We tested hosted PDF-to-markdown APIs on the same three hard documents: a long hybrid annual report, a table-heavy financial report, and an image-only scanned research paper. The goal was usable markdown with OCR, tables, charts, and reading order preserved well enough for downstream RAG, search, and reuse.
Tied with Extend AI on hybrid reports; best for programmatic chart extraction; scanned tables drop performance.
The scan text stayed readable on both dense prose and a blurred stamp, but the evidence is narrower than for a full 5 because there is not broad proof across many degraded pages.
The evidence-backed checks show the shape of the field; coverage explains the gaps.
Columns, left to right: Chart retention · Image retention · OCR quality on scans · Reading order & structure · Separate table/chart extraction (bonus) · Table structure preservation · Transparency
Ranking rule: tools measured on every decisive check rank above tools missing any, whatever their score. LlamaParse skipped Chart retention, Image retention, Transparency (scores 4.5 on the checks it ran); Tensorlake skipped Transparency (scores 3.7 on the checks it ran); Landing AI skipped Image retention, Separate table/chart extraction (bonus), Transparency (scores 3.5 on the checks it ran); Mistral AI skipped Separate table/chart extraction (bonus), Transparency (scores 3.4 on the checks it ran); Upstage AI skipped Chart retention, Image retention, OCR quality on scans, Transparency (scores 3.3 on the checks it ran); Nutrient.io skipped Separate table/chart extraction (bonus), Transparency (scores 2.2 on the checks it ran); Adobe PDF Extract API skipped Chart retention, Image retention, OCR quality on scans, Reading order & structure, Separate table/chart extraction (bonus), Table structure preservation, Transparency.
Pick the tools you care about, then compare what they returned or how they scored.
It did a solid job on the report hierarchy and a key multilevel table, but the more complex table and the table of contents lost some structure, so the overall result is good but not consistently strong.
llamaparse-financialpdf-title-page-6b8eb4c8c284.png
It kept the narrative section and the simpler segment table in good shape, but the more complex multi-header table broke down, so the overall result is solid but uneven.
4a1b21a5855f4064b049eaf39823a417.png
It preserves the report’s section flow and a core financial table, but the more complex grouped headers are flattened enough to keep it from scoring higher.
1967e1c5f9944428a4dbc41122a50e81.png
It handled the table-heavy financial report well enough to keep the segment tables and section flow readable, but the first page headings and a nested table header lost structure.
1036a7a2e37f4e41a92554b56a4d59ee.png
It kept the financial report readable and recovered the key tables, but the table of contents was flattened and one multilevel table lost header clarity, so the structure was only partly preserved.
f5cbd244ad274873a5b0d91b67d8e1ff.png
It preserved one narrative subsection cleanly, but the balance-sheet table and another section page lost structure, so the output is uneven rather than consistently strong.
cc1490a021354e44ad8192bc5e979001.png
It could recover at least one section page, but the disclaimer pages got fragmented and the multi-level table lost its header structure, so the report came out uneven.
0dc4cf41af00402b8936af637821d46f.png
We have no recorded result for Adobe PDF Extract API on this prompt, so there is nothing to compare here.
Not part of this comparison
All 7 recorded checks per tool. Open a tool to inspect every finding.
The charts are not kept as pictures, but their values and labels are carried over faithfully, which is strong enough for a high score without being a perfect visual retention result.
Scored, but no finding or artifact was recorded. We do not present the number as proof.
LlamaParse is the page’s winner, and that matches the scorecard: it leads on hierarchy/reading order, table handling, and scanned OCR, which are the decisive strengths on this page. The caveat is that it is only partly tested (4 of 10 decisive checks), so the evidence is narrower than for some lower-ranked tools; still, by the published rank order it is the overall pick. If you need document flow plus ordinary table reconstruction, Tensorlake is the next fit, but it is weaker on hierarchical tables and image retention. Extend AI is a solid alternative for preserving flow and charts, but its complex-table and image-handling results are less dependable. Landing AI stands out for chart/table recovery, yet its weaker headings and scanned OCR make it less balanced. Upstage AI is best when the goal is extracting tables and charts into usable text, but it does not preserve page layout as well. Adobe API has no decisive checks here, so there is not enough evidence to place it meaningfully.
The tools we tested for this use case — each card opens its full tested review.
If you are looking to build a custom PDF-to-markdown conversion, OCR extraction, or document parsing system for your business or internal workflow, email us at contact@futuresmart.ai.
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.
Comments (0)