--- title: "docTR" type: "AI Tool" url: "https://aidemos.com/tools/doctr" description: "We ran docTR on PDFs and got solid OCR for prose and digits, but it flattened tables, headings, images, and chart structure." category: "developer-tools" website: "https://github.com/mindee/doctr" published: "2026-08-19T17:11:08.792174+00:00" updated: "2026-09-11T10:51:01.585326+00:00" --- # docTR Fast OCR for PDFs when you need raw text, digits, and speed more than preserved Markdown structure. ## TL;DR Verdict **Fast OCR, but not a Markdown converter** **Where it wins:** - You need a self-hosted OCR engine for PDFs and can post-process the raw text yourself. - You care more about fast text and digit extraction than about preserving Markdown structure. - You need a library that handled long PDFs and a fully scanned PDF without crashing in this benchmark. **Main limitation:** You need real Markdown headings, tables, image references, or document hierarchy from the parser itself. **Pricing:** Open source Free `Open source` · `Scanned OCR` · `Plain text export` · `No Markdown structure` **Website:** [Visit docTR](https://github.com/mindee/doctr) ## How docTR behaved on the tested PDFs Same benchmark set across three document archetypes: hybrid annual report, table-heavy financial filing, and a scanned research note. - **1** Financial PDF — Best overall fit for docTR: numeric transcription, reading order, and footnotes were the strongest here. - **2** Hybrid earnings report — OCR is fast, but charts, two-column prose, and sparse-table structure degrade noticeably. - **3** Scanned research paper — Genuine OCR works, but page-1 reading order and some glyph-level recognition were the weakest here. > **Fast OCR, but not a Markdown converter** > > docTR is very fast and its OCR is solid on both digital and scanned PDFs, with most prose and digits surviving intact. The hard limit is structural: it only emits flat page text, so tables, headings, images, and chart structure are lost. For PDF-to-Markdown, it works best as a raw OCR layer you clean up yourself. ## Demo Recording [Video: docTR demo recording](https://cdn.futuresmart.ai/public/aidemos/340f5b40af384319841b4c71d08efdb5.mp4?v=1) *Video — Google Colab walkthrough of the docTR pipeline on the financial PDF; shows upload, model setup, page-count check, markdown writing, and download.* ## Feature-by-Feature Breakdown ### PDF OCR Text Extraction **Verdict:** Strong raw OCR, but the export stays flat and page-based. docTR converts hybrid and scanned PDFs into searchable plain text, as shown on the earnings report, the financial filing, and the scanned research note. The evidence covers long documents that complete without crashing, but the output remains flat text rather than Markdown. **Input:** > **Image** **Output:** > **Image** **Input:** > **Image** **Output:** > **Image** **Input:** Hybrid earnings PDF > **File** — Hybrid earnings PDF **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Financial PDF > **File** — Financial PDF **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Scanned research paper PDF > **File** — Scanned research paper PDF **Output:** docTR output markdown > **File** — docTR output markdown **Bottom line:** Good at turning PDFs into searchable text, including scans, but it stops at plain text rather than Markdown structure. ### Table OCR Transcription **Verdict:** Table numbers are usually transcribed correctly and kept in left-to-right order, but blank cells and some headers lose structure. docTR reads table contents from hybrid, dense, and scanned documents, preserving many numeric sequences and row values in left-to-right order. The examples also show structural limits such as omitted blank cells, sparse headers, and occasional noise on difficult scans. **Input:** Credit card expenses row crop > **File** — Credit card expenses row crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Repeated segment-table header crop > **File** — Repeated segment-table header crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Dense scanned table crop > **File** — Dense scanned table crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Quarterly results table > **Image** — Quarterly results table **Output:** docTR output > **Image** — docTR output **Input:** Aspen row crop > **Image** — Aspen row crop **Output:** docTR output > **Image** — docTR output **Bottom line:** Numeric transcription is strong, but structural table fidelity is only partial because empty cells and some header labels have no stable representation. ### Reading Order Reconstruction **Verdict:** Works on straightforward single-column pages, but two-column prose and mixed title/sidebar layouts can be interleaved or zipped together. docTR reconstructs reading order on straightforward single-column pages and row-major lists, but becomes unreliable when multiple text streams or mixed layouts are present. The evidence includes simple financial pages as well as two-column and masthead/title pages where text can interleave. **Input:** Two-column CEO letter crop > **File** — Two-column CEO letter crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Single-column assets table crop > **File** — Single-column assets table crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Scanned masthead/title crop > **File** — Scanned masthead/title crop **Output:** docTR output markdown > **File** — docTR output markdown **Bottom line:** Good on simple layouts, unreliable on multi-stream pages where content should stay separated. ### Caption and Footnote Association **Verdict:** Captions usually stay attached, but in-table reference markers can be garbled. docTR keeps figure/table captions and numbered footnotes near the content they describe in extracted output. The examples include financial tables and figure pages where nearby notes stay