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(api): stop accumulating duplicate extracted PDF images on repeated previews (#44684)

uhhgxge08b084b197f
diff --git a/api/core/rag/extractor/extract_processor.py b/api/core/rag/extractor/extract_processor.py
index fa9d0437bdd..5f26f16717a 100644
--- a/api/core/rag/extractor/extract_processor.py
+++ b/api/core/rag/extractor/extract_processor.py
@@ -153,6 +153,7 @@ class ExtractProcessor:
file_path,
upload_file.tenant_id if upload_file else None,
upload_file.created_by if upload_file else None,
+ source_file_id=upload_file.id if upload_file else None,
session=session,
)
elif file_extension in {".md", ".markdown", ".mdx"}:
@@ -203,6 +204,7 @@ class ExtractProcessor:
file_path,
upload_file.tenant_id if upload_file else None,
upload_file.created_by if upload_file else None,
+ source_file_id=upload_file.id if upload_file else None,
session=session,
)
elif file_extension in {".md", ".markdown", ".mdx"}:
diff --git a/api/core/rag/extractor/pdf_extractor.py b/api/core/rag/extractor/pdf_extractor.py
index 778514cd27d..4bf2d2b3b07 100644
--- a/api/core/rag/extractor/pdf_extractor.py
+++ b/api/core/rag/extractor/pdf_extractor.py
@@ -1,14 +1,16 @@
"""Abstract interface for document loader implementations."""
import contextlib
+import hashlib
import io
import logging
import uuid
from collections.abc import Iterator
-from typing import override
+from typing import TypedDict, override
import pypdfium2
import pypdfium2.raw as pdfium_c
+from sqlalchemy import select
from sqlalchemy.orm import Session
from configs import dify_config
@@ -25,6 +27,14 @@ from models.model import UploadFile
logger = logging.getLogger(__name__)
+class _PdfImageCandidate(TypedDict):
+ image_bytes: bytes
+ image_ext: str
+ mime_type: str
+ content_hash: str
+ file_key: str
+
+
class PdfExtractor(BaseExtractor):
"""
PdfExtractor is used to extract text and images from PDF files.
@@ -34,6 +44,9 @@ class PdfExtractor(BaseExtractor):
tenant_id: Workspace ID.
user_id: ID of the user performing the extraction.
file_cache_key: Optional cache key for the extracted text.
+ source_file_id: Optional id of the source upload file. When provided,
+ extracted images are persisted under deterministic storage keys so
+ repeated extractions of the same source reuse the same assets.
session: Session used to persist extracted images.
"""
@@ -51,6 +64,7 @@ class PdfExtractor(BaseExtractor):
)
MAX_MAGIC_LEN = max(len(m) for m, _, _ in IMAGE_FORMATS)
_session: Session | None
+ _source_file_id: str | None
def __init__(
self,
@@ -58,6 +72,7 @@ class PdfExtractor(BaseExtractor):
tenant_id: str | None = None,
user_id: str | None = None,
file_cache_key: str | None = None,
+ source_file_id: str | None = None,
*,
session: Session | None = None,
):
@@ -66,6 +81,7 @@ class PdfExtractor(BaseExtractor):
self._tenant_id = tenant_id
self._user_id = user_id
self._file_cache_key = file_cache_key
+ self._source_file_id = source_file_id
self._session = session
@override
@@ -106,7 +122,7 @@ class PdfExtractor(BaseExtractor):
content = text_page.get_text_range()
text_page.close()
- image_content = self._extract_images(page)
+ image_content = self._extract_images(page, page_number)
if image_content:
content += "\n" + image_content
@@ -116,13 +132,20 @@ class PdfExtractor(BaseExtractor):
finally:
pdf_reader.close()
- def _extract_images(self, page) -> str:
+ def _extract_images(self, page, page_number: int = 0) -> str:
"""
Extract images from a PDF page, save them to storage and database,
and return markdown image links.
+ When a source file id is available, images are persisted under
+ deterministic keys derived from the source upload file, page, object
+ index, and content hash so repeated extractions of the same unchanged
+ source reuse the existing UploadFile rows and storage objects instead
+ of accumulating duplicates.
+
Args:
page: pypdfium2 page object.
