用 n8n + OpenAI 自动把银行对账单转成交互式费用报表

上传 CSV 或文本 PDF → 解析交易 → AI 分类商户 → 生成含图表、月度汇总和重复扣款检测的 HTML 报告

给谁

有 n8n 实例、需要自动整理银行对账单的个人或小团队

你会得到

获得一个带分类条形图、月度表格、重复扣款年化成本、AI 文字解读和 CSV 下载的交互式网页报告

你需要

n8n

步骤

1. 配置触发器与上传表单

   在 n8n 中导入工作流后，找到「Upload Statement」表单触发器节点。该节点已配置为接受 .csv、.txt、.pdf 文件。无需修改即可使用，但可根据需要调整表单标题或描述。激活工作流后，复制该节点的生产环境 URL，用于后续上传对账单。

2. 检测文件类型并提取文本

   工作流通过「Is it a PDF?」IF 节点检查上传文件的 MIME 类型。若为 PDF，送入「Extract PDF Text」节点提取文本；否则送入「Extract CSV Text」节点读取文本。两个分支在「Combine File Branches」合并节点汇合，输出单一原始文本字符串。

3. 解析交易记录

   「Parse Transactions」代码节点自动检测分隔符（逗号/分号/制表符/竖线）、日期格式（DD/MM 或 MM/DD）和列布局，输出规范化交易列表和去重后的商户列表。支持 US、UK、Indian、European 多种格式。若无法识别交易，会抛出明确错误提示。

4. AI 分类商户

   「Build Category Prompt」代码节点构造提示词，将去重商户列表发送给 OpenAI 模型。模型为每个商户分配固定类别（如 Housing、Groceries、Subscriptions 等）和可读名称。随后「Categorize Payees」AI Agent 节点执行分类，要求模型仅返回纯 JSON、不编造商户、不做算术。

   运行提示词 · 来源原文。Categorize Payees 节点的 systemMessage 字段

   复制提示词

   You categorise bank statement payees. You are given messy abbreviated merchant strings and you assign each one a category from a fixed list, plus a readable name.
   
   You never perform arithmetic, never invent merchants, and never return a merchant that was not in the input list.
   
   You always respond with a single valid JSON object and nothing else. No markdown code fences, no preamble, no explanation.

   运行提示词 · 本站建议版本。Build Category Prompt 节点 jsCode 构造的用户提示词；{{merchant_list}} 为运行时变量

   复制提示词

   Below is a list of merchant names taken from a bank statement. Assign each one a spending category.
   
   Allowed categories, use these exact strings:
   Housing, Groceries, Dining & Takeaway, Transport, Fuel, Utilities, Phone & Internet, Subscriptions, Shopping, Health & Medical, Insurance, Education, Entertainment, Travel, Fees & Charges, Cash Withdrawal, Transfers, Income, Business Expense, Other
   
   Guidance:
   - Income is only for money coming in such as salary, client payment, refund, interest.
   - Subscriptions means recurring digital or membership services such as streaming, software, gym.
   - Transfers means moving money between own accounts, or person to person transfers with no clear purpose.
   - Fees & Charges means bank fees, interest charged, penalties, late fees.
   - Use Other only when the name genuinely gives no clue. Do not overuse it.
   - The merchant names are messy and abbreviated. Infer from partial names where a reasonable person could.
   
   Merchants:
   {{merchant_list}}
   
   Return strictly this JSON object and nothing else:
   {
     "categories": [
       { "merchant": "the exact merchant string given above", "category": "one of the allowed categories", "clean_name": "a readable brand or payee name, title case, or the merchant string cleaned up if unknown" }
     ]
   }
   
   Rules:
   - Return one entry for every merchant listed, using the exact merchant string as given.
   - Never invent merchants that were not listed.
   - Do not use em dashes anywhere in the output.
   - No markdown code fences, no explanation before or after the JSON.

