AI Prompt Engineering · 课时

由模式驱动的数据提取

Q: 「由模式驱动的数据提取」这节课中我会学到什么？

在提示中提供 JSON 模式，以确保结构化输出格式 你通过在浏览器中直接运行的动手代码来练习 AI Prompt Engineering，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

在提示中提供 JSON 模式，以确保结构化输出格式

由模式驱动的数据提取 是 CoddyKit 上的免费 AI Prompt Engineering 课时。这是第 2 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 AI Prompt Engineering 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 AI Prompt Engineering 课程共包含 4 节课。

为什么要采用模式驱动的抽取

当您告诉模型提取重要数据时，得到的输出会不一致且不可预测。当您提供 JSON 模式并说提取与此确切模式匹配的数据时，每次都能得到机器可读、一致且类型安全的输出。

模式驱动的抽取是生产系统中采用的模式，适用于处理发票、合同、医疗记录、会议记录，以及任何需要从非结构化文本中可靠抽取结构化数据的文档。

在提示词中提供模式

模式直接写在提示词中。模型将其作为输出契约使用：

import anthropic, json

client = anthropic.Anthropic(api_key='YOUR_API_KEY')

INVOICE_SCHEMA = '''
{
  "invoice_number": "string",
  "vendor_name": "string",
  "vendor_address": "string or null",
  "invoice_date": "YYYY-MM-DD",
  "due_date": "YYYY-MM-DD or null",
  "line_items": [
    {
      "description": "string",
      "quantity": "number",
      "unit_price": "number",
      "total": "number"
    }
  ],
  "subtotal": "number",
  "tax": "number or null",
  "total_amount": "number",
  "currency": "3-letter ISO code e.g. USD"
}
'''

def extract_invoice(invoice_text):
    prompt = f'Extract structured data from this invoice.\nReturn JSON matching this schema exactly:\n{INVOICE_SCHEMA}\n\nInvoice:\n{invoice_text}'
    r = client.messages.create(model='claude-opus-4-5', max_tokens=500, messages=[{'role': 'user', 'content': prompt}])
    return json.loads(r.content[0].text)

print('Invoice schema defined.')

发票抽取示例

将模式应用于从真实发票文本中抽取结构化数据：

invoice_text = '''
INVOICE #INV-2025-0342
From: Acme Software Ltd.
123 Tech Street, San Francisco, CA 94105

Date: March 15, 2025
Due: April 14, 2025

Items:
- Annual Pro License (5 seats) x1 @ $2,400.00 = $2,400.00
- Setup & Onboarding x2 @ $300.00 = $600.00

Subtotal: $3,000.00
Tax (8.5%): $255.00
TOTAL DUE: $3,255.00 USD
'''

result = extract_invoice(invoice_text)
print(f'Invoice: {result["invoice_number"]}')
print(f'Vendor: {result["vendor_name"]}')
print(f'Total: {result["currency"]} {result["total_amount"]}')
print(f'Line items: {len(result["line_items"])}')

会议记录抽取

将模式驱动的抽取应用于会议记录——这是一种结构化程度较低的文档类型：

MEETING_SCHEMA = '''
{
  "meeting_title": "string",
  "date": "YYYY-MM-DD",
  "attendees": ["string"],
  "decisions": ["string"],
  "action_items": [
    {
      "task": "string",
      "owner": "string or null",
      "due_date": "YYYY-MM-DD or null"
    }
  ],
  "next_meeting": "string or null"
}
'''

meeting_notes = '''
Product Sync - March 20, 2025
Attendees: Sarah (PM), Jake (Engineering), Priya (Design)

Decided to push the v2.0 launch to April 15.
Will not include the analytics dashboard in v2.0.

Actions:
- Jake to fix the login bug by March 25
- Priya to finalize mockups by March 22
- Sarah to send updated roadmap to stakeholders (no date set)

Next sync: March 27, same time.
'''

print(f'Meeting schema: {len(MEETING_SCHEMA)} chars')
print(f'Notes length: {len(meeting_notes)} chars')

产品规格抽取

从目录描述中抽取结构化的产品规格：

PRODUCT_SCHEMA = '''
{
  "product_name": "string",
  "sku": "string or null",
  "category": "string",
  "price": {"amount": "number", "currency": "string"},
  "dimensions": {
    "length_cm": "number or null",
    "width_cm": "number or null",
    "height_cm": "number or null",
    "weight_kg": "number or null"
  },
  "colors": ["string"],
  "materials": ["string"],
  "features": ["string"],
  "in_stock": true | false
}
'''

product_text = 'AlphaDesk Pro standing desk. SKU: AD-PRO-001. $899. Available in white and black. 120x60x75cm, 35kg. Steel frame, bamboo top. Features: memory height, anti-collision, app control. In stock.'

prompt = f'Extract product specs. Return JSON:\n{PRODUCT_SCHEMA}\n\nProduct: {product_text}'
r = client.messages.create(model='claude-opus-4-5', max_tokens=400, messages=[{'role': 'user', 'content': prompt}])
print(json.loads(r.content[0].text))

处理可选字段

模式必须妥善处理可选字段。对于缺失数据，请使用 null 作为默认值，而不是省略字段——这样可以保持输出结构一致：

prompt_optional = '''
Extract the data. For fields not present in the source text,
use null — do NOT omit the field.
Every field in the schema must appear in the output.

