Pythonによる請求書解析：REST API活用ガイド

import os
import base64
import json
import requests

# Get your API key from an environment variable
API_KEY = os.environ.get("SPACE_OCR_API_KEY")
IMAGE_PATH = "./path/to/your/invoice.jpg" # Path to your invoice image
API_URL = "https://api.space-ocr.com/ocr/fields"

if not API_KEY:
    raise ValueError("API key not found. Set the SPACE_OCR_API_KEY environment variable.")

# Declare the schema you want back.
# 'type' adds a deterministic parse under data.normalized; 'required' and 'pattern'
# are checked after extraction and show up in the review list when they fail.
# Declarations never reach the model: they do not change the value, only the signals.
fields_schema = [
    {"name": "supplier", "type": "string", "description": "The name of the company that issued the invoice."},
    {"name": "invoice_number", "type": "string", "required": True, "pattern": "^[0-9]{8}-[0-9]{3}$"},
    {"name": "issue_date", "type": "date", "required": True},
    {"name": "total_due", "type": "number", "required": True},
    {
        "name": "line_items",
        "type": "array",
        "description": "All items listed in the invoice table.",
        "children": [
            {"name": "item_description", "type": "string"},
            {"name": "unit_price", "type": "number"},
            {"name": "quantity", "type": "integer"},
            {"name": "line_total", "type": "number"}
        ]
    }
]

# Read the image file and encode it in base64
with open(IMAGE_PATH, "rb") as image_file:
    base64_image = base64.b64encode(image_file.read()).decode("utf-8")

headers = {
    "Authorization": f"Bearer {API_KEY}",
    "Content-Type": "application/json"
}

payload = {
    "image": base64_image,
    "imageType": "base64",
    "fields": fields_schema
}

print(f"Sending request for {IMAGE_PATH}...")
response = requests.post(API_URL, headers=headers, json=payload)

if response.status_code != 200:
    # Errors use one envelope: {"error": {"code", "message", "requestId"}, "details": {...}}
    try:
        err = response.json()["error"]
        print(f"HTTP {response.status_code} {err['code']}: {err['message']} (requestId {err['requestId']})")
    except (ValueError, KeyError):
        # A body over 32MiB is cut off before the app sees it and comes back as HTML
        print(f"HTTP {response.status_code}: {response.text[:200]}")
    if response.status_code == 429:
        print(f"Rate limited - Retry-After: {response.headers.get('Retry-After')} seconds")
    # 400 invalid_image is not charged and a retry will not help; 502 is refunded and worth
    # retrying; 504 means the 180s sync ceiling was hit, so send that page on its own.
    raise SystemExit(1)

data = response.json()["data"]
values = data["values"]                  # your schema, and nothing else
cells = data["cells"]                    # flat map: path -> box / quad / verified / review / evidence
review = data["review"]                  # what is worth a second look
normalized = data.get("normalized", {})  # deterministic parse of the types you declared
image = data["image"]                    # {"width", "height"} - the frame the coordinates use

print("\n--- Extracted Invoice Data ---")
print(json.dumps(values, ensure_ascii=False, indent=2))

# 'values' keeps the text as it was read; the parsed form lives here.
# A null leaf means the parse failed - the reason sits in cells[path]["normalized"]["error"].
print("\nissue_date parsed:", normalized.get("issue_date"))
print("total_due parsed :", normalized.get("total_due"))

# The review queue: every entry carries a path and a rank-ordered reasons array.
print(f"\n{len(review['flagged'])} of {review['declared']} declared fields were flagged:")
for item in review["flagged"]:
    path = item["path"]
    reasons = item["reasons"]      # reasons[0] is the primary one
    cell = cells.get(path)         # 'missing' and 'nobox' have no cell to point at
    box = cell["box"] if cell else None
    print(f"  {path}: {reasons[0]} ({', '.join(reasons)}) box={box}")

# Where a single value came from
total = cells.get("total_due")
if total:
    print("\ntotal_due")
    print("  box     :", total["box"])       # {"xmin", "ymin", "xmax", "ymax"}, 0-1000
    print("  quad    :", total["quad"])      # four {"x", "y"} points, following the page tilt
    print("  verified:", total["verified"])  # False if flagged, True if a check ran clean, None if there was nothing to check
    print("  evidence:", total["evidence"])  # text_match, source, match_ratio, printed_text ...
    left = total["box"]["xmin"] / 1000 * image["width"]
    top = total["box"]["ymin"] / 1000 * image["height"]
    print(f"  pixels  : ({left:.0f}, {top:.0f}) on a {image['width']}x{image['height']} page")

# Line items: values holds the rows, cells is keyed by the indexed path
print("\nline_items")
for i, row in enumerate(values.get("line_items") or []):
    cell = cells.get(f"line_items[{i}].line_total")
    verified = cell["verified"] if cell else None
    print(f"  row {i}: {row.get('item_description')} x{row.get('quantity')} "
          f"= {row.get('line_total')} (verified={verified})")