developer-tools

docling

Name: docling
Availability: InStock
Author: AI Demos

Open-source PDF-to-markdown conversion that is strong on text, headings, and standard tables, but drops charts and other visual assets.

Visit docling

Open sourceScanned OCRTable extractionVisuals dropped

TL;DR — our verdictUpdated July 2026 · 27 test artifacts

Excellent for text and tables, unsafe for visual fidelity

Where it wins

You need a self-hosted, open-source PDF-to-markdown pipeline for reports with lots of text, headings, and standard financial tables.
You want OCR text from scanned pages and can tolerate minor spacing, word-merging, or library-stamp noise.
You need footnotes, captions, and simple form checkboxes preserved in the converted output.

Main limitation

You need charts, figures, logos, signatures, or divider graphics to survive conversion.

Strongest test artifacts

Text Extraction and OCR →Table Reconstruction →Document Hierarchy Preservation →

Our take

Docling is a strong self-hosted PDF-to-markdown parser when your priority is readable text, preserved headings, and solid reconstruction of standard tables. It handled long documents and scanned pages well, but it consistently replaced charts, figures, logos, signatures, and other visual assets with placeholders, and its hardest scan-heavy tables still broke down.

Demo walkthrough of Docling converting a hybrid earnings report to markdown.

In-Depth Review

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

AI Demos Team

Expert Reviewer

Verified Review

Feature-by-Feature Breakdown

Text Extraction and OCR

Strong

▾

Test Summary

Feature tested: Text Extraction and OCR

Result: Passed — Strong

Feature tested: Text Extraction and OCR

Result: Passed

Verdict: Strong

Expected behavior: Docling extracts native PDF text and OCR text from scanned pages with high fidelity overall. In the hybrid earnings report, the table-heavy financial report, and the scanned research paper, the text stayed readable end to end, with only minor spacing/merging noise in tight crops.

Test case: PDF document → Text/code file

Input type: PDF document

Input used: Input artifact (PDF document): Hybrid earnings report combining native text, tables, charts, and a scanned signature page. — input1_hybridearnings.pdf

Observed output: Output artifact (Text/code file): Preserved the report text and section content with only minor spacing issues inside some table cells. — docling_input1_hybridearnings_output.md

Input artifact: Input artifact (PDF document): Hybrid earnings report combining native text, tables, charts, and a scanned signature page. — input1_hybridearnings.pdf

Output artifact: Output artifact (Text/code file): Preserved the report text and section content with only minor spacing issues inside some table cells. — docling_input1_hybridearnings_output.md

What changed: PDF document transformed into Text/code file

Test case: PDF document → Text/code file

Input type: PDF document

Input used: Input artifact (PDF document): Table-heavy financial report with narrative disclosures and quarterly statements. — input2_financialpdf.pdf

Observed output: Output artifact (Text/code file): Preserved the report text cleanly across all 18 pages, including the introductory disclaimer and operating-performance narrative. — docling_input2_financialpdf_output.md

Input artifact: Input artifact (PDF document): Table-heavy financial report with narrative disclosures and quarterly statements. — input2_financialpdf.pdf

Output artifact: Output artifact (Text/code file): Preserved the report text cleanly across all 18 pages, including the introductory disclaimer and operating-performance narrative. — docling_input2_financialpdf_output.md

What changed: PDF document transformed into Text/code file

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Scanned first-page crop showing the title and abstract area of the mountain pine beetle paper. — docling_input3_scannedpaper_page_1.png

Observed output: Output artifact (Image): OCR captured the page text, but the opening page introduced noise such as merged words and library-stamp artifacts, including the injected numbers/letters in the title area. — docling_input3_scannedpaper_page_1_parsed_failure.png

Input artifact: Input artifact (Image): Scanned first-page crop showing the title and abstract area of the mountain pine beetle paper. — docling_input3_scannedpaper_page_1.png

Output artifact: Output artifact (Image): OCR captured the page text, but the opening page introduced noise such as merged words and library-stamp artifacts, including the injected numbers/letters in the title area. — docling_input3_scannedpaper_page_1_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Strong OCR and text preservation across all three documents, with only small scan-related noise rather than silent text loss.

