Upstage AI
Solid on native financial tables, but unreliable for multi-column and scanned-document structure in markdown conversion.
Mixed result for complex PDF-to-markdown work
Upstage AI handled the API workflow cleanly and did its best work on native financial tables, where row/value placement stayed mostly intact. It also converted charts into text summaries with extracted values instead of dropping them outright. But across the broader use case, it was inconsistent: multi-column pages lost hierarchy, scanned signature pages flattened badly, and chart/table structure became less trustworthy once layouts got harder.
In-Depth Review
Our detailed analysis of Upstage AI — features, performance, and real-world testing.
Feature-by-Feature Breakdown
API-based PDF-to-markdown conversionReliable ingestion and export across the three tested PDFs.▾
Feature tested: API-based PDF-to-markdown conversion
Result: Passed
Verdict: Reliable ingestion and export across the three tested PDFs.
Expected behavior: Upstage accepted all three tested documents through an automated API workflow and returned downloadable markdown outputs: an 84-page hybrid earnings report, an 18-page table-heavy financial report, and a scanned research report. The researcher did not need manual correction, UI interaction, or post-processing to obtain the markdown files.
Test case: PDF document → Text/code file
Input type: PDF document
Input used: Input artifact (PDF document): Hybrid earnings report (84 pages) with native text, financial tables, charts, and a scanned signature page. — llamaparse-hybrid-earnings-pdf-1.pdf
Observed output: Output artifact (Text/code file): Upstage accepted the hybrid earnings report and returned a downloadable markdown file through a fully automated API flow. The researcher did not report any manu — upstage-ai-upstage-hybrid-earningspdf-output-1.md
Input artifact: Input artifact (PDF document): Hybrid earnings report (84 pages) with native text, financial tables, charts, and a scanned signature page. — llamaparse-hybrid-earnings-pdf-1.pdf
Output artifact: Output artifact (Text/code file): Upstage accepted the hybrid earnings report and returned a downloadable markdown file through a fully automated API flow. The researcher did not report any manu — upstage-ai-upstage-hybrid-earningspdf-output-1.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 (18 pages). — llamaparse-sumitomo-financial-pdf-1.pdf
Observed output: Output artifact (Text/code file): Upstage accepted the table-heavy financial PDF and exported a markdown file directly, matching the report's requirement for programmatic markdown output. — upstage-ai-upstage-financialpdf-parsed.md
Input artifact: Input artifact (PDF document): Table-heavy financial report (18 pages). — llamaparse-sumitomo-financial-pdf-1.pdf
Output artifact: Output artifact (Text/code file): Upstage accepted the table-heavy financial PDF and exported a markdown file directly, matching the report's requirement for programmatic markdown output. — upstage-ai-upstage-financialpdf-parsed.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): Scanned research report used to test OCR and layout handling. — llamaparse-scanned-research-pdf-1.pdf
Observed output: Output artifact (Text/code file): Upstage also processed the scanned research PDF and returned markdown without any manual repair step in the documented workflow. — upstage-ai-upstage-scannedpdf-output.md
Input artifact: Input artifact (PDF document): Scanned research report used to test OCR and layout handling. — llamaparse-scanned-research-pdf-1.pdf
Output artifact: Output artifact (Text/code file): Upstage also processed the scanned research PDF and returned markdown without any manual repair step in the documented workflow. — upstage-ai-upstage-scannedpdf-output.md
What changed: PDF document transformed into Text/code file
Why it matters / Conclusion: If your first question is simply whether the service will accept varied PDFs and give you markdown back through an API, Upstage passed that baseline cleanly in all three tests.
Upstage accepted all three tested documents through an automated API workflow and returned downloadable markdown outputs: an 84-page hybrid earnings report, an 18-page table-heavy financial report, and a scanned research report. The researcher did not need manual correction, UI interaction, or post-processing to obtain the markdown files.
