Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(mineru): honor dataset language in figure prompts (#18188)

Elias Shieh9784814a9b3e
diff --git a/deepdoc/parser/mineru_parser.py b/deepdoc/parser/mineru_parser.py
index 18921f1d96..cd69cc1770 100644
--- a/deepdoc/parser/mineru_parser.py
+++ b/deepdoc/parser/mineru_parser.py
@@ -876,7 +876,7 @@ class MinerUParser(RAGFlowPdfParser):
def _transfer_to_tables(self, outputs: list[dict[str, Any]]):
return []
- def _enhance_images_with_vlm(self, outputs: list[dict[str, Any]], vision_model, callback: Optional[Callable] = None):
+ def _enhance_images_with_vlm(self, outputs: list[dict[str, Any]], vision_model, callback: Optional[Callable] = None, language: str = "English"):
"""Generate semantic descriptions for image blocks via the tenant's
VISION model, mirroring deepdoc's VisionFigureParser. Each
IMAGE block with a readable img_path gets a ``vlm_description``
@@ -894,7 +894,7 @@ class MinerUParser(RAGFlowPdfParser):
if callback:
callback(0.78, f"[MinerU] Generating VLM descriptions for {len(image_jobs)} images...")
- prompt = vision_llm_figure_describe_prompt()
+ prompt = vision_llm_figure_describe_prompt(language=language or "English")
def worker(idx, item):
try:
@@ -935,7 +935,7 @@ class MinerUParser(RAGFlowPdfParser):
created_tmp_dir = False
parser_cfg = kwargs.get("parser_config", {})
- lang = parser_cfg.get("mineru_lang") or kwargs.get("lang", "English")
+ lang = parser_cfg.get("mineru_lang") or kwargs.get("lang") or "English"
mineru_lang_code = LANGUAGE_TO_MINERU_MAP.get(lang, "ch") # Defaults to Chinese if not matched
mineru_method_raw_str = parser_cfg.get("mineru_parse_method", "auto")
enable_formula = parser_cfg.get("mineru_formula_enable", True)
@@ -998,7 +998,7 @@ class MinerUParser(RAGFlowPdfParser):
vision_model = kwargs.get("vision_model")
if vision_model is not None:
try:
- self._enhance_images_with_vlm(outputs, vision_model, callback=callback)
+ self._enhance_images_with_vlm(outputs, vision_model, callback=callback, language=lang)
except Exception as e:
self.logger.warning(f"[MinerU] VLM image enhancement failed: {e}. Continuing without descriptions.")
diff --git a/test/unit_test/deepdoc/parser/test_mineru_parser.py b/test/unit_test/deepdoc/parser/test_mineru_parser.py
index f76add7f71..bf7090c4cf 100644
--- a/test/unit_test/deepdoc/parser/test_mineru_parser.py
+++ b/test/unit_test/deepdoc/parser/test_mineru_parser.py
@@ -4,8 +4,11 @@ import sys
from io import BytesIO
from pathlib import Path
from types import ModuleType
+from unittest.mock import Mock
import json
+import pytest
+
def _load_mineru_parser(monkeypatch):
repo_root = Path(__file__).resolve().parents[4]
@@ -39,6 +42,73 @@ def _load_mineru_parser(monkeypatch):
return module
+ ("language", "expected_language"),
+ [
+ ("Japanese", "Japanese"),
+ ("", "English"),
+ (None, "English"),
+ ],
+)
+def test_enhance_images_with_vlm_passes_dataset_language_to_prompt(monkeypatch, tmp_path, language, expected_language):
+ module = _load_mineru_parser(monkeypatch)
+ parser = module.MinerUParser()
+ image_path = tmp_path / "figure.png"
+ module.Image.new("RGB", (1, 1)).save(image_path)
+
+ picture_module = ModuleType("rag.app.picture")