Docling extracts native PDF text and OCR text from scanned pages with high fidelity overall. In the hybrid earnings report, the table-heavy financial report, and the scanned research paper, the text stayed readable end to end, with only minor spacing/merging noise in tight crops.

INPUT

input1_hybridearnings.pdf

Hybrid earnings report combining native text, tables, charts, and a scanned signature page.

↓→

OUTPUT

docling_input1_hybridearnings_output.md

Loading file...

Preserved the report text and section content with only minor spacing issues inside some table cells.

INPUT

input2_financialpdf.pdf

Table-heavy financial report with narrative disclosures and quarterly statements.

↓→

OUTPUT

docling_input2_financialpdf_output.md

Loading file...

Preserved the report text cleanly across all 18 pages, including the introductory disclaimer and operating-performance narrative.

INPUT

Scanned first-page crop showing the title and abstract area of the mountain pine beetle paper.

↓→

OUTPUT

OCR captured the page text, but the opening page introduced noise such as merged words and library-stamp artifacts, including the injected numbers/letters in the title area.

Bottom Line

Strong OCR and text preservation across all three documents, with only small scan-related noise rather than silent text loss.

Table Reconstruction

Mixed but generally strong

▾

Test Summary

Feature tested: Table Reconstruction

Result: Passed — Mixed but generally strong

Feature tested: Table Reconstruction

Result: Passed

Verdict: Mixed but generally strong

Expected behavior: Docling reconstructs standard tables well, including multi-column financial layouts and clean balance-sheet structures. It also preserves many smaller scanned tables, though multi-level headers can flatten or duplicate and dense scan-heavy tables can become garbled.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Financial Summary table from the hybrid earnings report. — docling_input1_hybridearnings_financial_table.png

Observed output: Output artifact (Image): Reconstructed the yearly financial summary as a clean table with values aligned across 2015 through 2011. — docling_input1_hybridearnings_financial_table_parsed_success.png

Input artifact: Input artifact (Image): Financial Summary table from the hybrid earnings report. — docling_input1_hybridearnings_financial_table.png

Output artifact: Output artifact (Image): Reconstructed the yearly financial summary as a clean table with values aligned across 2015 through 2011. — docling_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): Quarterly consolidated balance sheet table from the financial report. — docling_input2_financialpdf_quarterly_table.png

Observed output: Output artifact (Image): Preserved the balance-sheet rows and the two reporting dates in a readable table layout. — docling_input2_financialpdf_quarterly_table_parsed_success.png

Input artifact: Input artifact (Image): Quarterly consolidated balance sheet table from the financial report. — docling_input2_financialpdf_quarterly_table.png

Output artifact: Output artifact (Image): Preserved the balance-sheet rows and the two reporting dates in a readable table layout. — docling_input2_financialpdf_quarterly_table_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Business Results table comparing first quarter FY2025 versus FY2024. — docling_input2_financialpdf_business_results_table.png

Observed output: Output artifact (Image): The table stayed recognizable, but the multi-level headers were duplicated and misaligned, which made column mapping confusing. — docling_input2_financialpdf_business_results_table_parsed_failure.png

Input artifact: Input artifact (Image): Business Results table comparing first quarter FY2025 versus FY2024. — docling_input2_financialpdf_business_results_table.png

Output artifact: Output artifact (Image): The table stayed recognizable, but the multi-level headers were duplicated and misaligned, which made column mapping confusing. — docling_input2_financialpdf_business_results_table_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Scanned Table 1 with stand data and proposed cuts. — docling_input3_scannedpaper_table_1.png

Observed output: Output artifact (Image): Extracted the upper portion of the table with some values preserved, but the structure was only partially recovered. — docling_input3_scannedpaper_table_1_parsed_partial_success.png

Input artifact: Input artifact (Image): Scanned Table 1 with stand data and proposed cuts. — docling_input3_scannedpaper_table_1.png

Output artifact: Output artifact (Image): Extracted the upper portion of the table with some values preserved, but the structure was only partially recovered. — docling_input3_scannedpaper_table_1_parsed_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Scanned Table 4 with trees killed per acre by cutting block, year, cause, and diameter. — docling_input3_scannedpaper_table_4.png

Observed output: Output artifact (Image): The table structure broke badly: headers expanded awkwardly and the row/column layout became unusable. — docling_input3_scannedpaper_table_4_parsed_failure.png

Input artifact: Input artifact (Image): Scanned Table 4 with trees killed per acre by cutting block, year, cause, and diameter. — docling_input3_scannedpaper_table_4.png

Output artifact: Output artifact (Image): The table structure broke badly: headers expanded awkwardly and the row/column layout became unusable. — docling_input3_scannedpaper_table_4_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Excellent on standard tables, but header flattening and scan-heavy complexity can still break the extraction.