Table reconstructionStrong on cleaner native tables; weaker on harder headers and scanned layouts.▾
Feature tested: Table reconstruction
Result: Partial
Verdict: Strong on cleaner native tables; weaker on harder headers and scanned layouts.
Expected behavior: Upstage reconstructs tables into readable markdown-like structure, but quality depends heavily on source complexity. It preserved the Target financial summary table from the hybrid earnings report with strong row/column fidelity and only minor symbol loss. On the Sumitomo quarterly balance sheet and the scanned forestry table, body values survived better than header structure, and grouped headers became misaligned or duplicated.
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Target 2015 annual report financial summary table. — landing-ai-target-annual-report-financial-summary-table-2.png
Observed output: Output artifact (Image): From the Target financial summary table, Upstage preserved the year columns from 2015 to 2011 and kept line items such as Sales, SG&A, EBIT, taxes, and net earn — upstage-ai-target-2015-financial-results-parsed-table.png
Input artifact: Input artifact (Image): Target 2015 annual report financial summary table. — landing-ai-target-annual-report-financial-summary-table-2.png
Output artifact: Output artifact (Image): From the Target financial summary table, Upstage preserved the year columns from 2015 to 2011 and kept line items such as Sales, SG&A, EBIT, taxes, and net earn — upstage-ai-target-2015-financial-results-parsed-table.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Sumitomo quarterly consolidated balance sheet table. — upstage-ai-sumitomo-quarterly-consolidated-balance-sheets.png
Observed output: Output artifact (Image): On the quarterly balance sheet page, the asset rows and numeric values remained visible, but the column headers became misaligned with their corresponding data — upstage-ai-parsed-balance-sheet.png
Input artifact: Input artifact (Image): Sumitomo quarterly consolidated balance sheet table. — upstage-ai-sumitomo-quarterly-consolidated-balance-sheets.png
Output artifact: Output artifact (Image): On the quarterly balance sheet page, the asset rows and numeric values remained visible, but the column headers became misaligned with their corresponding data — upstage-ai-parsed-balance-sheet.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned treatment table with original diameter and diameter after harvest. — mistral-ai-scanned-treatment-diameter-table.png
Observed output: Output artifact (Image): On the scanned treatment table, Upstage retained the four treatment rows and their numeric values, but header reconstruction broke down. 'Original diameter' and — upstage-ai-parsed-diameter-treatment-table.png
Input artifact: Input artifact (Image): Scanned treatment table with original diameter and diameter after harvest. — mistral-ai-scanned-treatment-diameter-table.png
Output artifact: Output artifact (Image): On the scanned treatment table, Upstage retained the four treatment rows and their numeric values, but header reconstruction broke down. 'Original diameter' and — upstage-ai-parsed-diameter-treatment-table.png
What changed: Image transformed into Image
Why it matters / Conclusion: Upstage is credible for readable extraction of simpler native financial tables, but once headers get more complex or the source is scanned, the markdown table structure becomes much less dependable.
Upstage reconstructs tables into readable markdown-like structure, but quality depends heavily on source complexity. It preserved the Target financial summary table from the hybrid earnings report with strong row/column fidelity and only minor symbol loss. On the Sumitomo quarterly balance sheet and the scanned forestry table, body values survived better than header structure, and grouped headers became misaligned or duplicated.
Target 2015 annual report financial summary table.
From the Target financial summary table, Upstage preserved the year columns from 2015 to 2011 and kept line items such as Sales, SG&A, EBIT, taxes, and net earnings in the correct rows. The researcher noted only minor loss of currency symbols, so the table stayed readable but was not perfectly faithful.
Sumitomo quarterly consolidated balance sheet table.
On the quarterly balance sheet page, the asset rows and numeric values remained visible, but the column headers became misaligned with their corresponding data regions. That made the extracted table structurally inconsistent with the source even though much of the content was still present.
Scanned treatment table with original diameter and diameter after harvest.
