Skip to content

armgit / dify

Public Imported from GitHub
0

Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack.

fix(rag): identify DOCX images by relationship type (#44905)

Winoren04aee9d90658
diff --git a/api/core/rag/extractor/word_extractor.py b/api/core/rag/extractor/word_extractor.py
index b9f2547fff1..c15eacf5a9e 100644
--- a/api/core/rag/extractor/word_extractor.py
+++ b/api/core/rag/extractor/word_extractor.py
@@ -14,6 +14,7 @@ from typing import override
from urllib.parse import urlparse
from docx import Document as DocxDocument
+from docx.opc.constants import RELATIONSHIP_TYPE as RT
from docx.oxml.ns import qn
from docx.table import Table
from docx.text.paragraph import Paragraph
@@ -123,7 +124,7 @@ class WordExtractor(BaseExtractor):
base_url = dify_config.FILES_URL
for r_id, rel in doc.part.rels.items():
- if "image" in rel.target_ref:
+ if rel.reltype == RT.IMAGE:
image_count += 1
if rel.is_external:
url = rel.target_ref
diff --git a/api/tests/unit_tests/core/rag/extractor/test_word_extractor.py b/api/tests/unit_tests/core/rag/extractor/test_word_extractor.py
index a4d2f10a10d..54071a9dd50 100644
--- a/api/tests/unit_tests/core/rag/extractor/test_word_extractor.py
+++ b/api/tests/unit_tests/core/rag/extractor/test_word_extractor.py
@@ -17,11 +17,13 @@ from docx.opc.part import Part
from docx.oxml import OxmlElement
from docx.oxml.ns import qn
from httpx import Response
+from PIL import Image
from sqlalchemy import event, select
from sqlalchemy.orm import Session
import core.rag.extractor.word_extractor as we
from core.rag.extractor.word_extractor import WordExtractor
+from extensions.storage.opendal_storage import OpenDALStorage
from models.model import UploadFile
from tests.helpers.unit.config_override import apply_config_overrides
@@ -742,3 +744,63 @@ def test_parse_cell_paragraph_hyperlink_in_table_cell_mailto():
finally:
if os.path.exists(tmp_path):
os.remove(tmp_path)
+
+
+def local_image_storage(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> OpenDALStorage:
+ local_storage = OpenDALStorage(scheme="fs", root=str(tmp_path / "storage"))
+ monkeypatch.setattr(we.storage, "storage_runner", local_storage, raising=False)
+ apply_config_overrides(monkeypatch, FILES_URL="http://files.local", STORAGE_TYPE="local")
+ return local_storage
+
+
[email protected]("part_name", ["image1.png", "logo.png", "Logo.PNG"])
+def test_extract_preserves_image_relationships_regardless_of_part_name(
+ tmp_path: Path, sqlite_session: Session, local_image_storage: OpenDALStorage, part_name: str
+) -> None:
+ png = io.BytesIO()
+ Image.new("RGB", (2, 2), "red").save(png, format="PNG")
+ image_bytes = png.getvalue()
+ png.seek(0)
+ doc = Document()
+ doc.add_paragraph("Before logo")
+ doc.add_picture(png)
+ doc.add_paragraph("After logo")
+ image_rel = next(rel for rel in doc.part.rels.values() if rel.reltype == RT.IMAGE)
+ image_rel.target_part._partname = PackURI(f"/word/media/{part_name}")
+ path = tmp_path / "logo.docx"
+ doc.save(str(path))
+ assert len(Document(str(path)).inline_shapes) == 1
+
+ content = (
+ WordExtractor(
+ str(path),
+ "00000000-0000-0000-0000-000000000001",
+ "00000000-0000-0000-0000-000000000002",
+ session=sqlite_session,
+ )
+ .extract()[0]
+ .page_content
+ )
+
+ upload = sqlite_session.scalars(select(UploadFile)).one()
+ image_link = f"![image](http://files.local/files/{upload.id}/file-preview)"
+ assert content.index("Before logo") < content.index(image_link) < content.index("After logo")
+ assert local_image_storage.load_once(upload.key) == image_bytes
+
+
+def test_extract_ignores_non_image_relationship_with_image_in_target(
+ tmp_path: Path, sqlite_session: Session, local_image_storage: OpenDALStorage
+) -> None:
+ doc = Document()
+ doc.add_paragraph("Visible text")
+ metadata = Part(PackURI("/word/image-metadata.xml"), "application/xml", b"<metadata/>")
+ doc.part.rels.add_relationship(RT.CUSTOM_XML, metadata, "rIdMetadata")
+ path = tmp_path / "metadata.docx"
+ doc.save(str(path))
+
+ content = WordExtractor(str(path), "tenant", "user", session=sqlite_session).extract()[0].page_content
+
+ assert content == "Visible text"
+ assert sqlite_session.scalars(select(UploadFile)).all() == []
+ assert not local_image_storage.exists("image_files/")