Unstructured icon
developer-tools

Unstructured

Good for text-first PDF-to-Markdown workflows, but not for faithful tables, charts, or images.

Visit Unstructured
Open sourceScanned OCRComplex tablesReading order
TL;DR — our verdictUpdated August 2026 · 19 test artifacts

Strong on prose, weak on structure-heavy PDF elements

Where it wins
  • You want an open-source PDF-to-Markdown parser you can run locally or in your own infrastructure.
  • Your PDFs are mostly prose with scanned pages, and you care about preserving reading order.
  • You can tolerate flattened headings and missing images if the text extraction is strong.
Main limitation
  • You need faithful chart preservation or chart-to-data extraction.
Pricing (verified plans)
Open source $0
Strongest test artifacts
PDF Text Extraction to Markdown →Table Reconstruction →Heading Detection →

Our take

Unstructured is reliable for narrative OCR and reading order, including scanned pages, but this benchmark shows flattened headings, dropped images, noisy charts, and fragile reconstruction for dense tables. It works best when your PDF is mostly prose with only modest table complexity.

Download video: Demo videoWatch video (streaming): Demo video
Colab notebook recording of the partition_pdf workflow processing the benchmark PDFs and writing Markdown outputs.

In-Depth Review

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

AD
AI Demos Team
Expert Reviewer
Verified Review

Feature-by-Feature Breakdown

PDF Text Extraction to Markdown
Mostly preserved, with punctuation and superscript fragility.
▾
Test Summary
Feature tested: PDF Text Extraction to Markdown
Result: Partial — Mostly preserved, with punctuation and superscript fragility.

Feature tested: PDF Text Extraction to Markdown

Result: Partial

Verdict: Mostly preserved, with punctuation and superscript fragility.

Expected behavior: Extracts prose from both native-text and scanned PDFs into markdown, including legal disclosures and checkbox-style lines. It handles ordinary body text well, though typographic punctuation and superscript markers may not be preserved exactly.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input1_hybridearnings_checkbox_text.png

Observed output: Output artifact (Image): The checkbox lines are preserved with the source's x/o markers intact, showing accurate extraction of native text that already encodes checkbox state. — unstructured_input1_hybridearnings_checkbox_text_parsed_success.png

Input artifact: Input artifact (Image): Input — unstructured_input1_hybridearnings_checkbox_text.png

Output artifact: Output artifact (Image): The checkbox lines are preserved with the source's x/o markers intact, showing accurate extraction of native text that already encodes checkbox state. — unstructured_input1_hybridearnings_checkbox_text_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input2_financialpdf_disclaimer_text.png

Observed output: Output artifact (Image): The disclaimer paragraph is mostly preserved, but the curly apostrophe in the forward-looking-statements text is corrupted into mojibake. — unstructured_input2_financialpdf_disclaimer_text_parsed_partial_success.png

Input artifact: Input artifact (Image): Input — unstructured_input2_financialpdf_disclaimer_text.png

Output artifact: Output artifact (Image): The disclaimer paragraph is mostly preserved, but the curly apostrophe in the forward-looking-statements text is corrupted into mojibake. — unstructured_input2_financialpdf_disclaimer_text_parsed_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_superscript2.png

Observed output: Output artifact (Image): The scanned-paper sentence is otherwise accurate, but every superscript-2 unit marker in the sentence is misread as a question mark. — unstructured_input3_scannedpaper_superscript2_parsed_partial_success.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_superscript2.png

Output artifact: Output artifact (Image): The scanned-paper sentence is otherwise accurate, but every superscript-2 unit marker in the sentence is misread as a question mark. — unstructured_input3_scannedpaper_superscript2_parsed_partial_success.png

What changed: Image transformed into Image

Why it matters / Conclusion: Good at ordinary prose OCR, but exact-string workflows will trip over non-ASCII punctuation and superscripts.

