把发票图片交给视觉模型，确实能省掉手输；但"模型说是 128.00 元"不等于系统该付款。更稳妥的做法是：模型只产生候选字段，本地程序以 Decimal 重算含税金额、检查币种和上限，再把通过或拒绝的理由交给界面。你将得到一个可以复制的最小链路，也能把同样的门禁迁移到合同、收据和质检单。
OpenAI 的 Images and vision 文档说明 Responses API 可接收图像输入。本文用 REST 而非 SDK，目的是让请求、超时和错误分支对初学者可见。模型 ID 请以目标账号的 Models 页面为准；代码中的 gpt-6 是示例占位，调用前替换为你账户可用模型。
#mermaid-svg-3thsV6f2MnrgkRWV{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-3thsV6f2MnrgkRWV .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-3thsV6f2MnrgkRWV .error-icon{fill:#552222;}#mermaid-svg-3thsV6f2MnrgkRWV .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-3thsV6f2MnrgkRWV .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-3thsV6f2MnrgkRWV .marker{fill:#333333;stroke:#333333;}#mermaid-svg-3thsV6f2MnrgkRWV .marker.cross{stroke:#333333;}#mermaid-svg-3thsV6f2MnrgkRWV svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-3thsV6f2MnrgkRWV p{margin:0;}#mermaid-svg-3thsV6f2MnrgkRWV .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-3thsV6f2MnrgkRWV .cluster-label text{fill:#333;}#mermaid-svg-3thsV6f2MnrgkRWV .cluster-label span{color:#333;}#mermaid-svg-3thsV6f2MnrgkRWV .cluster-label span p{background-color:transparent;}#mermaid-svg-3thsV6f2MnrgkRWV .label text,#mermaid-svg-3thsV6f2MnrgkRWV span{fill:#333;color:#333;}#mermaid-svg-3thsV6f2MnrgkRWV .node rect,#mermaid-svg-3thsV6f2MnrgkRWV .node circle,#mermaid-svg-3thsV6f2MnrgkRWV .node ellipse,#mermaid-svg-3thsV6f2MnrgkRWV .node polygon,#mermaid-svg-3thsV6f2MnrgkRWV .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-3thsV6f2MnrgkRWV .rough-node .label text,#mermaid-svg-3thsV6f2MnrgkRWV .node .label text,#mermaid-svg-3thsV6f2MnrgkRWV .image-shape .label,#mermaid-svg-3thsV6f2MnrgkRWV .icon-shape .label{text-anchor:middle;}#mermaid-svg-3thsV6f2MnrgkRWV .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-3thsV6f2MnrgkRWV .rough-node .label,#mermaid-svg-3thsV6f2MnrgkRWV .node .label,#mermaid-svg-3thsV6f2MnrgkRWV .image-shape .label,#mermaid-svg-3thsV6f2MnrgkRWV .icon-shape .label{text-align:center;}#mermaid-svg-3thsV6f2MnrgkRWV .node.clickable{cursor:pointer;}#mermaid-svg-3thsV6f2MnrgkRWV .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-3thsV6f2MnrgkRWV .arrowheadPath{fill:#333333;}#mermaid-svg-3thsV6f2MnrgkRWV .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-3thsV6f2MnrgkRWV .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-3thsV6f2MnrgkRWV .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-3thsV6f2MnrgkRWV .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-3thsV6f2MnrgkRWV .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-3thsV6f2MnrgkRWV .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-3thsV6f2MnrgkRWV .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-3thsV6f2MnrgkRWV .cluster text{fill:#333;}#mermaid-svg-3thsV6f2MnrgkRWV .cluster span{color:#333;}#mermaid-svg-3thsV6f2MnrgkRWV div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-3thsV6f2MnrgkRWV .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-3thsV6f2MnrgkRWV rect.text{fill:none;stroke-width:0;}#mermaid-svg-3thsV6f2MnrgkRWV .icon-shape,#mermaid-svg-3thsV6f2MnrgkRWV .