Skip to content

Transforms complex documents like PDFs and Office docs into LLM-ready markdown/JSON for your Agentic workflows.

Fix URL encoding for file paths in Gradio API and update comments for clarity

myhloli8de2e84c21d8
diff --git a/mineru/kit/gradio/app.py b/mineru/kit/gradio/app.py
index 5e3eaa2a..ffde428f 100644
--- a/mineru/kit/gradio/app.py
+++ b/mineru/kit/gradio/app.py
@@ -1076,7 +1076,7 @@ def _build_office_preview_html(file_path: str | Path, request: object | None = N
headers = getattr(request, "headers", None) or {}
host = headers.get("x-forwarded-host") or headers.get("host") or "localhost:7860"
protocol = headers.get("x-forwarded-proto") or "http"
- public_url = f"{protocol}://{host}/gradio_api/file={quote(str(source_path), safe='/:')}"
+ public_url = f"{protocol}://{host}/gradio_api/file={quote(source_path.as_posix(), safe='/:')}"
short_name = f"{source_path.stem[-12:]}{source_path.suffix}" if source_path.stem else source_path.name
short_public_url = f"{protocol}://{host}/....{short_name}"
viewer_url = "https://view.officeapps.live.com/op/embed.aspx?src=" + quote(public_url, safe="")
diff --git a/mineru/kit/gradio/artifacts.py b/mineru/kit/gradio/artifacts.py
index 4147fd1b..f2991584 100644
--- a/mineru/kit/gradio/artifacts.py
+++ b/mineru/kit/gradio/artifacts.py
@@ -323,7 +323,7 @@ def _render_html(artifacts: RunArtifacts, *, public_base_url: str) -> str:
if urlsplit(public_base_url).scheme not in {"http", "https"} or not urlsplit(public_base_url).netloc:
raise ValueError("HTML rendering requires the current Gradio HTTP base URL")
result = ParseResult.from_json(artifacts.middle_json_path.read_text(encoding="utf-8"))
- asset_base_url = f"{public_base_url.rstrip('/')}/gradio_api/file={quote(str(artifacts.root), safe='/:')}"
+ asset_base_url = f"{public_base_url.rstrip('/')}/gradio_api/file={quote(artifacts.root.as_posix(), safe='/:')}"
return cast(
str,
render(
diff --git a/mineru/model/_internal/pytorchocr/base_ocr_v20.py b/mineru/model/_internal/pytorchocr/base_ocr_v20.py
index 1b2b672d..7412ef65 100644
--- a/mineru/model/_internal/pytorchocr/base_ocr_v20.py
+++ b/mineru/model/_internal/pytorchocr/base_ocr_v20.py
@@ -7,7 +7,7 @@ import torch
from .modeling.architectures.base_model import BaseModel
-# OCR 推理精度开关:auto 表示 CPU 使用 fp32,非 CPU 自动使用 fp16。
+# OCR rec/cls 推理精度开关:auto 表示 CPU 使用 fp32,非 CPU 自动使用 fp16;det 固定使用 fp32。
OCR_INFERENCE_PRECISION = "auto"
diff --git a/mineru/model/_internal/pytorchocr/infer/predict_det.py b/mineru/model/_internal/pytorchocr/infer/predict_det.py
index 4752ea7d..c035e440 100644
--- a/mineru/model/_internal/pytorchocr/infer/predict_det.py
+++ b/mineru/model/_internal/pytorchocr/infer/predict_det.py
@@ -126,9 +126,10 @@ class TextDetector(BaseOCRV20):
super(TextDetector, self).__init__(network_config, **kwargs)
self.load_pytorch_weights(self.weights_path)
self.net.eval()
+ # Tiny/seal det 的 BatchNorm 统计值超出 fp16 范围,所有设备均固定 fp32;rec/cls 仍使用各自精度策略。
self._apply_inference_precision(
self.device,
- precision_override=getattr(args, "det_inference_precision", None),
+ precision_override="fp32",
)
for module in self.net.modules():
if hasattr(module, 'rep'):
diff --git a/mineru/model/ocr/pytorch_paddle.py b/mineru/model/ocr/pytorch_paddle.py
index 5c7363ce..937764be 100644
--- a/mineru/model/ocr/pytorch_paddle.py
+++ b/mineru/model/ocr/pytorch_paddle.py
@@ -70,8 +70,6 @@ class PytorchPaddleOCR(TextSystem):
kwargs['rec_char_dict_path'] = str(PPOCRV6_DICT_PATH)
kwargs['rec_batch_num'] = 6
if self.is_seal:
- # Seal detector 的 BatchNorm 统计量会在 fp16 下溢出,初始化时固定使用 fp32。
- kwargs['det_inference_precision'] = 'fp32'
kwargs['det_limit_side_len'] = 736
kwargs['det_limit_type'] = 'min'
kwargs['det_max_side_limit'] = 4000
