任务：把 PDF 或扫描页转成可核对的表格Playbook: turn a PDF or scan into a checkable table

以下是本站制作的虚构订单，CC BY 4.0。两份 PDF 内容相同，一份有文字层，另一份是光栅图像。参考表按设计数据编写，不是运行 OCR 得到的结果，也不代表任何工具的识别率。输入为英文，便于比较字符与列。

• 可选中文字的 PDF
• 只有图像的 PDF
• 参考 CSV
• 校验说明

先只下载输入完成练习，再看参考表。真实文件先确认使用权限、敏感字段、上传目的地和保留政策；不确定时先用这里的虚构文件。

These are fictional orders created for this site under CC BY 4.0. Both PDFs have the same visual content; one has text and the other contains a raster image. The reference table is a designed answer, not OCR output or a tool accuracy claim. The input is in English so characters and columns are easy to compare.

• Text PDF
• Image-only PDF
• Reference CSV
• Checks

Try the input before opening the reference. For real documents, establish rights, sensitive fields, upload destination and retention rules first. Use the fictional input while those are unresolved.

交付一份 UTF-8 CSV 和一份异常清单。列为 item_id、description、quantity、unit_price、currency、line_amount、page、source_row、status。编号是文本；金额保留两位小数；缺失单价和金额留空并标状态，不写 0。page 是原 PDF 的页号，source_row 是该页数据行号，不是 CSV 文件行号。

只接受原页提供的信息。税、运费和应付总额均未提供；已知行金额之和不等于应付总额。

Deliver a UTF-8 CSV plus an exception log. Columns: item_id, description, quantity, unit_price, currency, line_amount, page, source_row, status. Treat IDs as text, retain two decimal places, and leave missing prices/amounts blank with an explicit status instead of zero. page is the original PDF page; source_row is its data-row index, not a CSV file line.

Use only source information. Tax, shipping and the payable total are absent. The sum of known line amounts is not a payable total.

1. 尝试选中并复制一句文字，再对照原页。能复制不等于顺序和表格结构正确；检查是否有隐藏 OCR 层。
2. 有可靠文字层时先抽取文字和表格；只有图像时才走 OCR。保留原文件和每页图像，不覆盖原件。
3. 扫描有旋转、裁边或模糊时，先看原图，再调整角度与分辨率。改善后的图仍需回原件核对。Tesseract 文档单列了表格识别限制，识别字符并不等于恢复列结构。
4. 先试这一页：编号 001 的名称换行仍是一条商品；003 的空单价仍为空。逐格校对后才推广到更多页。
5. 输出表与异常单，保留页号、原行号和修改记录。真实多页材料另查重复表头、跨页行、脚注和合并单元格。

1. Try selecting and copying a sentence, then compare it with the page. Copyable text does not prove correct reading order or table structure; check for a hidden OCR layer.
2. Extract text/tables from a usable text layer first; use OCR for image-only input. Preserve originals and page images.
3. Inspect rotation, cropping and blur before changing image angle or resolution. Check processed images against originals. Tesseract documents table-recognition limitations; recognizing characters does not establish column structure.
4. Trial this page: the wrapped description for 001 is one item, and the missing price for 003 stays blank. Check individual cells before scaling up.
5. Export the table and exceptions with page/row evidence and a change log. For real multi-page input, also check repeated headers, split rows, footnotes and merged cells.

将提供的订单页整理成 CSV，不查询外部信息。
字段：item_id,description,quantity,unit_price,currency,line_amount,page,source_row,status。
编号按文本保留前导零；换行名称合并为同一行。
缺失值留空并写 status，不猜价格、税率或总额。
另列无法确认的单元格和对应原页位置。先给一页结果供核对。

工具若不支持原页定位，另外保存一份逐行对照记录；没有定位时不要宣称结果可追溯。

Convert the supplied order page to CSV without external lookup.
Fields: item_id,description,quantity,unit_price,currency,line_amount,page,source_row,status.
Keep leading zeros in text IDs; join a wrapped description within one item.
Leave missing values blank and give a status. Do not guess prices, tax or totals.
List unresolved cells with their source-page locations. Return one page for review first.

If the tool cannot retain locations, maintain a separate row-by-row evidence record. Do not call the result traceable without one.

• 正好 3 条数据行，编号 001、002、003 保持原样。
• 001 的 description 包含 Notebook, dotted，不能拆成两个商品；CSV 中的逗号需正确引用。
• 2 × 12.50 = 25.00；3 × 4.00 = 12.00；已知金额小计为 CNY 37.00。
• 003 的 unit_price、line_amount 为空，status 为 unit_price_missing。
• 每行能回到第 1 页的对应数据行；另列未提供税和运费。

可用 CSV 解析器重新打开导出文件，确认列数和字符串类型；在表格软件中导入编号列时指定文本。参考答案只是验收口径，接受等价列名，但不能丢失证据或未知项。

• Exactly three data rows, with IDs 001, 002 and 003 unchanged.
• The description for 001 contains Notebook, dotted; it is not two products. Quote the comma correctly in CSV.
• 2 × 12.50 = 25.00; 3 × 4.00 = 12.00; known amounts total CNY 37.00.
• Row 003 has blank unit_price and line_amount, with unit_price_missing status.
• Each row maps to its original data row on page 1. Note that tax and shipping are absent.

Reopen the export with a CSV parser to check column counts and string values. In spreadsheet import, set the ID column to text. Equivalent column names are acceptable; losing evidence or unknowns is not.

• 无法选中文字：确认是否为图像页，再选 OCR；文件损坏或权限问题看 文件排查。
• 名称断行、列错位：先用原页确定行列，再修抽取规则，不让模型补齐看不见的格。
• 金额对不上：区分识别错误、计算错误和源文件本身不一致；异常留在记录中。
• CSV 无法解析：查分隔符、引号、编码与换行，再看 结构化输出排查。

• No selectable text: establish whether the page is image-only, then choose OCR. For damaged or restricted files, use file troubleshooting.
• Broken names or shifted columns: establish the source grid before fixing extraction rules; do not fill unseen cells.
• Wrong totals: distinguish recognition, arithmetic and source inconsistencies; keep exceptions visible.
• CSV parse failures: check delimiter, quoting, encoding and newlines; see structured-output troubleshooting.

把这张表作为输入，继续 多表合并练习。需要判断输出是否有依据，读 怎么验证 AI 的输出。

2026-11-04 核对了本文列出的官方说明范围；本站仅生成并检查示例 PDF 的文字层、图像版和参考文件，未运行 OCR 或测量任何工具表现。

Continue with the multi-table merge exercise. For evidence checks, read verifying AI output.

The cited documentation scope was checked on 2026-11-04. This site generated and checked the text/image PDF fixtures and reference files; it did not run OCR or measure a product.