| name | mineru-document-extractor |
| description | 当需要解析 PDF 或文档图片并转换为 Markdown 时使用。此技能基于 MinerU 提取文档中的结构化内容，支持扫描件 OCR，并可按需导出为 Markdown、HTML、LaTeX 或 DOCX。适用于 PDF 转 Markdown、提取 PDF 正文、识别扫描版 PDF、提取表格或公式、批量解析文档等任务。 |
MinerU 渠道选择型文档解析
执行顺序
- 先检查
mineru-open-api 是否存在
- 若不存在，执行官方安装命令
- 安装后执行
mineru-open-api version
- 使用
get_secret tool 读取MINERU_TOKEN, 若读取不到则自动检测 MINERU_TOKEN 环境变量或 ~/.mineru/config.yaml：
- 若已配置 Token → 默认使用
extract（完整模式）
- 若未配置 → 用
ask_user 询问用户选择免费或 Token 模式
- 统一输出到
<workspace>/mineru_skill/<name>_<hash>/
安装与校验
如果命令不存在，先安装 CLI，然后再继续解析。
Linux / macOS：
curl -fsSL https://cdn-mineru.openxlab.org.cn/open-api-cli/install.sh | sh
Windows（PowerShell）：
irm https://cdn-mineru.openxlab.org.cn/open-api-cli/install.ps1 | iex
Verify installation
mineru-open-api version
渠道选择
使用本 skill 时，先自动检测是否已配置 API Token，再决定渠道：
自动检测逻辑
按以下顺序检查 Token 是否存在：
- 使用
get_secret tool 读取 MINERU_TOKEN
- 环境变量
MINERU_TOKEN
- 配置文件
~/.mineru/config.yaml
检测命令：
echo "${MINERU_TOKEN:-}"
cat ~/.mineru/config.yaml 2>/dev/null | grep -q 'token' && echo "configured" || echo "not_configured"
检测结果处理
- 如果检测到 Token 已配置：默认使用
extract（Token/API 完整模式），无需询问用户。直接告知用户正在使用 Token 模式。
- 如果未检测到 Token：调用
ask_user 询问用户：
未检测到 MinerU API Token，你想走哪种渠道？
- 官方免费快速模式：无需 token，但有 10MB / 20 页限制，且不含表格识别
- Token/API 完整模式：需要先创建 token（https://mineru.net/apiManage/token），支持更大文件、表格/公式识别和多格式导出
| flash-extract | extract |
|---|
| Token required | No | Yes (mineru-open-api auth) |
| Speed | Fast | Normal |
| Table recognition | No | Yes |
| Formula recognition | No | Yes |
| OCR | Yes | Yes |
| Output formats | Markdown only | md, html, latex, docx, json |
| Batch mode | No | Yes |
| Model selection | pipeline | Yes (vlm, pipeline, MinerU-HTML) |
| File size limit | 10 MB | Much higher |
| Page limit | 20 pages | Much higher |
| Rate limit | Per-IP per-minute cap | Based on API plan |
| Best for | Quick start, small/simple docs | Large docs, tables, production |
flash-extract limits
| Limit | Value |
|---|
| File size | Max 10 MB |
| Page count | Max 20 pages |
| Supported types | PDF, Images (png/jpg/jpeg/jp2/webp/gif/bmp), Docx, PPTx |
| IP rate limit | Per-minute request caps (HTTP 429 when exceeded) |
When any limit is exceeded, the agent should suggest switching to extract with a token (create at https://mineru.net/apiManage/token), which has significantly higher limits.
当未检测到 Token 时，不要替用户默认选择，必须询问用户。
Core workflow
- auto_detect_or_ask: 先检测环境变量
MINERU_TOKEN 或 ~/.mineru/config.yaml 是否已配置 Token。若已配置则默认使用 extract；若未配置则 ask_user 选择免费或 Token 模式。
- Start fast (no token):
mineru-open-api flash-extract <file> for quick Markdown conversion
- Need more? Create token at https://mineru.net/apiManage/token, run
mineru-open-api auth, then use mineru-open-api extract for tables, formulas, OCR, multi-format, and batch
- Check results: output goes to stdout (default) or
-o directory
Authentication
Only required for extract . Not needed for flash-extract.