+ page_number: Zero-based index of the page within the PDF.
Returns:
Markdown string containing links to the extracted images.
@@ -133,13 +156,12 @@ class PdfExtractor(BaseExtractor):
# image extraction and return text only.
return ""
- image_content = []
- upload_files = []
base_url = dify_config.FILES_URL
+ candidates: list[_PdfImageCandidate] = []
try:
image_objects = page.get_objects(filter=(pdfium_c.FPDF_PAGEOBJ_IMAGE,))
- for obj in image_objects:
+ for obj_index, obj in enumerate(image_objects):
try:
# Extract image bytes
img_byte_arr = io.BytesIO()
@@ -163,37 +185,81 @@ class PdfExtractor(BaseExtractor):
if not image_ext or not mime_type:
continue
- file_uuid = str(uuid.uuid4())
- file_key = "image_files/" + self._tenant_id + "/" + file_uuid + "." + image_ext
-
- storage.save(file_key, img_bytes)
-
- # save file to db
- upload_file = UploadFile(
- tenant_id=self._tenant_id,
- storage_type=StorageType(dify_config.STORAGE_TYPE),
- key=file_key,
- name=file_key,
- size=len(img_bytes),
- extension=image_ext,
- mime_type=mime_type,
- created_by=self._user_id,
- created_by_role=CreatorUserRole.ACCOUNT,
- created_at=naive_utc_now(),
- used=True,
- used_by=self._user_id,
- used_at=naive_utc_now(),
+ content_hash = hashlib.sha256(img_bytes).hexdigest()
+ if self._source_file_id:
+ # Deterministic key so retried or repeated extractions
+ # reuse the same assets (mirrors ExcelExtractor).
+ file_key = (
+ f"image_files/{self._tenant_id}/{self._source_file_id}/"
+ f"p{page_number}_{obj_index}_{content_hash}.{image_ext}"
+ )
+ else:
+ # Without a source upload file id there is no stable
+ # identity to deduplicate against; keep unique keys.
+ file_key = f"image_files/{self._tenant_id}/{uuid.uuid4()}.{image_ext}"
+
+ candidates.append(
+ _PdfImageCandidate(
+ image_bytes=img_bytes,
+ image_ext=image_ext,
+ mime_type=mime_type,
+ content_hash=content_hash,
+ file_key=file_key,
+ )
)
- upload_files.append(upload_file)
- image_content.append(f"![image]({base_url}/files/{upload_file.id}/file-preview)")
except Exception as e:
logger.warning("Failed to extract image from PDF: %s", e)
continue
except Exception as e:
logger.warning("Failed to get objects from PDF page: %s", e)
- if upload_files:
- session = self._session or db.session
- session.add_all(upload_files)
+
+ if not candidates:
+ return ""
+
+ session = self._session or db.session
+
+ upload_files_by_key: dict[str, UploadFile] = {}
+ if self._source_file_id:
+ existing_upload_files = session.scalars(
+ select(UploadFile).where(
+ UploadFile.tenant_id == self._tenant_id,
+ UploadFile.key.in_([candidate["file_key"] for candidate in candidates]),
+ )
+ ).all()
+ upload_files_by_key = {upload_file.key: upload_file for upload_file in existing_upload_files}
+
+ image_content = []
+ new_upload_files: list[UploadFile] = []
+ for candidate in candidates:
+ upload_file = upload_files_by_key.get(candidate["file_key"])
+ if upload_file is None:
+ try:
+ storage.save(candidate["file_key"], candidate["image_bytes"])
+ except Exception as e:
+ logger.warning("Failed to persist extracted PDF image: %s", e)
+ continue
+ upload_file = UploadFile(
+ tenant_id=self._tenant_id,
+ storage_type=StorageType(dify_config.STORAGE_TYPE),
+ key=candidate["file_key"],
+ name=candidate["file_key"],
+ size=len(candidate["image_bytes"]),
+ extension=candidate["image_ext"],
+ mime_type=candidate["mime_type"],
+ created_by=self._user_id,
+ created_by_role=CreatorUserRole.ACCOUNT,
+ created_at=naive_utc_now(),
+ used=True,
+ used_by=self._user_id,
+ used_at=naive_utc_now(),
+ hash=candidate["content_hash"],
+ )
+ upload_files_by_key[candidate["file_key"]] = upload_file
+ new_upload_files.append(upload_file)
+ image_content.append(f"![image]({base_url}/files/{upload_file.id}/file-preview)")
+
+ if new_upload_files:
+ session.add_all(new_upload_files)