Extracts prose from both native-text and scanned PDFs into markdown, including legal disclosures and checkbox-style lines. It handles ordinary body text well, though typographic punctuation and superscript markers may not be preserved exactly.

image
Input artifact for "PDF Text Extraction to Markdown" test: Input, unstructured_input1_hybridearnings_checkbox_text.png
↓
image
Output artifact for "PDF Text Extraction to Markdown" test: The checkbox lines are preserved with the source's x/o markers intact, showing accurate extraction of native text that already encodes checkbox state., unstructured_input1_hybridearnings_checkbox_text_parsed_success.png
The checkbox lines are preserved with the source's x/o markers intact, showing accurate extraction of native text that already encodes checkbox state.
image
Input artifact for "PDF Text Extraction to Markdown" test: Input, unstructured_input2_financialpdf_disclaimer_text.png
↓
image
Output artifact for "PDF Text Extraction to Markdown" test: The disclaimer paragraph is mostly preserved, but the curly apostrophe in the forward-looking-statements text is corrupted into mojibake., unstructured_input2_financialpdf_disclaimer_text_parsed_partial_success.png
The disclaimer paragraph is mostly preserved, but the curly apostrophe in the forward-looking-statements text is corrupted into mojibake.
image
Input artifact for "PDF Text Extraction to Markdown" test: Input, unstructured_input3_scannedpaper_superscript2.png
↓
image
Output artifact for "PDF Text Extraction to Markdown" test: The scanned-paper sentence is otherwise accurate, but every superscript-2 unit marker in the sentence is misread as a question mark., unstructured_input3_scannedpaper_superscript2_parsed_partial_success.png
The scanned-paper sentence is otherwise accurate, but every superscript-2 unit marker in the sentence is misread as a question mark.
Bottom Line
Good at ordinary prose OCR, but exact-string workflows will trip over non-ASCII punctuation and superscripts.
Table Reconstruction
Simple tables hold up; dense tables are brittle.
▾
Test Summary
Feature tested: Table Reconstruction
Result: Partial — Simple tables hold up; dense tables are brittle.

Feature tested: Table Reconstruction

Result: Partial

Verdict: Simple tables hold up; dense tables are brittle.

Expected behavior: Converts tables into markdown, with cleaner results on simple 2-3 column tables. Dense financial tables and uneven side-by-side layouts can lose columns, misalign cells, or spill values into nearby text.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input1_hybridearnings_us_stores_table.png

Observed output: Output artifact (Image): The side-by-side U.S. Stores layout is merged for paired rows, but the unpaired Missouri row is silently dropped. — unstructured_input1_hybridearnings_us_stores_table_omitted_failed.png

Input artifact: Input artifact (Image): Input — unstructured_input1_hybridearnings_us_stores_table.png

Output artifact: Output artifact (Image): The side-by-side U.S. Stores layout is merged for paired rows, but the unpaired Missouri row is silently dropped. — unstructured_input1_hybridearnings_us_stores_table_omitted_failed.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input2_financialpdf_business_results_table.png

Observed output: Output artifact (Image): The Business Results table loses the entire percent-change data column and one header cell is garbled into OCR noise. — unstructured_input2_financialpdf_business_results_table_parsed_partial_success.png

Input artifact: Input artifact (Image): Input — unstructured_input2_financialpdf_business_results_table.png

Output artifact: Output artifact (Image): The Business Results table loses the entire percent-change data column and one header cell is garbled into OCR noise. — unstructured_input2_financialpdf_business_results_table_parsed_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input2_financialpdf_balance_sheet_table.png

Observed output: Output artifact (Image): The balance sheet keeps some rows correct, but others lose their second-column values to garbled glyphs and the table's own total-assets row is missing. — unstructured_input2_financialpdf_balance_sheet_table_parsed_partial_success.png

Input artifact: Input artifact (Image): Input — unstructured_input2_financialpdf_balance_sheet_table.png

Output artifact: Output artifact (Image): The balance sheet keeps some rows correct, but others lose their second-column values to garbled glyphs and the table's own total-assets row is missing. — unstructured_input2_financialpdf_balance_sheet_table_parsed_partial_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_table4.png

Observed output: Output artifact (Image): The widest table in the scanned paper suffers row-label merging and concatenated cell values, making a meaningful fraction of the table unusable without cleanup. — unstructured_input3_scannedpaper_table4_omitted_failed.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_table4.png

Output artifact: Output artifact (Image): The widest table in the scanned paper suffers row-label merging and concatenated cell values, making a meaningful fraction of the table unusable without cleanup. — unstructured_input3_scannedpaper_table4_omitted_failed.png

What changed: Image transformed into Image

Why it matters / Conclusion: Simple 2-3 column tables can come through well, but denser or uneven structures can lose columns, rows, or cell-to-value alignment.