Docling reconstructs standard tables well, including multi-column financial layouts and clean balance-sheet structures. It also preserves many smaller scanned tables, though multi-level headers can flatten or duplicate and dense scan-heavy tables can become garbled.

INPUT

Financial Summary table from the hybrid earnings report.

↓→

OUTPUT

Reconstructed the yearly financial summary as a clean table with values aligned across 2015 through 2011.

INPUT

Quarterly consolidated balance sheet table from the financial report.

↓→

OUTPUT

Preserved the balance-sheet rows and the two reporting dates in a readable table layout.

INPUT

Business Results table comparing first quarter FY2025 versus FY2024.

↓→

OUTPUT

The table stayed recognizable, but the multi-level headers were duplicated and misaligned, which made column mapping confusing.

INPUT

Scanned Table 1 with stand data and proposed cuts.

↓→

OUTPUT

Extracted the upper portion of the table with some values preserved, but the structure was only partially recovered.

INPUT

Scanned Table 4 with trees killed per acre by cutting block, year, cause, and diameter.

↓→

OUTPUT

The table structure broke badly: headers expanded awkwardly and the row/column layout became unusable.

Bottom Line

Excellent on standard tables, but header flattening and scan-heavy complexity can still break the extraction.

Document Hierarchy Preservation

Strong

▾

Test Summary

Feature tested: Document Hierarchy Preservation

Result: Passed — Strong

Feature tested: Document Hierarchy Preservation

Result: Passed

Verdict: Strong

Expected behavior: Docling preserves sectioning, headings, and nested list structure so the output reads like a usable markdown document rather than a flat text dump. This held across the annual-report business section, the financial report summary headings, and the scanned paper's structured prescriptions list.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Annual-report section header for Item 1. Business and General. — docling_input1_hybridearnings_item_header.png

Observed output: Output artifact (Image): Preserved the Item 1. Business section as a clear heading with the opening paragraph below it. — docling_input1_hybridearnings_item_header_parsed_success.png

Input artifact: Input artifact (Image): Annual-report section header for Item 1. Business and General. — docling_input1_hybridearnings_item_header.png

Output artifact: Output artifact (Image): Preserved the Item 1. Business section as a clear heading with the opening paragraph below it. — docling_input1_hybridearnings_item_header_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Summary of Operating Performance section from the financial report. — docling_input2_financialpdf_summary_section-2.png

Observed output: Output artifact (Image): Kept the section heading and paragraph structure intact, including the concluding sentence that introduces the segment discussion. — docling_input2_financialpdf_summary_section_parsed_success-2.png

Input artifact: Input artifact (Image): Summary of Operating Performance section from the financial report. — docling_input2_financialpdf_summary_section-2.png

Output artifact: Output artifact (Image): Kept the section heading and paragraph structure intact, including the concluding sentence that introduces the segment discussion. — docling_input2_financialpdf_summary_section_parsed_success-2.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Scanned STAND PRESCRIPTIONS section with numbered and lettered list items. — docling_input3_scannedpaper_stand_prescriptions_list.png

Observed output: Output artifact (Image): Recovered the heading and nested list structure into readable markdown-style prose and bullets. — docling_input3_scannedpaper_stand_prescriptions_list_parsed_success.png

Input artifact: Input artifact (Image): Scanned STAND PRESCRIPTIONS section with numbered and lettered list items. — docling_input3_scannedpaper_stand_prescriptions_list.png

Output artifact: Output artifact (Image): Recovered the heading and nested list structure into readable markdown-style prose and bullets. — docling_input3_scannedpaper_stand_prescriptions_list_parsed_success.png

What changed: Image transformed into Image

Why it matters / Conclusion: Section headings and nested lists are consistently preserved well across native and scanned inputs.

