Files
PolyGateway/src/polygateway/transports/monkey_ocr.py
T

294 lines
12 KiB
Python

"""MonkeyOCR transport: 双端点协议细节(M3 设计 §4;不含任何治理)。
蓝本: CHS `invokers.py:427-552`(两段协议 + 数值防御,逐段保真)与
VT `adapters/ocr.py`(/ocr/text 请求形态)。已在设计声明的偏离: 全元素
提取、`success!=true` 附 status_code=200、429 细分放弃、健康判定收紧 2xx、
`content` 显式校验。服务无鉴权——不带 Authorization 头,api_key 是占位。
"""
from __future__ import annotations
import asyncio
import json
import math
import zipfile
from contextlib import contextmanager
from io import BytesIO
from typing import TYPE_CHECKING, Any
import httpx
from polygateway.errors import (
RequestRejectedError,
ResultInvalidError,
SourceDeadError,
TransientError,
)
from polygateway.types import (
OcrLayoutElement,
OcrLayoutTransportResult,
OcrTextTransportResult,
)
if TYPE_CHECKING:
from collections.abc import Callable, Iterator
from polygateway.types import SourceConfig
_HEALTH_TIMEOUT_S = 5.0 # 探测常量(同 jitter 系数先例,不进配置)
def _default_client_factory(source: SourceConfig) -> httpx.AsyncClient:
return httpx.AsyncClient(
base_url=source.base_url,
timeout=httpx.Timeout(source.timeout_s),
trust_env=source.trust_env, # VT LAN 直连绕代理场景配 false(R9)
)
@contextmanager
def _translate_http_errors(source_name: str, operation: str) -> Iterator[None]:
"""httpx 异常 → 错误四分类(CHS `_translate_http_errors` 语义主体保真)。
429 细分(insufficient_quota/Retry-After)有意放弃——MonkeyOCR 无鉴权
无计费(设计 §10.2);`from exc` 保留 __cause__ 供 RetryMW 归类 timeout。
"""
try:
yield
except httpx.TimeoutException as exc:
raise TransientError(
f"{source_name} OCR {operation} 超时: {exc}",
source_name=source_name,
operation=operation,
) from exc
except httpx.HTTPStatusError as exc:
raise _classify_status(exc, source_name, operation) from exc
except httpx.HTTPError as exc:
raise TransientError(
f"{source_name} OCR {operation} 网络错误: {exc}",
source_name=source_name,
operation=operation,
) from exc
def _classify_status(
exc: httpx.HTTPStatusError, source_name: str, operation: str
) -> TransientError | SourceDeadError | RequestRejectedError:
status = exc.response.status_code
ctx: dict[str, Any] = {
"source_name": source_name,
"status_code": status,
"operation": operation,
}
message = f"{source_name} OCR {operation} HTTP {status}"
if status >= 500 or status == 429:
return TransientError(message, **ctx)
if status in (401, 403):
return SourceDeadError(message, **ctx)
return RequestRejectedError(message, **ctx)
def _finite_number(value: object) -> float:
"""JSON 数值 → 有限浮点;布尔与非有限数拒绝(CHS invokers.py:427 逐字)。"""
if isinstance(value, bool) or not isinstance(value, (int, float)):
raise ValueError("字段不是数值")
number = float(value)
if not math.isfinite(number):
raise ValueError("字段不是有限数值")
return number
def _parse_page_size(page: dict) -> tuple[float, float]:
page_size = page.get("page_size")
if not isinstance(page_size, list) or len(page_size) != 2:
raise ValueError("page_size 无效")
width, height = (_finite_number(item) for item in page_size)
if width <= 0 or height <= 0:
raise ValueError("page_size 必须为正数")
return (width, height)
def _parse_bbox(bbox: object) -> tuple[float, float, float, float]:
if not isinstance(bbox, list) or len(bbox) != 4:
raise ValueError("bbox 无效")
x1, y1, x2, y2 = (_finite_number(item) for item in bbox)
if x2 <= x1 or y2 <= y1:
raise ValueError("bbox 边界顺序无效")
# 库返回 float 原生 bbox;整数化退化校验专为 CHS shim 的 int() 裁剪路径
# 兜底(亚像素宽的框裁剪即空图)——不是死代码,勿删(设计 §4)
if int(x2) <= int(x1) or int(y2) <= int(y1):
raise ValueError("bbox 整数化后退化")
return (x1, y1, x2, y2)
def _parse_block(block: object, page_index: int) -> OcrLayoutElement:
if not isinstance(block, dict):
raise ValueError("块结构无效")
block_type = block.get("type")
if not isinstance(block_type, str) or not block_type.strip():
raise ValueError("块缺少 type")
return OcrLayoutElement(
type=block_type, bbox=_parse_bbox(block.get("bbox")), page_index=page_index
)
def _parse_pages(payload: object) -> tuple[list[OcrLayoutElement], list[tuple[float, float]]]:
if not isinstance(payload, dict) or not isinstance(payload.get("pdf_info"), list):
raise ValueError("_middle.json 缺少 pdf_info 列表")
elements: list[OcrLayoutElement] = []
page_sizes: list[tuple[float, float]] = []
for fallback_index, page in enumerate(payload["pdf_info"]):
if not isinstance(page, dict):
raise ValueError("页面结构无效")
page_sizes.append(_parse_page_size(page))
raw_index = page.get("page_idx", fallback_index)
# bool 是 int 子类、负页号无意义——与 _finite_number 拒 bool 的口径一致
valid_idx = (
isinstance(raw_index, int) and not isinstance(raw_index, bool) and raw_index >= 0
)
page_index = raw_index if valid_idx else fallback_index
blocks = page.get("para_blocks", [])
if not isinstance(blocks, list):
raise ValueError("para_blocks 无效")