diff --git a/tests/unittest/test_kit_gradio_html.py b/tests/unittest/test_kit_gradio_html.py
index e681e62e..7801821d 100644
--- a/tests/unittest/test_kit_gradio_html.py
+++ b/tests/unittest/test_kit_gradio_html.py
@@ -8,7 +8,8 @@ import io
import json
import re
import zipfile
-from pathlib import Path
+from dataclasses import replace
+from pathlib import Path, PurePosixPath, PureWindowsPath
from types import SimpleNamespace
from urllib.parse import quote, unquote
@@ -21,11 +22,12 @@ from PIL import Image
from starlette.requests import Request
from mineru.backend.analyze import doc_analyze
+from mineru.kit.gradio import artifacts as artifacts_module
from mineru.kit.gradio.app import _download_handler, _gradio_public_base_url, build_gradio_app
from mineru.kit.gradio.artifacts import _prepare_preview_links, persist_parse_result, render_download, render_html_preview
from mineru.kit.gradio.client import V1ServerCapabilities
from mineru.parser.base import ParseResult
-from mineru.types import BlockType, EquationBlock, TableBlock, TableBodyBlock
+from mineru.types import BlockType, ChartBlock, ChartBodyBlock, EquationBlock, TableBlock, TableBodyBlock
from test_kit_gradio import _middle_json
@@ -36,6 +38,77 @@ def _image_bytes(color: str) -> bytes:
return output.getvalue()
[email protected]("entrypoint", ["preview", "download"])
+ "platform_root",
+ [
+ PurePosixPath("/tmp/输出 文件/gradio/授权书"),
+ PureWindowsPath("C:/Users/测试用户/输出 文件/gradio/授权书"),
+ PureWindowsPath("//fileserver/共享目录/输出 文件/gradio/授权书"),
+ ],
+ ids=["posix", "windows-drive", "windows-unc"],
+)
+def test_platform_asset_urls_preserve_seal_images_and_details(
+ tmp_path: Path,
+ monkeypatch: pytest.MonkeyPatch,
+ platform_root: PurePosixPath | PureWindowsPath,
+ entrypoint: str,
+) -> None:
+ """跨平台资源根保留印章、普通图片、图表和表内图片,预览及下载均不会被路径校验过滤。"""
+ middle = _middle_json(file_suffix="docx")
+ seal = middle.pages[0].blocks[1]
+ seal.sub_type = "seal"
+ seal.content[0].content = "上海人工智能创新中心"
+ seal.content[0].image_path = "images/印章 1.png"
+ photo = seal.model_copy(deep=True, update={"index": 2, "sub_type": None})
+ photo.content[0].index = 2
+ middle.pages[0].blocks.extend(
+ [
+ photo,
+ ChartBlock(
+ type="chart",
+ index=3,
+ content=[ChartBodyBlock(type="chart_body", index=3, content="", image_path="images/chart.png")],
+ ),
+ TableBlock(
+ type="table",
+ index=4,
+ content=[
+ TableBodyBlock(
+ type="table_body", index=4, content='<table><tr><td><img src="images/cell.png"></td></tr></table>'
+ )
+ ],
+ ),
+ ]
+ )
+ artifacts = artifacts_module.create_run_artifacts(tmp_path / "授权书.docx", tmp_path)
+ artifacts.middle_json_path.write_text(ParseResult(middle_json=middle).to_json(), encoding="utf-8")
+ original_render_html = artifacts_module._render_html
+
+ def render_with_platform_root(artifacts: artifacts_module.RunArtifacts, *, public_base_url: str) -> str:
+ """文件读写与边界校验仍在本机执行,仅在实际 URL 构造处注入目标平台路径。"""
+ return original_render_html(replace(artifacts, root=platform_root), public_base_url=public_base_url)
+
+ monkeypatch.setattr(artifacts_module, "_render_html", render_with_platform_root)
+ public_base = "https://demo.example.test/mineru"
+ if entrypoint == "preview":
+ frame = BeautifulSoup(render_html_preview(artifacts, public_base_url=public_base), "html.parser").iframe
+ document = frame["srcdoc"]
+ else:
+ path = render_download(artifacts.as_state(), "html", allowed_root=tmp_path, public_base_url=public_base)
+ document = Path(path).read_text(encoding="utf-8")
+ soup = BeautifulSoup(document, "html.parser")
+ images = soup.find_all("img")
+ assert len(images) == 4
+ expected_base = f"{public_base}/gradio_api/file={quote(platform_root.as_posix(), safe='/:')}/images/"