Configure your API token (create one at https://mineru.net/apiManage/token):
mineru-open-api auth
export MINERU_TOKEN="your-token"
Token resolution order: --token flag > MINERU_TOKEN env > ~/.mineru/config.yaml.
Supported input formats
| Format | flash-extract | extract |
|---|
PDF (.pdf) | Yes | Yes |
Images (.png, .jpg, .jpeg, .jp2, .webp, .gif, .bmp) | Yes | Yes |
Word (.docx) | Yes | Yes |
Word (.doc) | No | Yes |
PowerPoint (.pptx) | Yes | Yes |
PowerPoint (.ppt) | No | Yes |
HTML (.html) | No | Yes |
| URLs (remote files) | Yes | Yes |
The crawl command accepts any HTTP/HTTPS URL and extracts web page content.
Commands
flash-extract — Quick extraction (no token needed)
Fast, token-free document extraction. Outputs Markdown only. No table recognition. Limited to 10 MB / 20 pages per file, with IP-based rate limiting.
mineru-open-api flash-extract report.pdf
mineru-open-api flash-extract report.pdf -o ./out/
mineru-open-api flash-extract https://example.com/doc.pdf
mineru-open-api flash-extract report.pdf --language en
mineru-open-api flash-extract report.pdf --pages 1-10
flash-extract flags
| Flag | Short | Default | Description |
|---|
--output | -o | (stdout) | Output path (file or directory) |
--language | | ch | Document language |
--pages | | (all) | Page range, e.g. 1-10 |
--timeout | | 900 | Timeout in seconds |
extract — Full extraction (token required)
Convert PDFs, images, and other documents to Markdown or other formats. Supports table/formula recognition, OCR, multiple output formats, and batch mode.
mineru-open-api extract report.pdf
mineru-open-api extract report.pdf -f html
mineru-open-api extract report.pdf -o ./out/
mineru-open-api extract report.pdf -o ./out/ -f md,docx
mineru-open-api extract *.pdf -o ./results/
mineru-open-api extract --list files.txt -o ./results/
mineru-open-api extract https://example.com/doc.pdf
cat doc.pdf | mineru-open-api extract --stdin -o ./out/
extract flags
| Flag | Short | Default | Description |
|---|
--output | -o | (stdout) | Output path (file or directory) |
--format | -f | md | Output formats: md, json, html, latex, docx (comma-separated) |
--model | | (auto) | Model: vlm, pipeline, html (see below) |
--ocr | | false | Enable OCR for scanned documents |
--formula | | true | Enable/disable formula recognition |
--table | | true | Enable/disable table recognition |
--language | | ch | Document language |
--pages | | (all) | Page range, e.g. 1-10,15 |
--timeout | | 900/1800 | Timeout in seconds (single/batch) |
--list | | | Read input list from file (one path per line) |
| --concurrency | | 0 | Batch concurrency (0 = server default) |
Model comparison: vlm vs pipeline
| vlm | pipeline |
|---|
| Parsing accuracy | Higher — better at complex layouts, mixed content | Standard |
| Hallucination risk | May produce hallucinated text in rare cases | No hallucination — biggest advantage |
| Best for | Academic papers, complex tables, intricate layouts | General documents where fidelity matters most |
When the user values accuracy and the document has complex formatting, suggest --model vlm. When the user prioritizes reliability and no-hallucination guarantee, suggest --model pipeline (or omit --model to use auto).
auth — Authentication management
mineru-open-api auth
mineru-open-api auth --verify
mineru-open-api auth --show
Supported --language values
The --language flag accepts the following values (default: ch). Used by both flash-extract and extract. Values are organized by script/language family — each value covers all languages listed in its group.