+ picture_module.vision_llm_chunk = Mock(return_value="description")
+ prompt = Mock(return_value="prompt")
+ generator_module = ModuleType("rag.prompts.generator")
+ generator_module.vision_llm_figure_describe_prompt = prompt
+ monkeypatch.setitem(sys.modules, "rag.app.picture", picture_module)
+ monkeypatch.setitem(sys.modules, "rag.prompts.generator", generator_module)
+
+ outputs = [{"type": module.MinerUContentType.IMAGE, "img_path": str(image_path)}]
+ parser._enhance_images_with_vlm(outputs, vision_model=object(), language=language)
+
+ prompt.assert_called_once_with(language=expected_language)
+ assert outputs[0]["vlm_description"] == "description"
+
+
+ ("language", "expected_language"),
+ [
+ ("Japanese", "Japanese"),
+ ("", "English"),
+ (None, "English"),
+ ],
+)
+def test_parse_pdf_forwards_normalized_dataset_language_to_image_enhancement(monkeypatch, tmp_path, language, expected_language):
+ module = _load_mineru_parser(monkeypatch)
+ parser = module.MinerUParser()
+ pdf_path = tmp_path / "document.pdf"
+ pdf_path.write_bytes(b"%PDF-1.4 fake")
+ output_dir = tmp_path / "output"
+ vision_model = object()
+
+ monkeypatch.setattr(module, "extract_pdf_outlines", Mock(return_value=[]))
+ monkeypatch.setattr(parser, "__images__", Mock())
+ monkeypatch.setattr(parser, "_run_mineru", Mock(return_value=output_dir))
+ monkeypatch.setattr(parser, "_read_output", Mock(return_value=[]))
+ enhance = Mock()
+ monkeypatch.setattr(parser, "_enhance_images_with_vlm", enhance)
+
+ language_kwargs = {} if language is None else {"lang": language}
+ parser.parse_pdf(
+ filepath=pdf_path,
+ binary=None,
+ output_dir=str(output_dir),
+ delete_output=False,
+ vision_model=vision_model,
+ **language_kwargs,
+ )
+
+ enhance.assert_called_once_with([], vision_model, callback=None, language=expected_language)
+
+
def test_sanitize_section_text_removes_escaped_html_tags(monkeypatch):
module = _load_mineru_parser(monkeypatch)
text = "<table><tr><td>Alpha</td><td>Beta</td></tr></table>"
diff --git a/test/unit_test/rag/app/test_vision_language_callers.py b/test/unit_test/rag/app/test_vision_language_callers.py
index b39f498684..87e6a2627d 100644
--- a/test/unit_test/rag/app/test_vision_language_callers.py
+++ b/test/unit_test/rag/app/test_vision_language_callers.py
@@ -125,3 +125,30 @@ def test_mistral_ocr_forwards_language_to_parser(monkeypatch):
assert tables == []
assert returned_parser is parser
assert parser.parse_pdf.call_args.kwargs["lang"] == "Japanese"
+
+
+def test_mineru_forwards_dataset_language_to_parser(monkeypatch):
+ parser = Mock()
+ parser.parse_pdf.return_value = (["section"], [])
+ ocr_model = Mock(mdl=parser)
+ monkeypatch.setattr(naive, "resolve_model_config", Mock(return_value={"llm_name": "mineru"}))
+ monkeypatch.setattr(naive, "LLMBundle", Mock(return_value=ocr_model))
+
+ sections, tables, returned_parser = naive.by_mineru(
+ "document.pdf",
+ binary=b"pdf",
+ from_page=2,
+ to_page=5,
+ lang="Japanese",
+ callback=lambda *_args, **_kwargs: None,
+ parse_method="raw",
+ mineru_llm_name="mineru",
+ tenant_id="tenant-id",
+ vision_model=object(),
+ )
+
+ assert sections == ["section"]
+ assert tables == []
+ assert returned_parser is parser
+ assert parser.parse_pdf.call_args.kwargs["lang"] == "Japanese"