Every output below is the verbatim Markdown returned by the production API on 2026-10-19 — no retyping, no cleanup. Source files are linked so you can reproduce each conversion yourself.
1 · Business report PDF → Markdown
A single-column quarterly report with two data tables. Headings map to H1/H2, both tables come out as GitHub-flavored Markdown, lists are preserved. This is the shape most business documents convert in.
# Q3 2026 Regional Sales Report Prepared by the Analytics Team · October 2026. This report summarizes quarterly revenue by region and product line, based on closed-won deals recorded in the CRM. ## Revenue by region | Region | Q2 ($M) | Q3 ($M) | QoQ change | | - | - | - | - | | AMER | 4.20 | 5.10 | +21.4% | | EMEA | 2.70 | 3.30 | +22.2% | | APAC | 1.90 | 2.80 | +47.4% | ## Outlook APAC growth is driven by two enterprise logos signed in September. We expect Q4 revenue between $12.0M and $12.8M, assuming no churn among the top ten accounts. - Hiring: two account executives for EMEA in November. - Risk: FX exposure on APAC contracts; hedging policy under review.
2 · CSV → Markdown table
Spreadsheets become pipe tables, cell for cell — ready to paste into a README, docs page or LLM prompt.
# headcount.csv ## Sheet1 | employee | department | start_date | annual_cost | |--------------|-------------|------------|-------------| | Alice Chen | Engineering | 2023-03-14 | 142000 | | Bruno Silva | Sales | 2024-07-01 | 98000 | | Chika Okafor | Support | 2022-11-20 | 76500 | | Dana Petrov | Engineering | 2025-01-06 | 151000 |
3 · Two-column academic paper — and where it struggles
We also converted Attention Is All You Need (arXiv 1706.03762) and are showing the result as-is. Body text and section structure come through, but the complex multi-column comparison table gets linearized into run-on text instead of a Markdown table. If your documents look like this, test with your own files before committing — that trade-off is real today.
### Page 6 Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. n is the sequence length, d is the representation dimension, k is the kernel size of convolutions and r the size of the neighborhood in restricted self-attention. Layer Type Complexity per Layer Sequential Maximum Path Length Operations Self-Attention O(n2 · d) O(1) O(1)Recurrent O(n · d2) O(n) O(n)Convolutional O(k · n · d2) O(1) O(logk(n))Self-Attention (restricted) O(r · n · d) O(1) O(n/r)
The full honest-limits list
- Tables in single-column documents survive; tables in dense multi-column academic layouts may be linearized (sample 3).
- Output is GitHub-flavored Markdown — no LaTeX, no MathML. Formulas come out as plain text approximations.
- Files are capped at 15 MB each, on every plan.
- 1 page of quota ≈ 4,000 characters of output Markdown; PDF, EPUB and PPTX page breaks are counted directly.
- Failed conversions never consume quota — you only spend pages on output you receive.