Docling preserves sectioning, headings, and nested list structure so the output reads like a usable markdown document rather than a flat text dump. This held across the annual-report business section, the financial report summary headings, and the scanned paper's structured prescriptions list.

INPUT

Annual-report section header for Item 1. Business and General.

↓→

OUTPUT

Preserved the Item 1. Business section as a clear heading with the opening paragraph below it.

INPUT

Summary of Operating Performance section from the financial report.

↓→

OUTPUT

Kept the section heading and paragraph structure intact, including the concluding sentence that introduces the segment discussion.

INPUT

Scanned STAND PRESCRIPTIONS section with numbered and lettered list items.

↓→

OUTPUT

Recovered the heading and nested list structure into readable markdown-style prose and bullets.

Bottom Line

Section headings and nested lists are consistently preserved well across native and scanned inputs.

Reading Order Linearization

Mostly good, with one clear scanned-page exception

▾

Test Summary

Feature tested: Reading Order Linearization

Result: Partial — Mostly good, with one clear scanned-page exception

Feature tested: Reading Order Linearization

Result: Partial

Verdict: Mostly good, with one clear scanned-page exception

Expected behavior: Docling usually linearizes mixed-layout pages into a coherent reading sequence, including multi-column text and side-by-side brand lists. The main exception was the scanned paper's first page, where the abstract appeared before the title, and some multi-column layouts merged.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Scanned first page of the research note, with title above the abstract in the source layout. — docling_input3_scannedpaper_page_1.png

Observed output: Output artifact (Image): The abstract was extracted before the paper title, showing a first-page reading-order error even though the surrounding OCR text was recovered. — docling_input3_scannedpaper_page_1_parsed_failure.png

Input artifact: Input artifact (Image): Scanned first page of the research note, with title above the abstract in the source layout. — docling_input3_scannedpaper_page_1.png

Output artifact: Output artifact (Image): The abstract was extracted before the paper title, showing a first-page reading-order error even though the surrounding OCR text was recovered. — docling_input3_scannedpaper_page_1_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Two-column brand inventory page with owned brands and exclusive brands. — docling_input3_hybridearnings_brand_list.png

Observed output: Output artifact (Image): The list was reconstructed into a usable layout, but the two brand columns were merged into a single flow rather than kept fully distinct. — docling_input3_hybridearnings_brand_list_partial_success.png

Input artifact: Input artifact (Image): Two-column brand inventory page with owned brands and exclusive brands. — docling_input3_hybridearnings_brand_list.png

Output artifact: Output artifact (Image): The list was reconstructed into a usable layout, but the two brand columns were merged into a single flow rather than kept fully distinct. — docling_input3_hybridearnings_brand_list_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Segments and notes section from the financial report. — docling_input2_financialpdf_segments.png

Observed output: Output artifact (Image): The segment list and explanatory notes stayed in a logical sequence after extraction. — docling_input2_financialpdf_segments_parsed_success.png

Input artifact: Input artifact (Image): Segments and notes section from the financial report. — docling_input2_financialpdf_segments.png

Output artifact: Output artifact (Image): The segment list and explanatory notes stayed in a logical sequence after extraction. — docling_input2_financialpdf_segments_parsed_success.png

What changed: Image transformed into Image

Why it matters / Conclusion: Reading order is usually coherent, but the scanned paper's opening page and some multi-column layouts show the limits.

Docling usually linearizes mixed-layout pages into a coherent reading sequence, including multi-column text and side-by-side brand lists. The main exception was the scanned paper's first page, where the abstract appeared before the title, and some multi-column layouts merged.

INPUT

Scanned first page of the research note, with title above the abstract in the source layout.

↓→

OUTPUT

The abstract was extracted before the paper title, showing a first-page reading-order error even though the surrounding OCR text was recovered.

INPUT

Two-column brand inventory page with owned brands and exclusive brands.

↓→

OUTPUT

The list was reconstructed into a usable layout, but the two brand columns were merged into a single flow rather than kept fully distinct.

INPUT

Segments and notes section from the financial report.

↓→

OUTPUT

The segment list and explanatory notes stayed in a logical sequence after extraction.

Bottom Line

Reading order is usually coherent, but the scanned paper's opening page and some multi-column layouts show the limits.

