本地优先推理：让80%的文档零API成本完成提取

import re
import json
from dataclasses import dataclass, field
from typing import Optional

# ── 置信度配置 ──
LOCAL_THRESHOLD = 0.85      # ≥ 此值直接输出
API_THRESHOLD = 0.50        # ≥ 此值送云端，< 此值送人工
LOW_CONFIDENCE_PENALTY = 0.3  # 云端返回仍低置信度时的惩罚

# ── 模板定义：工程图纸标题栏 ──
DRAWING_TEMPLATE = {
    "drawing_number": {
        "regex": r"DWG[-\s]?(\d{4,12})",
        "position": {"row": 3, "col": 2},
        "expected_len": (4, 12),
    },
    "revision": {
        "regex": r"REV[-\s]?([A-Z])",
        "position": {"row": 3, "col": 4},
        "expected_len": (1, 1),
    },
    "date": {
        "regex": r"(\d{4}[-/]\d{2}[-/]\d{2})",
        "position": {"row": 4, "col": 2},
        "expected_range": ("2000-01-01", "2030-12-31"),
    },
    "title": {
        "regex": r"TITLE[:\s]+(.+?)(?:\n|$)",
        "position": {"row": 1, "col": 1},
        "expected_len": (5, 200),
    },
}


@dataclass
class ExtractedField:
    name: str
    value: Optional[str] = None
    confidence: float = 0.0
    source: str = "local"  # local | api | human


@dataclass
class DocumentResult:
    doc_id: str
    fields: list[ExtractedField] = field(default_factory=list)
    route: str = "local"  # local | api | human


def compute_field_confidence(field_name: str, value: str, text: str) -> float:
    """组合多个信号计算单字段置信度"""
    template = DRAWING_TEMPLATE.get(field_name, {})
    conf = 0.0
    signals = 0

    # 信号1：正则是否匹配
    regex = template.get("regex")
    if regex and re.search(regex, text):
        match = re.search(regex, text)
        if match.group(1).strip() == value.strip():
            conf += 0.40
            signals += 1

    # 信号2：值长度是否在预期范围
    expected_len = template.get("expected_len")
    if expected_len and expected_len[0] <= len(value) <= expected_len[1]:
        conf += 0.20
        signals += 1

    # 信号3：值是否在预期值域（日期/数值类）
    expected_range = template.get("expected_range")
    if expected_range and value >= expected_range[0] and value <= expected_range[1]:
        conf += 0.25
        signals += 1

    # 信号4：位置是否命中（简化：检查值是否出现在文本对应行）
    position = template.get("position")
    if position:
        lines = text.split("\n")
        row_idx = position["row"] - 1
        if row_idx < len(lines) and value in lines[row_idx]:
            conf += 0.15
            signals += 1

    # 无信号则给低底分
    return conf if signals > 0 else 0.10


def local_extract(doc_id: str, text: str) -> DocumentResult:
    """第一层：确定性本地提取"""
    result = DocumentResult(doc_id=doc_id)
    for field_name, template in DRAWING_TEMPLATE.items():
        regex = template.get("regex", r"(.+)")
        match = re.search(regex, text)
        if match:
            value = match.group(1).strip()
            confidence = compute_field_confidence(field_name, value, text)
            result.fields.append(ExtractedField(
                name=field_name, value=value,
                confidence=confidence, source="local"
            ))
        else:
            result.fields.append(ExtractedField(
                name=field_name, confidence=0.0, source="local"
            ))
    return result


def route_document(result: DocumentResult) -> DocumentResult:
    """根据置信度决定路由"""
    avg_conf = sum(f.confidence for f in result.fields) / len(result.fields) if result.fields else 0.0

    if avg_conf >= LOCAL_THRESHOLD:
        result.route = "local"
    elif avg_conf >= API_THRESHOLD:
        result.route = "api"
    else:
        result.route = "human"
    return result


# ── 模拟云端提取（实际中调用 Azure OpenAI）──
def api_extract(result: DocumentResult, text: str) -> DocumentResult:
    """第二层：云端大模型提取，仅处理低置信度字段"""
    # 实际实现中，这里构造 prompt 调用 Azure OpenAI
    # 示例仅模拟：对低置信度字段重新赋值
    prompt = f"""Extract the following fields from this engineering drawing text.
Return JSON with keys: {', '.join(f.name for f in result.fields if f.confidence < LOCAL_THRESHOLD)}.

Text:
{text[:2000]}
"""
    # 模拟API返回（实际代码用 openai.ChatCompletion.create）
    mock_api_response = {
        "drawing_number": "DWG-20240101",
        "revision": "C",
        "date": "2024-06-15",
        "title": "Pump Assembly Detail"
    }

    for field in result.fields:
        if field.confidence < LOCAL_THRESHOLD:
            api_value = mock_api_response.get(field.name)
            if api_value:
                field.value = api_value
                field.confidence = 0.75  # API 返回给中等置信度
                field.source = "api"

    # API 返回后仍有低置信度字段 → 标记人工
    if any(f.confidence < API_THRESHOLD for f in result.fields):
        result.route = "human"
    else:
        result.route = "api"
    return result


# ── 运行示例 ──
sample_text = """TITLE: Hydraulic Pump Assembly - Section A
DWG-20240101
Material: Stainless Steel 316
REV-C
Date: 2024-06-15
Scale: 1:2
Approved by: J. Smith
"""

# Step 1: 本地提取
result = local_extract("ENG-001", sample_text)
print("=== 本地提取结果 ===")
for f in result.fields:
    print(f"  {f.name}: value={f.value}, conf={f.confidence:.2f}")

# Step 2: 路由决策
result = route_document(result)
print(f"\n=== 路由决策: {result.route} (avg_conf={sum(f.confidence for f in result.fields)/len(result.fields):.2f}) ===")

# Step 3: 如果路由到API层
if result.route == "api":
    result = api_extract(result, sample_text)
    print("\n=== API提取后 ===")
    for f in result.fields:
        print(f"  {f.name}: value={f.value}, conf={f.confidence:.2f}, source={f.source}")
    print(f"  最终路由: {result.route}")