Converts tables into markdown, with cleaner results on simple 2-3 column tables. Dense financial tables and uneven side-by-side layouts can lose columns, misalign cells, or spill values into nearby text.

image
Input artifact for "Table Reconstruction" test: Input, unstructured_input1_hybridearnings_us_stores_table.png
↓
image
Output artifact for "Table Reconstruction" test: The side-by-side U.S. Stores layout is merged for paired rows, but the unpaired Missouri row is silently dropped., unstructured_input1_hybridearnings_us_stores_table_omitted_failed.png
The side-by-side U.S. Stores layout is merged for paired rows, but the unpaired Missouri row is silently dropped.
image
Input artifact for "Table Reconstruction" test: Input, unstructured_input2_financialpdf_business_results_table.png
↓
image
Output artifact for "Table Reconstruction" test: The Business Results table loses the entire percent-change data column and one header cell is garbled into OCR noise., unstructured_input2_financialpdf_business_results_table_parsed_partial_success.png
The Business Results table loses the entire percent-change data column and one header cell is garbled into OCR noise.
image
Input artifact for "Table Reconstruction" test: Input, unstructured_input2_financialpdf_balance_sheet_table.png
↓
image
Output artifact for "Table Reconstruction" test: The balance sheet keeps some rows correct, but others lose their second-column values to garbled glyphs and the table's own total-assets row is missing., unstructured_input2_financialpdf_balance_sheet_table_parsed_partial_success.png
The balance sheet keeps some rows correct, but others lose their second-column values to garbled glyphs and the table's own total-assets row is missing.
image
Input artifact for "Table Reconstruction" test: Input, unstructured_input3_scannedpaper_table4.png
↓
image
Output artifact for "Table Reconstruction" test: The widest table in the scanned paper suffers row-label merging and concatenated cell values, making a meaningful fraction of the table unusable without cleanup., unstructured_input3_scannedpaper_table4_omitted_failed.png
The widest table in the scanned paper suffers row-label merging and concatenated cell values, making a meaningful fraction of the table unusable without cleanup.
Bottom Line
Simple 2-3 column tables can come through well, but denser or uneven structures can lose columns, rows, or cell-to-value alignment.
Heading Detection
Headings are detected, but nesting is flattened.
▾
Test Summary
Feature tested: Heading Detection
Result: Partial — Headings are detected, but nesting is flattened.

Feature tested: Heading Detection

Result: Partial

Verdict: Headings are detected, but nesting is flattened.

Expected behavior: Recognizes section headings and places them in markdown. In the tested PDFs it can flatten multi-level structure to a single depth and sometimes elevate non-heading text into heading form.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input1_hybridearnings_item1a_heading.png

Observed output: Output artifact (Image): The Item 1A heading and its subordinate category heading are both rendered at the same markdown level, flattening the source hierarchy. — unstructured_input1_hybridearnings_item1a_heading_parsed_failure.png

Input artifact: Input artifact (Image): Input — unstructured_input1_hybridearnings_item1a_heading.png

Output artifact: Output artifact (Image): The Item 1A heading and its subordinate category heading are both rendered at the same markdown level, flattening the source hierarchy. — unstructured_input1_hybridearnings_item1a_heading_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input2_financialpdf_heading_levels_top.png

Observed output: Output artifact (Image): The top-level section, its subsection, and the nested '(i) Mechatronics' sub-subsection are all flattened to the same heading depth. — unstructured_input2_financialpdf_heading_levels_parsed_failure.png

Input artifact: Input artifact (Image): Input — unstructured_input2_financialpdf_heading_levels_top.png

Output artifact: Output artifact (Image): The top-level section, its subsection, and the nested '(i) Mechatronics' sub-subsection are all flattened to the same heading depth. — unstructured_input2_financialpdf_heading_levels_parsed_failure.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_logo_header.png

Observed output: Output artifact (Image): Only part of the source's sidebar text is promoted to a heading, and the author byline is also treated like a heading rather than plain metadata text. — unstructured_input3_scannedpaper_logo_header_omitted_failed.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_logo_header.png

Output artifact: Output artifact (Image): Only part of the source's sidebar text is promoted to a heading, and the author byline is also treated like a heading rather than plain metadata text. — unstructured_input3_scannedpaper_logo_header_omitted_failed.png

What changed: Image transformed into Image

Why it matters / Conclusion: The tool finds headings, but the source's real nesting depth does not survive the conversion.