Caption and Footnote Association

Strong

▾

Test Summary

Feature tested: Caption and Footnote Association

Result: Passed — Strong

Docling keeps footnotes and captions adjacent to the tables or figures they explain, even when other layout elements degrade. In the hybrid report the financial table footnotes stayed nearby, and in the scanned paper the Table 2 caption remained attached to the table header area.

INPUT

Dense financial-table footnotes block from the hybrid earnings report.

↓→

OUTPUT

The footnotes were extracted into readable paragraph text and stayed directly below the financial summary table.

INPUT

Scanned Table 2 caption and top rows from the research paper.

↓→

OUTPUT

The caption was preserved and remained adjacent to the table structure in the extracted output.

Bottom Line

Captions and footnotes are preserved close to their source tables, which makes the markdown easier to trust.

Embedded Image and Chart Preservation

Hard fail

▾

Test Summary

Feature tested: Embedded Image and Chart Preservation

Result: Failed — Hard fail

Feature tested: Embedded Image and Chart Preservation

Result: Failed

Verdict: Hard fail

Expected behavior: Docling does not retain charts, figures, logos, signatures, or divider graphics in a usable way. In the tested documents, those assets were replaced with generic placeholders such as ``, with no chart data reconstruction and no surviving visual file references.

Test case: Image → Text/code file

Input type: Image

Input used: Input artifact (Image): Four-panel Financial Highlights chart page from the hybrid earnings report. — docling_input1_hybridearnings_financial_charts.png

Observed output: Output artifact (Text/code file): The output markdown replaced the Financial Highlights charts with `` placeholders and did not recover any chart values. — docling_input1_hybridearnings_output.md

Input artifact: Input artifact (Image): Four-panel Financial Highlights chart page from the hybrid earnings report. — docling_input1_hybridearnings_financial_charts.png

Output artifact: Output artifact (Text/code file): The output markdown replaced the Financial Highlights charts with `` placeholders and did not recover any chart values. — docling_input1_hybridearnings_output.md

What changed: Image transformed into Text/code file

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Brian Cornell signature image from the hybrid earnings report. — docling_input1_hybridearnings_ceo_sign.png

Observed output: Output artifact (Image): The signature was omitted and replaced by a placeholder instead of being retained as an image. — docling_input1_hybridearnings_ceo_sign_omitted_failure.png

Input artifact: Input artifact (Image): Brian Cornell signature image from the hybrid earnings report. — docling_input1_hybridearnings_ceo_sign.png

Output artifact: Output artifact (Image): The signature was omitted and replaced by a placeholder instead of being retained as an image. — docling_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): CEO portrait/title-page image from the hybrid earnings report. — docling_input1_hybridearnings_ceo_image.png

Observed output: Output artifact (Image): The portrait image was omitted from the extracted output and reduced to text-only content. — docling_input1_hybridearnings_ceo_image_omitted_failed.png

Input artifact: Input artifact (Image): CEO portrait/title-page image from the hybrid earnings report. — docling_input1_hybridearnings_ceo_image.png

Output artifact: Output artifact (Image): The portrait image was omitted from the extracted output and reduced to text-only content. — docling_input1_hybridearnings_ceo_image_omitted_failed.png

What changed: Image transformed into Image

Test case: Image → Text/code file

Input type: Image

Input used: Input artifact (Image): Divider graphic from the financial report front matter. — docling_input2_financialpdf_divider_graphic.png

Observed output: Output artifact (Text/code file): The divider graphic was replaced by a placeholder in the markdown output rather than being retained as an image. — docling_input2_financialpdf_output.md

Input artifact: Input artifact (Image): Divider graphic from the financial report front matter. — docling_input2_financialpdf_divider_graphic.png

Output artifact: Output artifact (Text/code file): The divider graphic was replaced by a placeholder in the markdown output rather than being retained as an image. — docling_input2_financialpdf_output.md

What changed: Image transformed into Text/code file

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): USDA Forest Service logo and front-matter branding from the scanned paper. — docling_input3_scannedpaper_usda_logo.png

Observed output: Output artifact (Image): The logo was dropped and replaced with a placeholder in the extracted output. — docling_input3_scannedpaper_usda_logo_omitted_failed.png

Input artifact: Input artifact (Image): USDA Forest Service logo and front-matter branding from the scanned paper. — docling_input3_scannedpaper_usda_logo.png