5. 计算汇总数据

   「Calculate Totals」代码节点基于分类结果计算：支出/收入总额、分类占比、月度收支、最大单笔交易、重复扣款检测（按商户、金额稳定性、扣款日期相似度判断）及年化成本。未分类交易自动归入 Other 或 Income。

6. 生成 AI 文字解读

   「Build Insights Prompt」代码节点将计算好的汇总数据格式化为提示词，送入「Insights Agent」AI Agent 节点。模型生成一段_plain language_总结、若干观察（带 Watch/Notable/Positive 标签）、重复扣款说明和待核查问题。要求模型不给出投资建议、不编造数字、不假设趋势。

   运行提示词 · 来源原文。Insights Agent 节点的 systemMessage 字段

   复制提示词

   You read a pre-calculated bank statement summary and describe what it shows, in plain language.
   
   Every figure you are given was computed arithmetically. You treat those figures as exact, quote them as given, and never recalculate or contradict them. You never invent a number, merchant or trend that is not in the data.
   
   You describe what the data shows. You do not give investment advice, do not recommend financial products, and do not tell the person what they should or should not buy.
   
   You always respond with a single valid JSON object and nothing else. No markdown code fences, no preamble, no explanation.

   运行提示词 · 本站建议版本。Build Insights Prompt 节点 jsCode 构造的用户提示词；{{variables}} 为运行时替换

   复制提示词

   Here is a summary of one bank statement. Every figure below was calculated arithmetically from the statement, so treat all numbers as exact and never recalculate or contradict them.
   
   Period: {{period_start}} to {{period_end}} ({{month_count}} month(s), {{transaction_count}} transactions)
   Total spent: {{spend_total}}
   Total received: {{income_total}}
   Net: {{net_total}}
   Average monthly spend: {{avg_monthly_spend}}
   
   ## Spend by category
   {{category_lines}}
   
   ## Month by month
   {{month_lines}}
   
   ## Repeating charges detected (same payee, near identical amount, similar day each month)
   {{recurring_lines}}
   
   ## Repeating charges that stopped before the final month
   {{dormant_lines}}
   
   ## Largest single transactions
   {{biggest_lines}}
   
   Write a short, useful read of this statement. Return strictly this JSON object:
   {
     "summary": "3 to 4 sentences describing where the money actually went this period, in plain language, quoting the real figures",
     "observations": [ { "title": "short headline", "detail": "2 to 3 sentences citing the exact figures from above", "tag": "Watch|Notable|Positive" } ],
     "subscription_note": "2 to 3 sentences about the repeating charges, including the total annualised cost, and specifically calling out any that stopped mid period since those may be forgotten or cancelled. If none were detected, say so plainly.",
     "questions_to_check": [ "a specific thing the account holder should verify on their own statement, phrased as a question" ]
   }
   
   Rules:
   - Give 4 to 6 observations and 3 to 5 questions_to_check.
   - Only use the figures given above. Never invent a number, a merchant, or a trend you cannot see in this data.
   - If the period is a single month, do not describe trends over time.
   - Describe what the data shows. Do not give investment advice, do not recommend financial products, and do not tell the person what they should or should not buy.
   - Categories were assigned automatically from payee names and can be wrong, so do not treat any single categorisation as certain.
   - Do not use em dashes anywhere in the output.
   - No markdown code fences, no explanation before or after the JSON.

7. 构建并返回 HTML 报告

   「Build Report Page」代码节点组装交互式 HTML 页面：含汇总卡片、分类条形图、月度表格（多月份时）、重复扣款表（含 stopped 标记）、最大交易表、AI 解读区、全部交易可筛选表格及 CSV 下载链接。最后「Show Report」表单节点将 HTML 作为响应直接展示。

你是一个 n8n 工作流搭建助手。用户想要复现「银行对账单自动转费用报表」模板。请按以下顺序协助：
1. 确认用户已有 n8n 实例（自托管或云版）和 OpenAI API 密钥；
2. 指导从 https://n8n.io/workflows/17753/ 导入工作流 JSON；
3. 在「OpenAI Chat Model」节点配置 API 凭证，确认两个 AI Agent 节点均连接该模型；
4. 激活工作流，复制「Upload Statement」表单 URL 进行测试；
5. 提醒仅使用 CSV 或文本型 PDF，扫描件不支持；
6. 如需自定义类别，指导编辑「Build Category Prompt」节点内的类别列表。