On the scanned treatment table, Upstage retained the four treatment rows and their numeric values, but header reconstruction broke down. 'Original diameter' and 'Diameter after harvest' were duplicated and split awkwardly, so the grouped header structure was not faithfully rebuilt.
Chart and figure extraction to textBetter than dropping charts, but not clean enough to count as faithful chart preservation.▾
Feature tested: Chart and figure extraction to text
Result: Partial
Verdict: Better than dropping charts, but not clean enough to count as faithful chart preservation.
Expected behavior: Upstage converts charts and figures into prose summaries plus extracted values instead of discarding them. In the hybrid earnings report, it turned an SG&A waterfall chart into a narrative explanation and category/value list. In the scanned research report, it summarized a multi-series line chart and produced year-by-year values. The tradeoff is organization: the recovered data was described as raw or poorly structured rather than preserved in a clean, chart-like form.
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): SG&A rate waterfall chart from the hybrid earnings report. — llamaparse-sga-rate-waterfall-chart-1.png
Observed output: Output artifact (Image): For the SG&A waterfall chart, Upstage generated a prose description and listed category/value pairs from 2013 to 2015 instead of omitting the figure. The resear — upstage-ai-sgaa-rate-waterfall-text-description.png
Input artifact: Input artifact (Image): SG&A rate waterfall chart from the hybrid earnings report. — llamaparse-sga-rate-waterfall-chart-1.png
Output artifact: Output artifact (Image): For the SG&A waterfall chart, Upstage generated a prose description and listed category/value pairs from 2013 to 2015 instead of omitting the figure. The resear — upstage-ai-sgaa-rate-waterfall-text-description.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned Figure 3 line chart showing average radial growth by year. — upstage-ai-figure-3-average-radial-growth-line-chart.png
Observed output: Output artifact (Image): For the scanned line chart, Upstage produced a chart summary and a year-by-year value table for five series from 1972 to 1981. The figure content was therefore — upstage-ai-parsed-figure-3-radial-growth-summary.png
Input artifact: Input artifact (Image): Scanned Figure 3 line chart showing average radial growth by year. — upstage-ai-figure-3-average-radial-growth-line-chart.png
Output artifact: Output artifact (Image): For the scanned line chart, Upstage produced a chart summary and a year-by-year value table for five series from 1972 to 1981. The figure content was therefore — upstage-ai-parsed-figure-3-radial-growth-summary.png
What changed: Image transformed into Image
Why it matters / Conclusion: Upstage does retain chart information in text form, which is better than a silent drop, but the output is still too loosely structured for users who need faithful markdown representations of figures.
Upstage converts charts and figures into prose summaries plus extracted values instead of discarding them. In the hybrid earnings report, it turned an SG&A waterfall chart into a narrative explanation and category/value list. In the scanned research report, it summarized a multi-series line chart and produced year-by-year values. The tradeoff is organization: the recovered data was described as raw or poorly structured rather than preserved in a clean, chart-like form.
SG&A rate waterfall chart from the hybrid earnings report.
For the SG&A waterfall chart, Upstage generated a prose description and listed category/value pairs from 2013 to 2015 instead of omitting the figure. The researcher still judged the result incomplete because the values were not returned in a clean structured form that preserved the chart's original organization.
Scanned Figure 3 line chart showing average radial growth by year.
For the scanned line chart, Upstage produced a chart summary and a year-by-year value table for five series from 1972 to 1981. The figure content was therefore recoverable, but it came back as raw delimiter-separated data and even included stray text from another figure title, which hurt interpretability.
Reading order and document hierarchy preservationInconsistent on digital sections and weak on multi-column pages.▾
Feature tested: Reading order and document hierarchy preservation
Result: Failed
Verdict: Inconsistent on digital sections and weak on multi-column pages.