Output artifact: Output artifact (Image): The logo was dropped and replaced with a placeholder in the extracted output. — docling_input3_scannedpaper_usda_logo_omitted_failed.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Figure 1 chart from the scanned paper. — docling_input3_scannedpaper_figure_1.png

Observed output: Output artifact (Image): Only the figure caption survived; the chart image itself was replaced with a placeholder. — docling_input3_scannedpaper_figure_1_parsed_failure.png

Input artifact: Input artifact (Image): Figure 1 chart from the scanned paper. — docling_input3_scannedpaper_figure_1.png

Output artifact: Output artifact (Image): Only the figure caption survived; the chart image itself was replaced with a placeholder. — docling_input3_scannedpaper_figure_1_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Figure 4 chart from the scanned paper. — docling_input3_scannedpaper_figure_4.png

Observed output: Output artifact (Image): The chart was not reconstructed and the output kept only the caption plus a placeholder. — docling_input3_scannedpaper_figure_4_parsed_failure.png

Input artifact: Input artifact (Image): Figure 4 chart from the scanned paper. — docling_input3_scannedpaper_figure_4.png

Output artifact: Output artifact (Image): The chart was not reconstructed and the output kept only the caption plus a placeholder. — docling_input3_scannedpaper_figure_4_parsed_failure.png

What changed: Image transformed into Image

Why it matters / Conclusion: Visual assets are consistently lost across all three inputs, so Docling is not safe when charts or images must survive conversion.

Docling does not retain charts, figures, logos, signatures, or divider graphics in a usable way. In the tested documents, those assets were replaced with generic placeholders such as ``, with no chart data reconstruction and no surviving visual file references.

INPUT

Four-panel Financial Highlights chart page from the hybrid earnings report.

↓→

OUTPUT

docling_input1_hybridearnings_output.md

Loading file...

The output markdown replaced the Financial Highlights charts with `` placeholders and did not recover any chart values.

INPUT

Brian Cornell signature image from the hybrid earnings report.

↓→

OUTPUT

The signature was omitted and replaced by a placeholder instead of being retained as an image.

INPUT

CEO portrait/title-page image from the hybrid earnings report.

↓→

OUTPUT

The portrait image was omitted from the extracted output and reduced to text-only content.

INPUT

Divider graphic from the financial report front matter.

↓→

OUTPUT

docling_input2_financialpdf_output.md

Loading file...

The divider graphic was replaced by a placeholder in the markdown output rather than being retained as an image.

INPUT

USDA Forest Service logo and front-matter branding from the scanned paper.

↓→

OUTPUT

The logo was dropped and replaced with a placeholder in the extracted output.

INPUT

Figure 1 chart from the scanned paper.

↓→

OUTPUT

Only the figure caption survived; the chart image itself was replaced with a placeholder.

INPUT

Figure 4 chart from the scanned paper.

↓→

OUTPUT

The chart was not reconstructed and the output kept only the caption plus a placeholder.

Bottom Line

Visual assets are consistently lost across all three inputs, so Docling is not safe when charts or images must survive conversion.

Form and Checkbox Extraction

Successful

▾

Test Summary

Feature tested: Form and Checkbox Extraction

Result: Passed — Successful

Docling can preserve simple filing-form checkbox state in the extracted output. On the SEC Form 10-K cover page, it correctly rendered the annual-report option as selected rather than dropping the mark entirely.

INPUT

SEC Form 10-K cover page with a marked annual-report checkbox.

↓→

OUTPUT

The checked annual-report state was preserved in the parsed output.

Bottom Line

A useful bonus capability: checkbox state survives the conversion.

Long-Document Robustness

Stable

▾

Test Summary

Feature tested: Long-Document Robustness

Result: Passed — Stable

Feature tested: Long-Document Robustness

Result: Passed

Verdict: Stable

Expected behavior: Docling stayed stable on long documents in this research: an 84-page annual report, an 18-page financial report, and a 12-page scanned research paper all completed without truncation or collapse in the extracted structure. The main quality risks came from layout complexity and images, not document length.