Recognizes section headings and places them in markdown. In the tested PDFs it can flatten multi-level structure to a single depth and sometimes elevate non-heading text into heading form.

image
Input artifact for "Heading Detection" test: Input, unstructured_input1_hybridearnings_item1a_heading.png
↓
image
Output artifact for "Heading Detection" test: The Item 1A heading and its subordinate category heading are both rendered at the same markdown level, flattening the source hierarchy., unstructured_input1_hybridearnings_item1a_heading_parsed_failure.png
The Item 1A heading and its subordinate category heading are both rendered at the same markdown level, flattening the source hierarchy.
image
Input artifact for "Heading Detection" test: Input, unstructured_input2_financialpdf_heading_levels_top.png
↓
image
Output artifact for "Heading Detection" test: The top-level section, its subsection, and the nested '(i) Mechatronics' sub-subsection are all flattened to the same heading depth., unstructured_input2_financialpdf_heading_levels_parsed_failure.png
The top-level section, its subsection, and the nested '(i) Mechatronics' sub-subsection are all flattened to the same heading depth.
image
Input artifact for "Heading Detection" test: Input, unstructured_input3_scannedpaper_logo_header.png
↓
image
Output artifact for "Heading Detection" test: Only part of the source's sidebar text is promoted to a heading, and the author byline is also treated like a heading rather than plain metadata text., unstructured_input3_scannedpaper_logo_header_omitted_failed.png
Only part of the source's sidebar text is promoted to a heading, and the author byline is also treated like a heading rather than plain metadata text.
Bottom Line
The tool finds headings, but the source's real nesting depth does not survive the conversion.
Reading Order Preservation
Section flow and column order stay in sequence.
▾
Test Summary
Feature tested: Reading Order Preservation
Result: Passed — Section flow and column order stay in sequence.

Feature tested: Reading Order Preservation

Result: Passed

Verdict: Section flow and column order stay in sequence.

Expected behavior: Keeps content in the source's top-to-bottom reading order across pages, including back-page footer text. It also linearizes two-column scanned content left-to-right in reading sequence.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input1_hybridearnings_item1_end.png

Observed output: Output artifact (Image): The closing Item 1 disclosure is extracted in the same place it appears in the source, with the next section continuing in order. — unstructured_input1_hybridearnings_item1_end_parsed_success.png

Input artifact: Input artifact (Image): Input — unstructured_input1_hybridearnings_item1_end.png

Output artifact: Output artifact (Image): The closing Item 1 disclosure is extracted in the same place it appears in the source, with the next section continuing in order. — unstructured_input1_hybridearnings_item1_end_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input2_financialpdf_segment_sequence.png

Observed output: Output artifact (Image): The segment subsections are emitted in the same source order, without being shuffled across the report. — unstructured_input2_financialpdf_segment_sequence_parsed_success.png

Input artifact: Input artifact (Image): Input — unstructured_input2_financialpdf_segment_sequence.png

Output artifact: Output artifact (Image): The segment subsections are emitted in the same source order, without being shuffled across the report. — unstructured_input2_financialpdf_segment_sequence_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_title_abstract_order.png

Observed output: Output artifact (Image): The scanned paper preserves the opening sequence of title, subtitle, author line, abstract, and keywords in the same top-to-bottom order as the page. — unstructured_input3_scannedpaper_title_abstract_order_parsed_success.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_title_abstract_order.png

Output artifact: Output artifact (Image): The scanned paper preserves the opening sequence of title, subtitle, author line, abstract, and keywords in the same top-to-bottom order as the page. — unstructured_input3_scannedpaper_title_abstract_order_parsed_success.png

What changed: Image transformed into Image

Why it matters / Conclusion: Reading order is a clear strength, including column linearization and end-of-document footers.