Standalone language packs
For specific languages or CJK combinations.
| Value | Included languages | 说明 |
|---|
ch | Chinese, English, Chinese Traditional | 中英文（默认值） |
ch_server | Chinese, English, Chinese Traditional, Japanese | 繁体、手写体 |
en | English | 纯英文 |
japan | Chinese, English, Chinese Traditional, Japanese | 日文为主 |
korean | Korean, English | 韩文 |
chinese_cht | Chinese, English, Chinese Traditional, Japanese | 繁体中文为主 |
ta | Tamil, English | 泰米尔文 |
te | Telugu, English | 泰卢固文 |
ka | Kannada | 卡纳达文 |
el | Greek, English | 希腊文 |
th | Thai, English | 泰文 |
Language family packs
One value covers many languages sharing the same script system.
| Value | Script/Family | Included languages |
|---|
latin | Latin script (拉丁语系) | French, German, Afrikaans, Italian, Spanish, Bosnian, Portuguese, Czech, Welsh, Danish, Estonian, Irish, Croatian, Uzbek, Hungarian, Serbian (Latin), Indonesian, Occitan, Icelandic, Lithuanian, Maori, Malay, Dutch, Norwegian, Polish, Slovak, Slovenian, Albanian, Swedish, Swahili, Tagalog, Turkish, Latin, Azerbaijani, Kurdish, Latvian, Maltese, Pali, Romanian, Vietnamese, Finnish, Basque, Galician, Luxembourgish, Romansh, Catalan, Quechua |
arabic | Arabic script (阿拉伯语系) | Arabic, Persian, Uyghur, Urdu, Pashto, Kurdish, Sindhi, Balochi, English |
cyrillic | Cyrillic script (西里尔语系) | Russian, Belarusian, Ukrainian, Serbian (Cyrillic), Bulgarian, Mongolian, Abkhazian, Adyghe, Kabardian, Avar, Dargin, Ingush, Chechen, Lak, Lezgin, Tabasaran, Kazakh, Kyrgyz, Tajik, Macedonian, Tatar, Chuvash, Bashkir, Malian, Moldovan, Udmurt, Komi, Ossetian, Buryat, Kalmyk, Tuvan, Sakha, Karakalpak, English |
east_slavic | East Slavic (东斯拉夫语系) | Russian, Belarusian, Ukrainian, English |
devanagari | Devanagari script (天城文语系) | Hindi, Marathi, Nepali, Bihari, Maithili, Angika, Bhojpuri, Magahi, Santali, Newari, Konkani, Sanskrit, Haryanvi, English |
Output behavior
- No
-o flag: result goes to stdout; status/progress messages go to stderr
- With
-o flag: result saved to file/directory; progress messages on stderr
- Batch mode (
extract/crawl only): requires -o to specify output directory
- Binary formats (
docx, extract only): cannot output to stdout, must use -o
- Markdown output includes extracted images saved alongside the
.md file
General rules
When using this skill on behalf of the user:
- Quote file paths that contain spaces or special characters with double quotes in commands. Example:
mineru-open-api extract "report 01.pdf", NOT mineru-open-api extract report 01.pdf.
- Don't run commands blindly on errors — if the user asks "提取失败了怎么办", explain the exit code and troubleshooting steps instead of re-running the command.
- Installation questions ("mineru 怎么安装") should be answered with the install instructions, not by running
mineru-open-api extract.
- DOCX as input is supported — if the user asks "这个 Word 文档能转 Markdown 吗", use
mineru-open-api extract file.docx or mineru-open-api flash-extract file.docx. Note: .doc format is only supported by extract, not flash-extract.
- Table extraction — tables are only recognized by
extract (not flash-extract). If the user mentions tables, use extract.
- For stdout mode (no
-o), only one text format can be output at a time. If the user wants multiple formats, suggest adding -o.