Schema:
{
  "company": "string",
  "ceo": "string or null",
  "founded": "YYYY or null",
  "revenue": "string or null",
  "employees": "number or null"
}

Text: Vertex AI Solutions is a B2B SaaS company.
'''

# Expected output: ceo, founded, revenue, employees all set to null
# NOT omitted — null fields are still present in the JSON
print(prompt_optional)

使用同一模式进行多文档抽取

同一模式可以一致地应用于许多文档。这就是如何大规模地从非结构化文档构建结构化数据库：

def extract_many(documents, schema):
    results = []
    for i, doc in enumerate(documents):
        try:
            r = client.messages.create(
                model='claude-opus-4-5', max_tokens=400,
                messages=[{'role': 'user', 'content': f'Extract data. Return JSON matching schema:\n{schema}\n\nDocument:\n{doc}'}]
            )
            parsed = json.loads(r.content[0].text)
            parsed['_source_doc'] = i
            parsed['_extraction_ok'] = True
            results.append(parsed)
        except (json.JSONDecodeError, Exception) as e:
            results.append({'_source_doc': i, '_extraction_ok': False, '_error': str(e)})
    return results

invoices = ['Invoice from Acme, March 2025, $500', 'Invoice from Beta Corp, April 2025, $1200']
results = extract_many(invoices, INVOICE_SCHEMA)
print(f'Processed: {len([r for r in results if r["_extraction_ok"]])} success, {len([r for r in results if not r["_extraction_ok"]])} failed')

抽取后的模式验证

使用 Python 的 jsonschema 库或自定义验证器，根据预期模式验证抽取的数据：

def validate_extracted(data, required_fields, type_checks):
    errors = []

    # Check required fields
    for field in required_fields:
        if field not in data or data[field] is None:
            errors.append(f'Required field missing or null: {field}')

    # Check types
    for field, expected_type in type_checks.items():
        if field in data and data[field] is not None:
            if not isinstance(data[field], expected_type):
                errors.append(f'{field}: expected {expected_type.__name__}, got {type(data[field]).__name__}')

    return errors

extracted = {'invoice_number': 'INV-001', 'total_amount': 3255.0, 'vendor_name': 'Acme', 'invoice_date': '2025-03-15'}
required = ['invoice_number', 'total_amount', 'vendor_name']
types = {'total_amount': float, 'invoice_number': str, 'line_items': list}
errors = validate_extracted(extracted, required, types)
print('Validation errors:', errors)

迭代完善模式

模式通过反复测试不断演进。流程如下：

根据领域知识定义初始模式
在 20 份样本文档上运行抽取
检查输出——哪些字段经常出错或缺失？
完善模式描述并添加字段定义
在同样的 20 份文档上重新运行
重复此过程，直到质量达到阈值

向模式添加字段描述

当字段含义不明确时，请添加描述性注释来引导模型：

ANNOTATED_SCHEMA = '''
{
  "invoice_number": "string // The unique identifier for this invoice, e.g., INV-2025-001",
  "invoice_date": "YYYY-MM-DD // Date the invoice was issued",
  "due_date": "YYYY-MM-DD or null // Payment due date; null if not specified",
  "subtotal": "number // Amount before tax, as a decimal number",
  "tax": "number or null // Tax amount as a decimal; null if tax is not listed",
  "total_amount": "number // Final amount to pay, including tax",
  "payment_terms": "string or null // e.g., Net 30, Due on receipt; null if not mentioned"
}
'''

print('Annotated schema adds context per field.')
print(f'Schema length: {len(ANNOTATED_SCHEMA)} chars')

抽取字段的置信度分数

对于生产系统，请为每个字段包含置信度分数。低置信度的抽取结果可以转交人工审核：

SCHEMA_WITH_CONFIDENCE = '''
{
  "fields": {
    "invoice_number": {"value": "string", "confidence": "high|medium|low"},
    "total_amount": {"value": "number", "confidence": "high|medium|low"},
    "due_date": {"value": "YYYY-MM-DD or null", "confidence": "high|medium|low"}
  },
  "overall_confidence": "high|medium|low",
  "extraction_notes": "string or null // Any ambiguities encountered"
}
'''

prompt = f'Extract invoice data with confidence scores.\nReturn JSON:\n{SCHEMA_WITH_CONFIDENCE}\n\nInvoice: Payment due within 30 days. Total is approximately $500.'
r = client.messages.create(model='claude-opus-4-5', max_tokens=300, messages=[{'role': 'user', 'content': prompt}])
result = json.loads(r.content[0].text)
print('Overall confidence:', result.get('overall_confidence'))
print('Notes:', result.get('extraction_notes'))

快速检查

当源文档中不存在必填字段时，模式驱动的抽取提示词应该指示模型为该字段返回什么？

模式驱动的抽取——要点

模式驱动的抽取是生产环境中可靠处理文档的标准：

在提示词中提供完整准确的 JSON 模式——模型将其作为输出契约使用
为含义不明确的字段添加字段描述，引导模型进行解释
始终指示：缺失字段返回空值，绝不省略字段
在许多文档中应用同一模式，以获得一致且可直接用于数据库的输出
为每个字段加入置信度分数，以便将结果转交人工审核
每次抽取后都以编程方式验证抽取的数据
迭代完善模式：抽取 20 份样本、检查、改进、重复

常见问题解答

「由模式驱动的数据提取」课时是免费的吗？

是的 — 「由模式驱动的数据提取」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 AI Prompt Engineering 课程的其余内容，请升级到 CoddyKit PRO。 AI Prompt Engineering 课程共包含 4 节课。

「由模式驱动的数据提取」这节课中我会学到什么？

在提示中提供 JSON 模式，以确保结构化输出格式你通过在浏览器中直接运行的动手代码来练习 AI Prompt Engineering，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 AI Prompt Engineering 需要有经验吗？

无需任何先前经验。CoddyKit 上的 AI Prompt Engineering 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 2 节课，共 4 节。

「由模式驱动的数据提取」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 AI Prompt Engineering 课中编写并运行代码吗？

能。每节 AI Prompt Engineering 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 AI Prompt Engineering