if self._session is None:
session.commit()
return "\n".join(image_content)
diff --git a/api/tests/unit_tests/core/rag/extractor/test_pdf_extractor.py b/api/tests/unit_tests/core/rag/extractor/test_pdf_extractor.py
index 0b9a211c5d5..8a221fc883d 100644
--- a/api/tests/unit_tests/core/rag/extractor/test_pdf_extractor.py
+++ b/api/tests/unit_tests/core/rag/extractor/test_pdf_extractor.py
@@ -1,9 +1,14 @@
+import hashlib
import io
+from collections.abc import Iterator
+from contextlib import closing
from dataclasses import dataclass
from unittest.mock import MagicMock, patch
from uuid import uuid4
+import pypdfium2
import pytest
+from PIL import Image
from sqlalchemy import select
from sqlalchemy.orm import Session
@@ -157,7 +162,7 @@ def test_extract_calls_extract_images(monkeypatch: pytest.MonkeyPatch) -> None:
extractor = pe.PdfExtractor(file_path="test.pdf", tenant_id=TENANT_ID, user_id=USER_ID)
# Mock _extract_images to return a known string
- monkeypatch.setattr(extractor, "_extract_images", lambda _page: "![image](img_url)")
+ monkeypatch.setattr(extractor, "_extract_images", lambda _page, _page_number=0: "![image](img_url)")
documents = list(extractor.extract())
@@ -210,3 +215,150 @@ def test_extract_images_skipped_without_tenant_context() -> None:
assert result == ""
mock_page.get_objects.assert_not_called()
+
+
+def image_pdf() -> Iterator[tuple[pypdfium2.PdfDocument, bytes]]:
+ """Build real PDF pages so image extraction exercises PDFium, too."""
+ image_buffer = io.BytesIO()
+ with Image.new("RGB", (8, 8), "red") as image:
+ image.save(image_buffer, format="JPEG")
+ image_bytes = image_buffer.getvalue()
+
+ with pypdfium2.PdfDocument.new() as pdf:
+ for _ in range(2):
+ with closing(pdf.new_page(100, 100)) as page:
+ image_object = pypdfium2.PdfImage.new(pdf)
+ image_object.load_jpeg(io.BytesIO(image_bytes), inline=True)
+ page.insert_obj(image_object)
+ page.gen_content()
+ yield pdf, image_bytes
+
+
[email protected]("sqlite_session", [(UploadFile,)], indirect=True)
[email protected]("inject_session", [False, True])
+def test_extract_images_with_source_file_id_reuses_existing_assets(
+ mock_dependencies: _Dependencies,
+ image_pdf: tuple[pypdfium2.PdfDocument, bytes],
+ inject_session: bool,
+) -> None:
+ """Repeated extraction of the same source must not accumulate duplicate assets."""
+ source_file_id = str(uuid4())
+ pdf, jpeg_bytes = image_pdf
+
+ with closing(pdf[0]) as page:
+ first_extractor = pe.PdfExtractor(
+ file_path="test.pdf",
+ tenant_id=TENANT_ID,
+ user_id=USER_ID,
+ source_file_id=source_file_id,
+ session=mock_dependencies.session if inject_session else None,
+ )
+ first_link = first_extractor._extract_images(page, 2)
+
+ second_extractor = pe.PdfExtractor(
+ file_path="test.pdf",
+ tenant_id=TENANT_ID,
+ user_id=USER_ID,
+ source_file_id=source_file_id,
+ session=mock_dependencies.session if inject_session else None,
+ )
+ second_link = second_extractor._extract_images(page, 2)
+
+ rows = mock_dependencies.session.scalars(select(UploadFile)).all()
+ assert len(rows) == 1
+ upload_file = rows[0]
+ assert f"![image](http://files.local/files/{upload_file.id}/file-preview)" == first_link
+ assert first_link == second_link
+ assert mock_dependencies.storage.saves == [(upload_file.key, jpeg_bytes)]
+ expected_hash = hashlib.sha256(jpeg_bytes).hexdigest()
+ assert upload_file.key == f"image_files/{TENANT_ID}/{source_file_id}/p2_0_{expected_hash}.jpg"
+ assert upload_file.hash == expected_hash
+
+
[email protected]("sqlite_session", [(UploadFile,)], indirect=True)
+def test_extract_images_with_source_file_id_keeps_unique_keys_per_page(
+ mock_dependencies: _Dependencies,
+ image_pdf: tuple[pypdfium2.PdfDocument, bytes],
+) -> None:
+ """The same image content on different pages maps to distinct assets."""