+ assert all(image["src"].startswith(expected_base) for image in images)
+ assert all("%5c" not in image["src"].lower() and " " not in image["src"] for image in images)
+ details = soup.find("summary", string="seal").parent
+ assert details.name == "details" and "open" not in details.attrs
+ assert "上海人工智能创新中心" in details.get_text()
+ assert details.find_previous_sibling("img") is not None
+
+
@pytest.mark.parametrize("image_source", ["inline", "sidecar"])
def test_materialized_html_and_independent_archives(tmp_path: Path, image_source: str) -> None:
"""正文、表格图片和公式保持语义,HTTP 与 ZIP 各自使用正确的图片地址。"""
diff --git a/tests/unittest/test_kit_gradio_office_preview.py b/tests/unittest/test_kit_gradio_office_preview.py
index 4a0743a6..1362dd86 100644
--- a/tests/unittest/test_kit_gradio_office_preview.py
+++ b/tests/unittest/test_kit_gradio_office_preview.py
@@ -5,7 +5,7 @@ from __future__ import annotations
import asyncio
import html
import re
-from pathlib import Path
+from pathlib import Path, PureWindowsPath
from types import SimpleNamespace
from unittest.mock import AsyncMock
from urllib.parse import parse_qs, quote, urlsplit
@@ -61,7 +61,7 @@ def test_office_short_address_keeps_full_encoded_iframe_url(headers: dict[str, s
markup = app_module._build_office_preview_html(source, SimpleNamespace(headers=headers))
iframe_url = html.unescape(re.search(r'<iframe[^>]+src="([^"]+)"', markup)[1])
public_url = parse_qs(urlsplit(iframe_url).query)["src"][0]
- assert public_url == f"{base}/gradio_api/file={quote(str(source), safe='/:')}"
+ assert public_url == f"{base}/gradio_api/file={quote(source.as_posix(), safe='/:')}"
displayed = html.unescape(re.search(r'<div class="office-preview-source-link">(.*?)</div>', markup)[1])
assert f"{base}/....{source.stem[-12:]}{source.suffix}" in displayed
assert "/uploaded/private" not in displayed
@@ -72,6 +72,17 @@ def test_office_short_address_keeps_full_encoded_iframe_url(headers: dict[str, s
assert markup == app_module._build_office_preview_html(source)
+def test_office_windows_file_url_uses_forward_slashes(monkeypatch: pytest.MonkeyPatch) -> None:
+ """Windows 上传预览使用正斜杠 URL,保留盘符、中文及空格的正确编码。"""
+ source = PureWindowsPath("C:/Users/测试用户/上传 文件/报告.docx")
+ monkeypatch.setattr(app_module, "Path", PureWindowsPath)
+ markup = app_module._build_office_preview_html(str(source))
+ iframe_url = html.unescape(re.search(r'<iframe[^>]+src="([^"]+)"', markup)[1])
+ public_url = parse_qs(urlsplit(iframe_url).query)["src"][0]
+ assert public_url == "http://localhost:7860/gradio_api/file=" + quote(source.as_posix(), safe="/:")
+ assert "%5c" not in public_url.lower()
+
+
def test_office_copy_matches_345_release() -> None:
"""固定中英文文案属于本次恢复的用户可见契约。"""
assert MESSAGES["office_preview_title"] == ("Office online preview", "Office 在线预览")
diff --git a/tests/unittest/test_ocr_detector_precision.py b/tests/unittest/test_ocr_detector_precision.py
new file mode 100644
index 00000000..6bbd0d7f
--- /dev/null
+++ b/tests/unittest/test_ocr_detector_precision.py
@@ -0,0 +1,95 @@
+"""验证 det 的 FP32 约束独立于共享 OCR 精度和设备路由。"""
+
+from argparse import Namespace
+from typing import Any, Self
+from unittest.mock import Mock
+
+import numpy as np
+import pytest
+import torch
+
+from mineru.model._internal.pytorchocr import base_ocr_v20
+from mineru.model._internal.pytorchocr.infer import pytorchocr_utility
+from mineru.model._internal.pytorchocr.infer.predict_det import TextDetector
+from mineru.model._internal.pytorchocr.infer.predict_system import TextSystem
+
+
+class _PrecisionNet(torch.nn.Module):
+ """使用真实参数和大数值 BN 缓冲区,仅模拟设备迁移以覆盖本机不存在的硬件。"""
+
+ def __init__(self, config: dict[str, Any], **kwargs: Any) -> None:
+ """构造会在 FP16 转换时溢出的 BN 统计值。"""
+ super().__init__()
+ self.norm = torch.nn.BatchNorm2d(3)