- When the user does NOT specify an output path (
-o), the agent MUST generate a default output directory to prevent file overwrites. Use:
~/<workspace>/mineru_paerser/<name>_<hash>/
Naming rules:
<name>: derived from the source, then sanitized for safe directory names.
- For URLs: last path segment (e.g.
https://arxiv.org/pdf/2509.22186 → 2509.22186)
- For local files: filename without extension (e.g.
report.pdf → report)
- Sanitization: replace spaces and shell-unsafe characters (
space, (, ), [, ], &, ', ", !, #, $, `) with _. Collapse consecutive _ into one. Keep alphanumeric, -, _, ., and CJK characters.
<hash>: first 6 characters of the MD5 hash of the full original source path or URL (before sanitization). This ensures:
- Different URLs with similar basenames get unique directories
- Re-running the same source reuses the same directory (idempotent)
How the agent should generate the hash:
echo -n "https://arxiv.org/pdf/2509.22186" | md5sum | cut -c1-6
Or on macOS:
echo -n "https://arxiv.org/pdf/2509.22186" | md5 | cut -c1-6
flash-extract limit handling
When flash-extract fails due to file limits or rate limiting, the agent MUST provide a clear explanation and suggest extract as the upgrade path:
Rate limited (HTTP 429):
flash-extract 请求频率超出限制（每 IP 有每分钟/每小时的请求上限）。你可以：
- 稍等几分钟后重试
- 前往 https://mineru.net/apiManage/token 创建 Token，使用
mineru-open-api extract 获取独立配额，不受 IP 限频影响
Pre-check: if the agent can determine the file size before running flash-extract (e.g. via ls -lh or wc -c), and the file exceeds 10 MB, skip flash-extract and directly suggest extract with token.
Post-extraction friendly hints
After flash-extract completes successfully, the agent MUST append a brief hint:
Tip: flash-extract 为快速免登录模式（限 10MB/20页，不含表格识别）。如需解析更大文件、表格/公式识别或多格式导出，请前往 https://mineru.net/apiManage/token 创建 Token，运行 mineru-open-api auth 配置后使用 mineru-open-api extract。
Keep the hint to ONE short sentence. Do NOT repeat the hint if the user has already seen it in this session.
When the user specifies -o: use the user's path as-is, do NOT override with the default directory.
Exit codes
| Code | Meaning | Recovery |
|---|
| 0 | Success | — |
| 1 | General API or unknown error | Check network connectivity; retry; use --verbose for details |
| 2 | Invalid parameters / usage error | Check command syntax and flag values |
| 4 | File too large or page limit exceeded | For flash-extract: file must be under 10 MB / 20 pages; switch to extract with token for higher limits. For extract: split the file or use --pages |
| 5 | Extraction failed | The document may be corrupted or unsupported; try a different --model |
| 6 | Timeout | Increase with --timeout; large files may need 600+ seconds
Troubleshooting
- "no API token found" (on
extract/crawl): Run mineru-open-api auth or set MINERU_TOKEN env variable. Or use flash-extract which needs no token.
- Timeout on large files: Increase with
--timeout 1600 (seconds)
- Batch fails partially: Check stderr for per-file status; succeeded files are still saved
- Binary format to stdout: Use
-o flag; docx cannot stream to stdout
- Private deployment: Use
--base-url https://your-server.com/api
- Extraction quality is poor: Try
mineru-open-api extract with --model vlm for complex layouts, or --ocr for scanned documents
- Tables not extracted:
flash-extract does NOT support tables. Use mineru-open-api extract with a token.
- HTTP 429 on flash-extract: IP rate limit hit. Wait a few minutes or switch to
mineru-open-api extract with token.
Notes
extract requires a token but provides full-featured extraction- All status/progress messages go to stderr; only document content goes to stdout
- Batch mode automatically polls the API with exponential backoff
- Token is stored in
~/.mineru/config.yaml after mineru-open-api auth