elements.extend(_parse_block(block, page_index) for block in blocks)
return elements, page_sizes
def _parse_middle_json(
zip_bytes: bytes,
) -> tuple[list[OcrLayoutElement], list[tuple[float, float]]]:
"""ZIP → (elements, page_sizes);一切形态异常归 ResultInvalid(坏图≠坏服务)。
数值防御全量下沉自 CHS `_parse_table_result`(invokers.py:437-479),
提取面从"首表"泛化为 para_blocks 全元素(设计 §1.2 取证证明无损)。
"""
try:
with zipfile.ZipFile(BytesIO(zip_bytes)) as archive:
middle_files = [n for n in archive.namelist() if n.endswith("_middle.json")]
if not middle_files:
raise ValueError("结果包缺少 _middle.json")
with archive.open(middle_files[0]) as stream:
payload = json.load(stream)
return _parse_pages(payload)
except (ValueError, TypeError, KeyError, json.JSONDecodeError, zipfile.BadZipFile) as exc:
raise ResultInvalidError(f"OCR 结果包内容无效: {exc}") from exc
class MonkeyOcrTransport:
"""MonkeyOCR transport: 每源一个预配 httpx client,懒创建,aclose 统一释放。"""
def __init__(
self,
*,
client_factory: Callable[[SourceConfig], httpx.AsyncClient] | None = None,
) -> None:
self._client_factory = client_factory or _default_client_factory
self._clients: dict[str, httpx.AsyncClient] = {}
def _client_for(self, source: SourceConfig) -> httpx.AsyncClient:
client = self._clients.get(source.name)
if client is None:
client = self._client_factory(source)
self._clients[source.name] = client
return client
@staticmethod
def _files(image: bytes) -> dict[str, tuple[str, bytes, str]]:
# 固定文件名/类型(CHS invokers.py:496 原样;服务按内容处理,35 样本实测)
return {"file": ("image.jpg", image, "image/jpeg")}
def _response_json(self, resp: httpx.Response, source: SourceConfig, operation: str) -> dict:
"""响应体 JSON 化;坏 JSON/非 dict → Transient(设计 §4: 服务侧瞬时异常)。"""
try:
data = resp.json()
except (json.JSONDecodeError, ValueError) as exc:
raise TransientError(
f"{source.name} OCR {operation} 响应非合法 JSON",
source_name=source.name,
operation=operation,
) from exc
if not isinstance(data, dict):
raise TransientError(
f"{source.name} OCR {operation} JSON 结构无效",
source_name=source.name,
operation=operation,
)
return data
async def recognize_text(
self, *, image: bytes, source: SourceConfig, call_id: str
) -> OcrTextTransportResult:
"""POST /ocr/text → 多行纯文本;content 缺失/非 str → ResultInvalid。"""
del call_id # 协议无关,治理侧遥测使用
client = self._client_for(source)
with _translate_http_errors(source.name, "ocr_text"):
resp = await client.post("/ocr/text", files=self._files(image))
resp.raise_for_status()
data = self._response_json(resp, source, "ocr_text")
content = data.get("content")
if not isinstance(content, str):
# VT 的 .get("content","") 静默兜底有意替换为显式校验(P5)
raise ResultInvalidError(
f"{source.name} /ocr/text 响应 content 缺失或非字符串",
source_name=source.name,
operation="ocr_text",
)
return OcrTextTransportResult(
text=content, raw={k: v for k, v in data.items() if k != "content"}
)
async def parse_layout(
self, *, image: bytes, source: SourceConfig, call_id: str
) -> OcrLayoutTransportResult:
"""两段协议: POST /parse → GET download_url → ZIP → 全元素提取。"""
del call_id
client = self._client_for(source)
with _translate_http_errors(source.name, "parse"):
resp = await client.post("/parse", files=self._files(image))
resp.raise_for_status()
data = self._response_json(resp, source, "parse")
if data.get("success") is not True:
# 设计 §10.2 有意修复: 200 响应确证服务活着,附 status_code
# 让治理侧按"响应即健康"记熔断成功(CHS 原样会误走本地拒绝分支)
raise RequestRejectedError(
f"{source.name} OCR 解析失败: {data.get('message')}",
source_name=source.name,
status_code=resp.status_code,
operation="parse",
)
download_url = data.get("download_url")
if not isinstance(download_url, str) or not download_url.strip():
raise TransientError(
f"{source.name} /parse 响应缺少 download_url",
source_name=source.name,
operation="parse",
)
with _translate_http_errors(source.name, "download_result"):
# 实测 download_url 为相对路径(/static/*.zip),相对 base_url 解析;
# httpx 对绝对 URL 会忽略 base_url,两种形态天然兼容
result_resp = await client.get(download_url)
result_resp.raise_for_status()
elements, page_sizes = _parse_middle_json(result_resp.content)
return OcrLayoutTransportResult(
elements=elements,
page_sizes=page_sizes,
raw={"success": True, "download_url": download_url},
)
async def check_health(self, *, source: SourceConfig) -> bool:
"""GET /health 探测: 2xx → True,失败 → False,取消穿透(R10)。"""
client = self._client_for(source)
try:
resp = await client.get("/health", timeout=_HEALTH_TIMEOUT_S)
except asyncio.CancelledError:
raise
except Exception: # noqa: BLE001 — 探测不是调用,失败一律视为不健康
return False
return resp.is_success
async def aclose(self) -> None:
"""幂等释放全部 per-source client。"""
clients, self._clients = self._clients, {}
for client in clients.values():
await client.aclose()