+ self.norm.running_var.fill_(1_000_000)
+ self.requested_device: str | None = None
+ self.input_dtypes: list[torch.dtype] = []
+
+ def to(self, device: str | None = None, *, dtype: torch.dtype | None = None) -> Self:
+ """保留真实 dtype 转换,只记录目标设备而不调用加速卡运行时。"""
+ if device is not None:
+ self.requested_device = str(device)
+ return super().to(dtype=dtype) if dtype is not None else self
+
+ def forward(self, inputs: torch.Tensor) -> dict[str, torch.Tensor]:
+ """记录真实批处理输入精度,并返回可供 DB 后处理检测的概率图。"""
+ self.input_dtypes.append(inputs.dtype)
+ return {"maps": self.norm(inputs).mean(dim=1, keepdim=True).sigmoid() + 0.25}
+
+
+def ocr_args(monkeypatch: pytest.MonkeyPatch) -> Namespace:
+ """替换模型文件加载,保留 det/rec/cls 的真实初始化和精度设置。"""
+ monkeypatch.setattr(base_ocr_v20, "BaseModel", _PrecisionNet)
+ monkeypatch.setattr(pytorchocr_utility, "get_arch_config", Mock(return_value={}))
+ monkeypatch.setattr(base_ocr_v20.BaseOCRV20, "load_pytorch_weights", Mock())
+ monkeypatch.setattr(base_ocr_v20.BaseOCRV20, "read_pytorch_weights", Mock(return_value={}))
+ monkeypatch.setattr(base_ocr_v20.BaseOCRV20, "get_out_channels", Mock(return_value=1))
+ monkeypatch.setattr(base_ocr_v20.BaseOCRV20, "load_state_dict", Mock())
+ args = pytorchocr_utility.init_args().parse_args([])
+ args.use_angle_cls = True
+ args.det_inference_precision = "fp16"
+ return args
+
+
[email protected]("device", ["cpu", "mps", "cuda:0", "npu:0", "xpu:0"])
[email protected]("lang", ["ch", "seal", "seal_lite"])
[email protected]("precision", ["auto", "fp16", "fp32"])
+def test_detector_fp32_keeps_rec_cls_precision_and_device(
+ monkeypatch: pytest.MonkeyPatch, ocr_args: Namespace, device: str, lang: str, precision: str
+) -> None:
+ """所有设备的 det 均保留有限 FP32 权重与输入,且不改变共享参数及 rec/cls 策略。"""
+ monkeypatch.setattr(base_ocr_v20, "OCR_INFERENCE_PRECISION", precision)
+ ocr_args.device, ocr_args.lang = device, lang
+ system = TextSystem(ocr_args)
+ detector = system.text_detector
+ expected_device = "cpu" if device == "mps" and lang == "ch" else device
+ assert detector.device == detector.net.requested_device == expected_device
+ assert ocr_args.device == device
+ assert detector.ocr_inference_dtype == torch.float32
+ for tensor in (*detector.net.parameters(), *detector.net.buffers()):
+ if tensor.is_floating_point():
+ assert tensor.dtype == torch.float32
+ assert torch.isfinite(tensor).all()
+ assert detector._to_inference_dtype(torch.ones(1, dtype=torch.float16)).dtype == torch.float32
+ indices = torch.ones(1, dtype=torch.int64)
+ assert detector._to_inference_dtype(indices) is indices
+ expected_rec_dtype = torch.float32 if device == "cpu" or precision == "fp32" else torch.float16
+ for model in (system.text_recognizer, system.text_classifier):
+ assert model.device == model.net.requested_device == device
+ assert model.ocr_inference_dtype == expected_rec_dtype
+ assert next(model.net.parameters()).dtype == expected_rec_dtype
+ assert model._to_inference_dtype(torch.ones(1)).dtype == expected_rec_dtype
+
+
+def test_detector_single_and_mixed_size_batches_use_fp32(monkeypatch: pytest.MonkeyPatch, ocr_args: Namespace) -> None:
+ """单图及不同形状分桶的真实前后处理均使用 FP32,且批处理框与单图一致。"""
+ monkeypatch.setattr(base_ocr_v20, "OCR_INFERENCE_PRECISION", "fp16")
+ detector = TextDetector(ocr_args)
+ images = [np.full((height, width, 3), 255, dtype=np.uint8) for height, width in [(32, 64), (64, 64), (32, 64)]]
+ singles = [detector(image)[0] for image in images]
+ batches = detector.batch_predict(images, max_batch_size=2)
+ assert all(len(boxes) > 0 for boxes in singles)
+ for single, (batch, _elapsed) in zip(singles, batches):
+ np.testing.assert_array_equal(single, batch)
+ assert detector.net.input_dtypes == [torch.float32] * 5