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-3thsV6f2MnrgkRWV .icon-shape p,#mermaid-svg-3thsV6f2MnrgkRWV .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-3thsV6f2MnrgkRWV .icon-shape .label rect,#mermaid-svg-3thsV6f2MnrgkRWV .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-3thsV6f2MnrgkRWV .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-3thsV6f2MnrgkRWV .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-3thsV6f2MnrgkRWV :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 一致且未超限
异常
授权票据图片
base64视觉请求
模型返回JSON候选
Decimal本地复算
进入人工确认
拒绝并显示原因
环境准备与完整代码
Python 3.10+，安装 pip install requests，设置 export OPENAI_API_KEY='...'，再运行 python receipt_gate.py receipt.jpg。示例未在本次任务中实际调用线上 API；已在本机完成语法检查。
python
import base64, json, os, sys
from decimal import Decimal, InvalidOperation
import requests
API = "https://api.openai.com/v1/responses"
def gate(data: dict) -> tuple[bool, str]:
try:
subtotal = Decimal(str(data["subtotal"]))
tax = Decimal(str(data["tax"]))
total = Decimal(str(data["total"]))
except (KeyError, InvalidOperation):
return False, "字段缺失或金额不是数字"
if data.get("currency") != "CNY": return False, "仅接受CNY票据"
if total > Decimal("2000"): return False, "超过自动处理上限"
if abs(subtotal + tax - total) > Decimal("0.01"):
return False, "小计、税额与总额不一致"
return True, "允许进入人工确认"
def main(path: str) -> None:
key = os.environ.get("OPENAI_API_KEY")
if not key: raise SystemExit("请设置OPENAI_API_KEY")
image = base64.b64encode(open(path, "rb").read()).decode()
prompt = "识别票据，严格只返回JSON：subtotal,tax,total,currency"
body = {"model":"gpt-6", "input":[{"role":"user","content":[
{"type":"input_text","text":prompt},
{"type":"input_image","image_url":"data:image/jpeg;base64,"+image}]}]}
try:
r = requests.post(API, headers={"Authorization":f"Bearer {key}","Content-Type":"application/json"}, json=body, timeout=30)
r.raise_for_status()
text = r.json()["output"][0]["content"][0]["text"]
ok, reason = gate(json.loads(text)); print({"approved": ok, "reason": reason})
except (requests.RequestException, KeyError, json.JSONDecodeError) as e:
print({"approved": False, "reason": f"调用或解析失败: {e}"})
if __name__ == "__main__":
if len(sys.argv) != 2: raise SystemExit("用法: python receipt_gate.py receipt.jpg")
main(sys.argv[1])关键点有三个：不要用浮点数算钱；不要允许模型决定"批准"；不要把原始票据和密钥写进日志。成功时预期输出 {'approved': True, 'reason': '允许进入人工确认'}，仍不是自动付款。
常见失败：一是图片太大或格式不被账户支持，先压缩并核对视觉文档；二是模型返回 Markdown 围栏而非 JSON，应在生产中启用结构化输出或做更严格解析；三是网络超时，代码已设 30 秒，生产环境还应重试并加幂等键；四是发票有多税率，不能沿用单一小计公式。
它适合低风险预录入和人工前置核验，不适合绕过财务审批、税务判断或未获授权的证件处理。下一步可把字段置信度、供应商白名单和审计 ID 写入数据库。5分钟练习：把 2000 改成你的团队阈值，并给 gate 增加"日期不得晚于今天"的规则。
把它接进真实页面
页面至少应该同时展示原图缩略图、模型原始候选、本地复算值和拒绝原因。不要只用一枚绿色"已通过"徽标：使用者需要能看出是"识别到 128.00""规则算出 128.00"，还是"金额相差 0.02"。若模型无法读取图片，保留上传记录并让用户补图；若规则不通过，允许修正字段但要求说明原因。对发票号、身份证号等敏感字段，日志只保留哈希或局部掩码，原图按组织保留期自动删除。批量场景可先排队、限制单用户并发，并把同一文件的 SHA-256 作为幂等键，避免重试造成重复报销草稿。
官方文档：Images and vision。你的业务里，哪一个AI提取字段必须由本地规则二次确认？
关注「蜗牛聊AI」，一起看懂技术变化背后的真正机会。
本文首发于 java4u.cn，转载请注明出处。