Keeps content in the source's top-to-bottom reading order across pages, including back-page footer text. It also linearizes two-column scanned content left-to-right in reading sequence.

image
Input artifact for "Reading Order Preservation" test: Input, unstructured_input1_hybridearnings_item1_end.png
↓
image
Output artifact for "Reading Order Preservation" test: The closing Item 1 disclosure is extracted in the same place it appears in the source, with the next section continuing in order., unstructured_input1_hybridearnings_item1_end_parsed_success.png
The closing Item 1 disclosure is extracted in the same place it appears in the source, with the next section continuing in order.
image
Input artifact for "Reading Order Preservation" test: Input, unstructured_input2_financialpdf_segment_sequence.png
↓
image
Output artifact for "Reading Order Preservation" test: The segment subsections are emitted in the same source order, without being shuffled across the report., unstructured_input2_financialpdf_segment_sequence_parsed_success.png
The segment subsections are emitted in the same source order, without being shuffled across the report.
image
Input artifact for "Reading Order Preservation" test: Input, unstructured_input3_scannedpaper_title_abstract_order.png
↓
image
Output artifact for "Reading Order Preservation" test: The scanned paper preserves the opening sequence of title, subtitle, author line, abstract, and keywords in the same top-to-bottom order as the page., unstructured_input3_scannedpaper_title_abstract_order_parsed_success.png
The scanned paper preserves the opening sequence of title, subtitle, author line, abstract, and keywords in the same top-to-bottom order as the page.
Bottom Line
Reading order is a clear strength, including column linearization and end-of-document footers.
Chart Handling
Charts become unreadable noise instead of usable figures or data.
▾
Test Summary
Feature tested: Chart Handling
Result: Failed — Charts become unreadable noise instead of usable figures or data.

Feature tested: Chart Handling

Result: Failed

Verdict: Charts become unreadable noise instead of usable figures or data.

Expected behavior: Attempts to process chart regions from both native and scanned documents. In the tested outputs, charts were not reconstructed as structured data or clean placeholders, and instead became OCR noise or digit-soup in the markdown body.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input1_hybridearnings_financial_highlights_chart.png

Observed output: Output artifact (Image): The chart value labels are read character-by-character in the wrong order, turning the chart into interleaved digit-soup. — unstructured_input1_hybridearnings_financial_highlights_chart_omitted_failed.png

Input artifact: Input artifact (Image): Input — unstructured_input1_hybridearnings_financial_highlights_chart.png

Output artifact: Output artifact (Image): The chart value labels are read character-by-character in the wrong order, turning the chart into interleaved digit-soup. — unstructured_input1_hybridearnings_financial_highlights_chart_omitted_failed.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_figure1_chart.png

Observed output: Output artifact (Image): The scanned chart region is emitted as OCR noise with a few fragmentary labels, but no usable chart data survives. — unstructured_input3_scannedpaper_figure1_chart_omitted_failed.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_figure1_chart.png

Output artifact: Output artifact (Image): The scanned chart region is emitted as OCR noise with a few fragmentary labels, but no usable chart data survives. — unstructured_input3_scannedpaper_figure1_chart_omitted_failed.png

What changed: Image transformed into Image

Why it matters / Conclusion: Charts are not preserved usefully: they are neither converted to data nor kept cleanly as visuals.

Attempts to process chart regions from both native and scanned documents. In the tested outputs, charts were not reconstructed as structured data or clean placeholders, and instead became OCR noise or digit-soup in the markdown body.

image
Input artifact for "Chart Handling" test: Input, unstructured_input1_hybridearnings_financial_highlights_chart.png
↓
image
Output artifact for "Chart Handling" test: The chart value labels are read character-by-character in the wrong order, turning the chart into interleaved digit-soup., unstructured_input1_hybridearnings_financial_highlights_chart_omitted_failed.png
The chart value labels are read character-by-character in the wrong order, turning the chart into interleaved digit-soup.
image
Input artifact for "Chart Handling" test: Input, unstructured_input3_scannedpaper_figure1_chart.png
↓
image
Output artifact for "Chart Handling" test: The scanned chart region is emitted as OCR noise with a few fragmentary labels, but no usable chart data survives., unstructured_input3_scannedpaper_figure1_chart_omitted_failed.png
The scanned chart region is emitted as OCR noise with a few fragmentary labels, but no usable chart data survives.
Bottom Line
Charts are not preserved usefully: they are neither converted to data nor kept cleanly as visuals.
Embedded Image and Logo Preservation
Embedded images are dropped entirely.
▾
Test Summary
Feature tested: Embedded Image and Logo Preservation
Result: Failed — Embedded images are dropped entirely.

Feature tested: Embedded Image and Logo Preservation

Result: Failed

Verdict: Embedded images are dropped entirely.