Test case: PDF document → Text/code file

Input type: PDF document

Input used: Input artifact (PDF document): 84-page hybrid annual report. — input1_hybridearnings.pdf

Observed output: Output artifact (Text/code file): The report was processed end to end with stable formatting; the research notes the run completed in 4 minutes 23 seconds. — docling_input1_hybridearnings_output.md

Input artifact: Input artifact (PDF document): 84-page hybrid annual report. — input1_hybridearnings.pdf

Output artifact: Output artifact (Text/code file): The report was processed end to end with stable formatting; the research notes the run completed in 4 minutes 23 seconds. — docling_input1_hybridearnings_output.md

What changed: PDF document transformed into Text/code file

Test case: PDF document → Text/code file

Input type: PDF document

Input used: Input artifact (PDF document): 18-page table-heavy financial report. — input2_financialpdf.pdf

Observed output: Output artifact (Text/code file): The report parsed end to end in 1 minute 57 seconds without truncation or output collapse. — docling_input2_financialpdf_output.md

Input artifact: Input artifact (PDF document): 18-page table-heavy financial report. — input2_financialpdf.pdf

Output artifact: Output artifact (Text/code file): The report parsed end to end in 1 minute 57 seconds without truncation or output collapse. — docling_input2_financialpdf_output.md

What changed: PDF document transformed into Text/code file

Test case: Artifact → Text/code file

Input type: Artifact

Input used: Input artifact (Artifact): 12-page scanned research paper.

Observed output: Output artifact (Text/code file): The scanned paper also completed in 1 minute 57 seconds, with stable extraction despite OCR and layout complexity. — docling_input3_scannedpaper_output.md

Input artifact: Input artifact (Artifact): 12-page scanned research paper.

Output artifact: Output artifact (Text/code file): The scanned paper also completed in 1 minute 57 seconds, with stable extraction despite OCR and layout complexity. — docling_input3_scannedpaper_output.md

What changed: Artifact transformed into Text/code file

Why it matters / Conclusion: Length itself is not the problem; Docling remained stable on all three long documents.

Docling stayed stable on long documents in this research: an 84-page annual report, an 18-page financial report, and a 12-page scanned research paper all completed without truncation or collapse in the extracted structure. The main quality risks came from layout complexity and images, not document length.

INPUT

input1_hybridearnings.pdf

84-page hybrid annual report.

↓→

OUTPUT

docling_input1_hybridearnings_output.md

Loading file...

The report was processed end to end with stable formatting; the research notes the run completed in 4 minutes 23 seconds.

INPUT

input2_financialpdf.pdf

18-page table-heavy financial report.

↓→

OUTPUT

docling_input2_financialpdf_output.md

Loading file...

The report parsed end to end in 1 minute 57 seconds without truncation or output collapse.

INPUT

Input

↓→

OUTPUT

docling_input3_scannedpaper_output.md

Loading file...

The scanned paper also completed in 1 minute 57 seconds, with stable extraction despite OCR and layout complexity.

Bottom Line

Length itself is not the problem; Docling remained stable on all three long documents.

✓ Use This If

●You need a self-hosted, open-source PDF-to-markdown pipeline for reports with lots of text, headings, and standard financial tables.

●You want OCR text from scanned pages and can tolerate minor spacing, word-merging, or library-stamp noise.

●You need footnotes, captions, and simple form checkboxes preserved in the converted output.

✕ Skip This If

●You need charts, figures, logos, signatures, or divider graphics to survive conversion.

●You need complex nested or scan-heavy tables to remain perfectly structured.

●You need exact visual fidelity or strict first-page reading order on scanned multi-column pages.

developer-toolsdocumentation-toolstext

No. In this research, charts, figures, logos, signatures, and divider graphics were consistently replaced with placeholders, and no chart data was reconstructed.

It handled standard financial tables very well, including clean row and column alignment. But multi-level headers could flatten or duplicate, and a dense scanned table became badly garbled.

Yes. It extracted scanned text accurately overall, but the scanned paper showed minor word-merging and library-stamp noise in the OCR output.

Yes. The reports kept their section headings, nested list structure, and general document hierarchy in markdown form.

Yes. Financial footnotes and scanned-table captions stayed adjacent to the relevant tables, even when the surrounding layout was imperfect.

Yes. The SEC filing cover page preserved the checked annual-report state in the parsed output.

Very stable in this research. It processed an 84-page report, an 18-page report, and a 12-page scanned paper without truncation or output collapse.