+ source_file_id = str(uuid4())
+ pdf, jpeg_bytes = image_pdf
+ extractor = pe.PdfExtractor(
+ file_path="test.pdf",
+ tenant_id=TENANT_ID,
+ user_id=USER_ID,
+ source_file_id=source_file_id,
+ )
+ with closing(pdf[0]) as first_page, closing(pdf[1]) as second_page:
+ first_link = extractor._extract_images(first_page, 0)
+ second_link = extractor._extract_images(second_page, 1)
+
+ rows = mock_dependencies.session.scalars(select(UploadFile)).all()
+ assert len(rows) == 2
+ assert {row.key for row in rows} == {
+ f"image_files/{TENANT_ID}/{source_file_id}/p0_0_{hashlib.sha256(jpeg_bytes).hexdigest()}.jpg",
+ f"image_files/{TENANT_ID}/{source_file_id}/p1_0_{hashlib.sha256(jpeg_bytes).hexdigest()}.jpg",
+ }
+ assert first_link != second_link
+ assert len(mock_dependencies.storage.saves) == 2
+
+
[email protected]("sqlite_session", [(UploadFile,)], indirect=True)
[email protected]("inject_session", [False, True])
[email protected]("fail_all", [False, True])
+def test_extract_images_storage_failure_isolated_and_retryable(
+ mock_dependencies: _Dependencies,
+ image_pdf: tuple[pypdfium2.PdfDocument, bytes],
+ monkeypatch: pytest.MonkeyPatch,
+ inject_session: bool,
+ fail_all: bool,
+) -> None:
+ """Failed saves create no rows; other images and later retries still succeed."""
+ pdf, jpeg_bytes = image_pdf
+ source_file_id = str(uuid4())
+ extractor = pe.PdfExtractor(
+ file_path="test.pdf",
+ tenant_id=TENANT_ID,
+ user_id=USER_ID,
+ source_file_id=source_file_id,
+ session=mock_dependencies.session if inject_session else None,
+ )
+ save = mock_dependencies.storage.save
+ attempts = 0
+
+ def fail_save(key: str, data: bytes) -> None:
+ nonlocal attempts
+ attempts += 1
+ if fail_all or attempts == 1:
+ raise OSError("Storage unavailable")
+ save(key, data)
+
+ with closing(pdf[0]) as page:
+ image_object = pypdfium2.PdfImage.new(pdf)
+ image_object.load_jpeg(io.BytesIO(jpeg_bytes), inline=True)
+ page.insert_obj(image_object)
+ page.gen_content()
+ monkeypatch.setattr(mock_dependencies.storage, "save", fail_save)
+ first_result = extractor._extract_images(page)
+
+ rows = mock_dependencies.session.scalars(select(UploadFile)).all()
+ assert attempts == 2
+ assert len(rows) == (0 if fail_all else 1)
+ assert len(mock_dependencies.storage.saves) == len(rows)
+ if fail_all:
+ assert first_result == ""
+ else:
+ assert first_result == f"![image](http://files.local/files/{rows[0].id}/file-preview)"
+ assert "/p0_1_" in rows[0].key
+
+ monkeypatch.setattr(mock_dependencies.storage, "save", save)
+ retry_result = extractor._extract_images(page)
+ assert extractor._extract_images(page) == retry_result
+
+ rows = mock_dependencies.session.scalars(select(UploadFile)).all()
+ assert len(rows) == 2
+ assert len(mock_dependencies.storage.saves) == 2
+ assert retry_result.count("![image]") == 2
+ for row in rows:
+ assert f"/files/{row.id}/file-preview" in retry_result