Expected behavior: Handles embedded images and logos during markdown export, but in the tested PDFs they were dropped with no image reference, placeholder, or comment marker. Repeated logos across pages were also removed.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input2_financialpdf_header_logo.png

Observed output: Output artifact (Image): The recurring header logo is omitted with zero trace in the markdown output. — unstructured_image_omitted_blank_placeholder.png

Input artifact: Input artifact (Image): Input — unstructured_input2_financialpdf_header_logo.png

Output artifact: Output artifact (Image): The recurring header logo is omitted with zero trace in the markdown output. — unstructured_image_omitted_blank_placeholder.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_shield_logo.png

Observed output: Output artifact (Image): The printed shield logo is also dropped entirely, with no placeholder at all. — unstructured_image_omitted_blank_placeholder.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_shield_logo.png

Output artifact: Output artifact (Image): The printed shield logo is also dropped entirely, with no placeholder at all. — unstructured_image_omitted_blank_placeholder.png

What changed: Image transformed into Image

Why it matters / Conclusion: No embedded image or logo tested here survives the markdown export.

Handles embedded images and logos during markdown export, but in the tested PDFs they were dropped with no image reference, placeholder, or comment marker. Repeated logos across pages were also removed.

image
Input artifact for "Embedded Image and Logo Preservation" test: Input, unstructured_input2_financialpdf_header_logo.png
↓
image
Output artifact for "Embedded Image and Logo Preservation" test: The recurring header logo is omitted with zero trace in the markdown output., unstructured_image_omitted_blank_placeholder.png
The recurring header logo is omitted with zero trace in the markdown output.
image
Input artifact for "Embedded Image and Logo Preservation" test: Input, unstructured_input3_scannedpaper_shield_logo.png
↓
image
Output artifact for "Embedded Image and Logo Preservation" test: The printed shield logo is also dropped entirely, with no placeholder at all., unstructured_image_omitted_blank_placeholder.png
The printed shield logo is also dropped entirely, with no placeholder at all.
Bottom Line
No embedded image or logo tested here survives the markdown export.
Footnote and Caption Placement
Nearby footnotes and captions stay attached.
▾
Test Summary
Feature tested: Footnote and Caption Placement
Result: Passed — Nearby footnotes and captions stay attached.

Feature tested: Footnote and Caption Placement

Result: Passed

Verdict: Nearby footnotes and captions stay attached.

Expected behavior: Keeps footnotes directly beneath their source table and preserves figure captions in the correct relative position to the figure. This adjacency holds even when surrounding table or figure content is degraded.

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input1_hybridearnings_footnote_a.png

Observed output: Output artifact (Image): The footnotes remain directly attached below the financial summary table instead of drifting to another part of the document. — unstructured_input1_hybridearnings_footnote_a_parsed_success.png

Input artifact: Input artifact (Image): Input — unstructured_input1_hybridearnings_footnote_a.png

Output artifact: Output artifact (Image): The footnotes remain directly attached below the financial summary table instead of drifting to another part of the document. — unstructured_input1_hybridearnings_footnote_a_parsed_success.png

What changed: Image transformed into Image

Test case: Image → Image

Input type: Image

Input used: Input artifact (Image): Input — unstructured_input3_scannedpaper_figure1_chart.png

Observed output: Output artifact (Image): The Figure 1 caption stays directly below the chart position, even though the chart itself degrades into OCR noise. — unstructured_input3_scannedpaper_figure1_chart_omitted_failed.png

Input artifact: Input artifact (Image): Input — unstructured_input3_scannedpaper_figure1_chart.png

Output artifact: Output artifact (Image): The Figure 1 caption stays directly below the chart position, even though the chart itself degrades into OCR noise. — unstructured_input3_scannedpaper_figure1_chart_omitted_failed.png

What changed: Image transformed into Image

Why it matters / Conclusion: Adjacency is preserved for footnotes and captions, even when the underlying table or figure is damaged.

Keeps footnotes directly beneath their source table and preserves figure captions in the correct relative position to the figure. This adjacency holds even when surrounding table or figure content is degraded.

image
Input artifact for "Footnote and Caption Placement" test: Input, unstructured_input1_hybridearnings_footnote_a.png
↓
image
Output artifact for "Footnote and Caption Placement" test: The footnotes remain directly attached below the financial summary table instead of drifting to another part of the document., unstructured_input1_hybridearnings_footnote_a_parsed_success.png
The footnotes remain directly attached below the financial summary table instead of drifting to another part of the document.
image
Input artifact for "Footnote and Caption Placement" test: Input, unstructured_input3_scannedpaper_figure1_chart.png
↓
image
Output artifact for "Footnote and Caption Placement" test: The Figure 1 caption stays directly below the chart position, even though the chart itself degrades into OCR noise., unstructured_input3_scannedpaper_figure1_chart_omitted_failed.png
The Figure 1 caption stays directly below the chart position, even though the chart itself degrades into OCR noise.
Bottom Line
Adjacency is preserved for footnotes and captions, even when the underlying table or figure is damaged.

