished with no image placeholder or retained image element. — pymupdf4llm_input1_hybridearnings_ceo_image_omitted_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_ceo_image.png

Output artifact: Output artifact (Image): The CEO photo vanished with no image placeholder or retained image element. — pymupdf4llm_input1_hybridearnings_ceo_image_omitted_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_ceo_sign.png

Observed output: Output artifact (Image): The handwritten signature was omitted, leaving only surrounding text and the printed title. — pymupdf4llm_input1_hybridearnings_ceo_sign_omitted_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input1_hybridearnings_ceo_sign.png

Output artifact: Output artifact (Image): The handwritten signature was omitted, leaving only surrounding text and the printed title. — pymupdf4llm_input1_hybridearnings_ceo_sign_omitted_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_company_header.png

Observed output: Output artifact (Image): The company header/logo was converted into repeated garbled text rather than retained as a visual asset. — pymupdf4llm_input2_financialpdf_company_header_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input2_financialpdf_company_header.png

Output artifact: Output artifact (Image): The company header/logo was converted into repeated garbled text rather than retained as a visual asset. — pymupdf4llm_input2_financialpdf_company_header_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_usda_logo.png

Observed output: Output artifact (Image): The USDA header/logo area was flattened into corrupted OCR text instead of being kept as an image. — pymupdf4llm_input3_scannedpaper_usda_logo_parsed_failure.png

Input artifact: Input artifact (Image): Input — pymupdf4llm_input3_scannedpaper_usda_logo.png

Output artifact: Output artifact (Image): The USDA header/logo area was flattened into corrupted OCR text instead of being kept as an image. — pymupdf4llm_input3_scannedpaper_usda_logo_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: If visual assets matter, this is not the right parser.

Retains standalone visuals such as portraits, logos, signatures, and header art. The Target CEO photo and signature disappeared, and the Sumitomo and USDA page headers were flattened into garbled repeated OCR text instead of being kept as images.

image

↓→

image

The CEO photo vanished with no image placeholder or retained image element.

image

↓→

image

The handwritten signature was omitted, leaving only surrounding text and the printed title.

image

↓→

image

The company header/logo was converted into repeated garbled text rather than retained as a visual asset.

image

↓→

image

The USDA header/logo area was flattened into corrupted OCR text instead of being kept as an image.

Bottom Line

If visual assets matter, this is not the right parser.

Caption and Footnote Association

Footnotes stay attached on clean tables, but scanned captions can be absorbed into the wrong block.

▾

Test Summary

Feature tested: Caption and Footnote Association

Result: Passed — Footnotes stay attached on clean tables, but scanned captions can be absorbed into the wrong block.

Keeps footnotes grouped under clean native tables, but scanned figure/table captions can be folded into picture-text blocks or merged directly into table cells. The hybrid annual report preserved the table footnotes cleanly, while the scanned research paper merged the Table 6 caption into the table.

image

Input artifact for "Caption and Footnote Association" test: Input, pymupdf4llm_input1_hybridearnings_financial_table_footnotes.png

↓→

image

Footnotes (a) through (i) stayed directly associated with the financial summary table.

image

Input artifact for "Caption and Footnote Association" test: Input, pymupdf4llm_input3_scannedpaper_table_6_caption.png

↓→

image

Output artifact for "Caption and Footnote Association" test: The Table 6 caption was merged into the extracted table rather than remaining separate from it., pymupdf4llm_input3_scannedpaper_table_6_caption_parsed_failure.png

The Table 6 caption was merged into the extracted table rather than remaining separate from it.

Bottom Line

Precise on straightforward table footnotes, brittle on scanned captions and caption/table boundaries.

Pricing

Open source, no paid tiers evaluated.

TESTED

Open source

AGPL open-source library

The report describes pymupdf4llm as free and open source (AGPL).

✓ Use This If

●You are converting clean, native-digital PDFs locally and want usable markdown without a hosted service.

●Your priority is readable prose and straightforward tables from well-structured pages.

●You can tolerate charts being reduced to OCR text and images being dropped.

✕ Skip This If

●Your PDFs are scanned or OCR-heavy multi-column documents.

●You need faithful retention of charts, photos, logos, or signatures.

●You rely on dense financial statement tables or validated repeat-run consistency.

developer-toolsdocumentation-toolstext

Very well in this report. The hybrid annual report's prose, the business section, the growth-story section, and the summary section were extracted cleanly, with readable reading order and preserved structure on the digital pages.

Poorly. In the financial-report input, the consolidated and quarterly tables were either omitted or reduced to garbled picture-text fragments, and the structure was not recoverable.

Not reliably. The scanned research paper showed OCR spelling errors, heavy column interleaving, weak heading parsing, and degraded tables across the document.

Their text can survive, but the chart itself is not reconstructed. The annual report charts were reduced to OCR text fragments inside picture-text blocks, and scanned figures degraded into broken text.

No. The CEO photo and signature were omitted, and the company header and USDA logo areas were turned into garbled OCR text instead of retained image elements.

Footnotes on clean native tables were preserved and stayed attached below the table, but scanned table captions could be merged into the table extraction instead of remaining separate.

The report describes it as free/open-source (AGPL). It does not include a repeat-run or matched-layout consistency test, so stability across repeated runs remains unproven.