Expected behavior: Upstage preserved basic narrative flow on at least one straightforward digital section, but it struggled to keep hierarchy and navigation intact on harder layouts. The Mechatronics/Industrial Machinery section from the Sumitomo report remained readable with subsection ordering preserved. By contrast, the hybrid earnings report's two-column strategy page blurred headings into surrounding prose, the operating-performance page lost clear heading distinction, and the researcher reported paragraph segmentation and reading-order breakdown on the scanned multi-column research paper.
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Numbered Mechatronics / Industrial Machinery section from the financial report. — upstage-ai-financial-section-bulletins-mechatronics-industrial-machinery.png
Observed output: Output artifact (Image): On the Mechatronics / Industrial Machinery section, Upstage kept the numbered subsection headings and paragraph flow in the same general order as the source, so — upstage-ai-parsed-financial-mechatronics-industrial-machinery-dark.png
Input artifact: Input artifact (Image): Numbered Mechatronics / Industrial Machinery section from the financial report. — upstage-ai-financial-section-bulletins-mechatronics-industrial-machinery.png
Output artifact: Output artifact (Image): On the Mechatronics / Industrial Machinery section, Upstage kept the numbered subsection headings and paragraph flow in the same general order as the source, so — upstage-ai-parsed-financial-mechatronics-industrial-machinery-dark.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Two-column Target narrative page with several distinct subheads. — upstage-ai-target-two-column-narrative-with-highlighted-section.png
Observed output: Output artifact (Image): On the two-column Target narrative page, Upstage preserved much of the text content but blurred the original navigation structure. Headings such as 'Target.com — upstage-ai-target-earnings-parsed-strategy-and-merchandising.png
Input artifact: Input artifact (Image): Two-column Target narrative page with several distinct subheads. — upstage-ai-target-two-column-narrative-with-highlighted-section.png
Output artifact: Output artifact (Image): On the two-column Target narrative page, Upstage preserved much of the text content but blurred the original navigation structure. Headings such as 'Target.com — upstage-ai-target-earnings-parsed-strategy-and-merchandising.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Summary of Operating Performance page from the Sumitomo report. — upstage-ai-summary-operating-performance-quarterly-results.png
Observed output: Output artifact (Image): On the 'Summary of Operating Performance' page, Upstage retained the substantive text but reduced section-level distinction. The researcher reported that headin — upstage-ai-operating-performance-summary-annotated-callouts.png
Input artifact: Input artifact (Image): Summary of Operating Performance page from the Sumitomo report. — upstage-ai-summary-operating-performance-quarterly-results.png
Output artifact: Output artifact (Image): On the 'Summary of Operating Performance' page, Upstage retained the substantive text but reduced section-level distinction. The researcher reported that headin — upstage-ai-operating-performance-summary-annotated-callouts.png
What changed: Image transformed into Image
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned multi-column forestry page headed 'STUDY AREA' with 'STAND PRESCRIPTIONS' lower on the page. — landing-ai-scanned-two-column-text-study-area.png
Observed output: Output artifact (Image): The paragraph-level segmentation and reading order broke down on this scanned multi-column page, with structural formatting not maintained during extraction. — upstage_scannedpdf_parsed_hierarchy.png
Input artifact: Input artifact (Image): Scanned multi-column forestry page headed 'STUDY AREA' with 'STAND PRESCRIPTIONS' lower on the page. — landing-ai-scanned-two-column-text-study-area.png
Output artifact: Output artifact (Image): The paragraph-level segmentation and reading order broke down on this scanned multi-column page, with structural formatting not maintained during extraction. — upstage_scannedpdf_parsed_hierarchy.png
What changed: Image transformed into Image
Why it matters / Conclusion: Upstage can preserve straightforward section flow, but it was not reliable enough on multi-column or hierarchy-sensitive pages to trust for full-document markdown fidelity.