Pricing

TESTED
Open source
$0
Apache 2.0; self-hosted library.

The report describes Unstructured as free and open source under Apache 2.0. The benchmark run was done in Google Colab on a T4 GPU, and processing time was not recorded.

✓ Use This If
●You want an open-source PDF-to-Markdown parser you can run locally or in your own infrastructure.
●Your PDFs are mostly prose with scanned pages, and you care about preserving reading order.
●You can tolerate flattened headings and missing images if the text extraction is strong.
✕ Skip This If
●You need faithful chart preservation or chart-to-data extraction.
●You need embedded images and logos retained in the markdown output.
●Your workflow depends on complex tables, merged cells, or exact punctuation and superscript preservation.
developer-toolspdf-toolstextOther
Yes for ordinary prose. In the scanned research paper, full paragraphs were read accurately, but superscript ² markers and some footnote numerals were misread.
Charts were not reconstructed usefully. Native charts became interleaved digit-soup, and scanned chart regions became OCR noise with only fragments of labels surviving.
Not reliably. Simple tables were mostly fine, but dense tables could lose a column, garble cells, drop summary rows, or spill correct values into orphaned text after the table.
It detects headings, but the benchmark showed that multi-level structure collapses to one heading depth, and at least one byline was misclassified as a heading.
No. Embedded images and logos were omitted entirely in all three inputs, with no placeholder, comment, or image reference left behind.
Yes. The report describes it as open source under Apache 2.0, so there is no license fee; the benchmark itself was run in Google Colab on a T4 GPU.

Banner Preview

How the embed badge will look on your site

Unstructured featured on AI Demos

Embed HTML

Copy this code to your website source

<a target="_blank" href="https://aidemos.com/tools/unstructured?utm_source=unstructured_embed" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> <img src="https://aidemos-website-images.s3.amazonaws.com/featured.png" alt="Unstructured | Featured on AI Demos" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> </a>

Quick Integration Guide

  • 1Copy the HTML code block above.
  • 2Paste it into your site's HTML or CMS editor.
  • 3Banner appears instantly on your page.
  • 4Links back to your tool profile here.
Similar Tools

Similar Tools

Discover more AI tools like Unstructured to enhance your workflow.

🤖
Docling
Open-source PDF-to-markdown conversion that is strong on text, headings, and standard tables, but drops charts and other visual assets.
AI Tool
🤖
PyMuPDF4LLM
Open-source PDF-to-markdown for clean native PDFs, but unreliable on scans, dense tables, and images.
AI Tool
🤖
LiteParse
Open-source PDF-to-markdown parsing that works well on native-digital reports, but degrades on scans, tables, and charts.
AI Tool
🤖
Doc2Mark
Open-source PDF-to-markdown that preserves native text and headings, but still struggles with tables, charts, images, and scans.
AI Tool
🤖
MarkItDown
Fast native-PDF text extraction for markdown, but structure, charts, images, and scans are unreliable.
AI Tool
🤖
MinerU
Open-source PDF-to-markdown that keeps figures and charts in place, but tables and punctuation can get messy on complex files.
AI Tool
🤖
PaddleOCR
Open-source PDF-to-markdown conversion that keeps images and tables in place, but still corrupts text and complex structure on harder documents.
AI Tool
🤖
Dolphin
Self-hosted PDF-to-Markdown that preserves most text and tables, but drops charts and struggles with repeated images.
AI Tool
🤖
DocTR
AI Tool

Comments (0)

Please Log in to join the discussion.

Built by FutureSmart AI — the team behind AI Demos

Need a custom AI solution for this use case?

If you are looking to build a custom PDF to Markdown, text extraction, or document parsing system for your business or internal workflow, email us at contact@futuresmart.ai.

Get a custom build

Found something inaccurate or missing? We try to keep our AI research accurate and useful. If you found outdated information, an issue, or have a suggestion, email us at collaborate@aidemos.com.

Back to Top