grouped even when local OCR around the markers is imperfect. **Input:** Net debt footnote-marker crop > **File** — Net debt footnote-marker crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Segment notes crop > **File** — Segment notes crop **Output:** docTR output markdown > **File** — docTR output markdown **Input:** Figure 1 chart crop > **File** — Figure 1 chart crop **Output:** docTR output markdown > **File** — docTR output markdown **Bottom line:** Association is usually good, but corrupted in-table markers can still make the reference chain unreliable. ### Visual Element Preservation **Verdict:** Images are not retained as images, and charts lose their visual content. docTR attempts to carry non-text visuals such as cover art, photos, signatures, logos, seals, and charts through extraction, using examples from cover pages and portrait/signature pages. In practice, those visuals are often omitted or degraded into noise. **Input:** Cover art > **Image** — Cover art **Output:** docTR output > **Image** — docTR output **Input:** CEO photo and signature page > **Image** — CEO photo and signature page **Output:** docTR output > **Md** — docTR output **Input:** Header logo banner > **Image** — Header logo banner **Output:** docTR output > **Image** — docTR output **Input:** Masthead seal page > **Image** — Masthead seal page **Output:** docTR output > **Md** — docTR output **Input:** Financial highlights chart > **Image** — Financial highlights chart **Output:** docTR output > **Image** — docTR output **Bottom line:** If you need figures, logos, or charts preserved in the output, docTR does not do that; at best it drops them, and at worst it turns them into noise. ## Free and open source Apache 2.0; no API fees. | Plan | Price | Notes | | --- | --- | --- | | Open source ★ (tested) | Free | Apache 2.0. No recurring API cost; you pay only for your own compute. | *The benchmark ran in Google Colab on a T4 GPU, at roughly 0.5–0.7 seconds per page depending on the document.* ## Is It Right For You? **Use it if** - You need a self-hosted OCR engine for PDFs and can post-process the raw text yourself. - You care more about fast text and digit extraction than about preserving Markdown structure. - You need a library that handled long PDFs and a fully scanned PDF without crashing in this benchmark. **Skip it if** - You need real Markdown headings, tables, image references, or document hierarchy from the parser itself. - You need reliable reading order on two-column or mixed-layout pages. - You need charts, logos, or embedded images retained in the extracted output. ## Classification - **Category:** developer-tools - **Subcategory:** pdf-tools - **Type:** text - **Built for:** Other ## Frequently Asked Questions **Q: Does docTR output real Markdown?** No. In this benchmark it produced plain page-by-page OCR text, with page markers added by the notebook. Tables, headings, and images were not exported as Markdown structure. **Q: How well does docTR handle scanned PDFs?** It handled the fully scanned research paper successfully and completed the full document, but OCR errors still appeared on some glyphs and page-1 reading order was badly scrambled. **Q: Does docTR keep tables readable?** It usually preserves numeric values and row order, but it does not preserve full table structure. Blank cells vanish, sparse rows lose alignment, and some headers can disappear. **Q: Does docTR preserve charts, logos, or images?** No. Charts lose their data as structure, and embedded images or logos are not retained in the output. **Q: How does docTR handle two-column or mixed-layout pages?** Poorly. The tested two-column page was zipped line-by-line, and the scanned paper’s masthead/title/marginalia were interleaved into one stream. **Q: Is docTR free to use?** Yes. The report says docTR is open source under Apache 2.0 and free to use, with compute cost depending on your own hardware or Colab. ## Similar Tools AI tools similar to docTR: - [Docling](https://aidemos.com/tools/docling) — Open-source PDF-to-markdown conversion that is strong on text, headings, and standard tables, but drops charts and other visual assets. - [PyMuPDF4LLM](https://aidemos.com/tools/pymupdf4llm) — Open-source PDF-to-markdown for clean native PDFs, but unreliable on scans, dense tables, and images. - [LiteParse](https://aidemos.com/tools/liteparse) — Open-source PDF-to-markdown parsing that works well on native-digital reports, but degrades on scans, tables, and charts. - [Doc2Mark](https://aidemos.com/tools/doc2mark) — Open-source PDF-to-markdown that preserves native text and headings, but still struggles with tables, charts, images, and scans. - [MarkItDown](https://aidemos.com/tools/markitdown) — Fast native-PDF text extraction for markdown, but structure, charts, images, and scans are unreliable. - [MinerU](https://aidemos.com/tools/mineru) — Open-source PDF-to-markdown that keeps figures and charts in place, but tables and punctuation can get messy on complex files. - [PaddleOCR](https://aidemos.com/tools/paddleocr) — Open-source PDF-to-markdown conversion that keeps images and tables in place, but still corrupts text and complex structure on harder documents. - [Dolphin](https://aidemos.com/tools/dolphin) — Self-hosted PDF-to-Markdown that preserves most text and tables, but drops charts and struggles with repeated images. - [Unstructured](https://aidemos.com/tools/unstructured) — Good for text-first PDF-to-Markdown workflows, but not for faithful tables, charts, or images. ## Need a custom AI solution for this use case? If you are looking to build a custom OCR, PDF text extraction, or document processing system for your business or internal workflow, email us at [contact@futuresmart.ai](mailto: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](mailto:collaborate@aidemos.com).