Upstage preserved basic narrative flow on at least one straightforward digital section, but it struggled to keep hierarchy and navigation intact on harder layouts. The Mechatronics/Industrial Machinery section from the Sumitomo report remained readable with subsection ordering preserved. By contrast, the hybrid earnings report's two-column strategy page blurred headings into surrounding prose, the operating-performance page lost clear heading distinction, and the researcher reported paragraph segmentation and reading-order breakdown on the scanned multi-column research paper.
Numbered Mechatronics / Industrial Machinery section from the financial report.
On the Mechatronics / Industrial Machinery section, Upstage kept the numbered subsection headings and paragraph flow in the same general order as the source, so the extracted text remained readable as a structured section rather than a flat dump.
Two-column Target narrative page with several distinct subheads.
On the two-column Target narrative page, Upstage preserved much of the text content but blurred the original navigation structure. Headings such as 'Target.com & mobile,' 'Local relevance and flexible formats,' and 'Target rewards' ran into surrounding prose, so the page no longer read like clearly separated sections.
Summary of Operating Performance page from the Sumitomo report.
On the 'Summary of Operating Performance' page, Upstage retained the substantive text but reduced section-level distinction. The researcher reported that headings were no longer visually or structurally separated from the content they introduced.
Scanned multi-column forestry page headed 'STUDY AREA' with 'STAND PRESCRIPTIONS' lower on the page.
The paragraph-level segmentation and reading order broke down on this scanned multi-column page, with structural formatting not maintained during extraction.
Scanned-page OCR for printed textPrinted text was partly recovered, but signature blocks and page structure were not faithfully preserved.▾
Feature tested: Scanned-page OCR for printed text
Result: Partial
Verdict: Printed text was partly recovered, but signature blocks and page structure were not faithfully preserved.
Expected behavior: Upstage can OCR printed text on scanned pages inside a mixed PDF, but the result is much less faithful once signatures and local structure matter. The tested example was the Target signatures page from the hybrid earnings report: surrounding printed text, date, and names were retained, but the handwritten signature itself was not meaningfully preserved and the section collapsed into a flat block.
Test case: Image → Image
Input type: Image
Input used: Input artifact (Image): Scanned signatures page from the Target annual report. — landing-ai-target-annual-report-signatures-page-2.png
Observed output: Output artifact (Image): On the scanned signatures page, Upstage recovered most of the surrounding printed text, the date, and the printed name 'Catherine R. Smith,' but handwritten sig — upstage-ai-target-signatures-ocr-extraction-1.png
Input artifact: Input artifact (Image): Scanned signatures page from the Target annual report. — landing-ai-target-annual-report-signatures-page-2.png
Output artifact: Output artifact (Image): On the scanned signatures page, Upstage recovered most of the surrounding printed text, the date, and the printed name 'Catherine R. Smith,' but handwritten sig — upstage-ai-target-signatures-ocr-extraction-1.png
What changed: Image transformed into Image
Why it matters / Conclusion: For scanned pages with ordinary printed text, Upstage can recover usable content, but it is not a good fit when the exact structure of signature sections or handwriting-adjacent content matters.
Upstage can OCR printed text on scanned pages inside a mixed PDF, but the result is much less faithful once signatures and local structure matter. The tested example was the Target signatures page from the hybrid earnings report: surrounding printed text, date, and names were retained, but the handwritten signature itself was not meaningfully preserved and the section collapsed into a flat block.
Scanned signatures page from the Target annual report.
On the scanned signatures page, Upstage recovered most of the surrounding printed text, the date, and the printed name 'Catherine R. Smith,' but handwritten signatures were not clearly captured and the section hierarchy collapsed into a flat text block. Signature-like strings such as 'CHeSmith' appeared instead of a meaningful preserved signature element.
Pricing & Access
Is This Right For You?
A side-by-side guide based on our hands-on testing.
Use Case Track
Related Pages
Banner Preview
How the embed badge will look on your site
Embed HTML
Copy this code to your website source
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
Discover more AI tools like Upstage AI to enhance your workflow.