# 布局分析输出示例
{
    "blocks": [
        {
            "type": "table",
            "bbox": [120, 340, 780, 1200],
            "children": [...]
        },
        {
            "type": "text",
            "bbox": [120, 100, 600, 280],
            "content": "财务报表 - 2026年Q1"
        }
    ]
}

# 手写识别配置
chandra.config.hw_recognition = {
    "enabled": True,
    "confidence_threshold": 0.7,
    "supported_scripts": ["latin", "digit"]
}

from chandra import OCR

# 初始化
ocr = OCR(model="chandra-table-v2")

# 识别财务报表
result = ocr.process("financial_report.jpg")

# 输出表格
for table in result.tables:
    print(f"表格 {table.id}")
    print(table.to_csv())  # 输出 CSV 格式

# 输入：复杂合并单元格的财务表格
┌──────────────────────────────────────┐
│         │    2025    │    2026      │
│  项目   ├──────┬─────┼──────┬────── │
│         │  Q1  │ Q2  │  Q1  │  Q2   │
├─────────┼──────┼─────┼──────┼────── │
│ 收入    │ 100  │ 120 │ 130  │ 150   │
├─────────┼──────┼─────┼──────┼────── │
│ 成本    │  60  │  70 │  75  │  80   │
└─────────┴──────┴─────┴──────┴────── ┘

# 输出：结构化 CSV
项目,2025_Q1,2025_Q2,2026_Q1,2026_Q2
收入,100,120,130,150
成本,60,70,75,80

# 表单识别示例
form = ocr.process_form("medical_form.jpg")

# 输出表单字段
for field in form.fields:
    print(f"{field.label}: {field.value} (置信度: {field.confidence})")

# 输出示例
"""
姓名: 张三 (置信度: 0.98)
日期: 2026-03-29 (置信度: 0.95)
症状: 头痛、发热 (置信度: 0.87)
签名: [手写签名图片]
"""

# 手写识别配置
config = {
    "handwriting": {
        "enabled": True,
        "min_word_length": 2,
        "language": "zh_cn",  # 支持中文手写
        "confidence_threshold": 0.6
    }
}

result = ocr.process("handwritten_notes.jpg", **config)
print(result.to_text())  # 输出识别文本

from chandra import BatchProcessor

# 批量处理 PDF 或图片目录
processor = BatchProcessor(
    output_format="json",
    parallel_workers=4
)

results = processor.process(
    input_path="documents/",
    output_path="results/"
)

# 批量输出
for doc in results:
    print(f"{doc.filename}: {doc.page_count} 页")

# 方式一：pip 安装
pip install chandra-ocr

# 方式二：从源码安装
git clone https://github.com/datalab-to/chandra
cd chandra
pip install -e .

# 下载预训练模型
python -m chandra download-models --all

# 最简使用
from chandra import OCR

ocr = OCR()  # 自动下载默认模型
result = ocr.process("document.jpg")

# 输出纯文本
print(result.text)

# 输出结构化 JSON
print(result.to_json())

# 输出带坐标的文本
print(result.to_text_with_coords())

from chandra import OCR

ocr = OCR(
    model="chandra-v2",      # 模型名称
    device="cuda",            # "cuda" 或 "cpu"
    precision="fp16",         # "fp32", "fp16", "int8"
    table_strategy="auto"      # "auto", "grid", "line"
)

result = ocr.process("document.jpg")

# 访问结果
result.text          # 纯文本
result.json          # 结构化 JSON
result.tables         # 表格列表
result.forms         # 表单列表
result.metadata      # 元信息（置信度、耗时等）

# 详细配置示例
config = {
    # 布局分析
    "layout": {
        "detect_tables": True,
        "detect_forms": True,
        "detect_signatures": True,
        "reading_order": True
    },
    
    # 表格识别
    "table": {
        "strategy": "hybrid",  # "grid", "line", "hybrid"
        "min_rows": 2,
        "min_cols": 2,
        "merge_cells": True
    },
    
    # 手写识别
    "handwriting": {
        "enabled": True,
        "confidence_threshold": 0.6,
        "language": "auto"
    },
    
    # 输出格式
    "output": {
        "include_coords": True,
        "include_confidence": True,
        "include_structure": True
    }
}

result = ocr.process("document.jpg", **config)

# 处理发票
invoice = ocr.process("invoice.jpg")

# 提取关键字段
print(f"发票号码: {invoice.fields['invoice_no']}")
print(f"金额: {invoice.fields['amount']}")
print(f"日期: {invoice.fields['date']}")

# 验证表格完整性
table = invoice.tables[0]
assert table.validate_schema(required_columns=["项目", "金额", "税率"])

# 处理病历表单
medical_record = ocr.process_form("medical_form.jpg")

# 提取结构化信息
patient_info = {
    "name": medical_record["patient_name"],
    "id": medical_record["patient_id"],
    "diagnosis": medical_record["diagnosis"],
    "prescription": medical_record["prescription"]
}

# 识别手写处方
prescription_handwritten = medical_record.handwritten_fields

# 处理历史文档（可能有褪色、损坏）
config = {
    "preprocessing": {
        "denoise": True,
        "enhance_contrast": True,
        "deskew": True
    },
    "layout": {
        "detect_stamps": True,  # 检测印章
        "handle_rotated_pages": True
    }
}

historical_doc = ocr.process("historical_1920.jpg", **config)

from chandra import EnterpriseProcessor

processor = EnterpriseProcessor(
    ocr_engine="chandra-v2",
    output_adapter="elasticsearch"  # 直接写入 ES
)

# 处理整个文件夹的文档
processor.process_directory(
    input_dir="contracts/",
    index_name="contracts_2026",
    field_mapping={
        "party_a": "甲方",
        "party_b": "乙方",
        "amount": "合同金额",
        "date": "签订日期"
    }
)

# 使用 GPU 进行加速
ocr = OCR(
    device="cuda",
    precision="fp16"  # 半精度加速
)

# Benchmark
import time
start = time.time()
result = ocr.process("page.jpg")
print(f"GPU 耗时: {time.time() - start:.3f}s")  # ~0.15s/page

# CPU 优化配置
ocr = OCR(
    device="cpu",
    num_threads=8,
    use_quantized_models=True  # INT8 量化模型
)

# Benchmark
start = time.time()
result = ocr.process("page.jpg")
print(f"CPU 耗时: {time.time() - start:.3f}s")  # ~1.2s/page

from chandra import BatchProcessor

# 使用多进程批处理
processor = BatchProcessor(
    num_workers=4,        # 并行 worker 数
    prefetch_factor=2,    # 预取因子
    chunk_size=10         # 分块大小
)

# 处理 1000 页文档
results = processor.process(
    input_path="large_document.pdf",
    pages=list(range(0, 1000, 10))  # 每 10 页一批
)

ocr = OCR(model="chandra-zh-v2")

FROM datalab/chandra:latest

# 暴露 API 端口
EXPOSE 8080

# 运行 API 服务
CMD ["python", "-m", "chandra.api", "--port", "8080"]

# 构建和运行
docker build -t my-chandra .
docker run -p 8080:8080 --gpus all my-chandra

from chandra import LargeDocumentProcessor

processor = LargeDocumentProcessor(
    ocr=ocr,
    checkpoint_interval=100  # 每 100 页保存一次检查点
)

# 自动分批处理，支持断点续传
result = processor.process("huge_document.pdf")

# 下载离线模型
python -m chandra download-models --offline

# 初始化离线 OCR
ocr = OCR(offline=True)