Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(pdf): serialize PDFium rendering across parser instances (#5952)

Qin Haojie92285e5b373e
diff --git a/openviking/parse/parsers/pdf.py b/openviking/parse/parsers/pdf.py
index f46a5cec8..509fb55f3 100644
--- a/openviking/parse/parsers/pdf.py
+++ b/openviking/parse/parsers/pdf.py
@@ -17,6 +17,7 @@ import base64
import hashlib
import io
import re
+import threading
import time
from collections import Counter, defaultdict
from pathlib import Path
@@ -36,6 +37,9 @@ from openviking_cli.utils.config.parser_config import PDFConfig
logger = get_logger(__name__)
+# PDFium is not thread-safe, even when rendering different documents.
+_PDFIUM_LOCK = threading.Lock()
+
class PDFParser(BaseParser):
"""
@@ -671,7 +675,8 @@ class PDFParser(BaseParser):
return None
cropped = page.crop(bbox)
- page_image = cropped.to_image(resolution=self.config.image_resolution)
+ with _PDFIUM_LOCK:
+ page_image = cropped.to_image(resolution=self.config.image_resolution)
buffer = io.BytesIO()
page_image.save(buffer, format="PNG")
diff --git a/tests/parse/test_pdf_bookmark_extraction.py b/tests/parse/test_pdf_bookmark_extraction.py
index 441540e51..b7cf3de47 100644
--- a/tests/parse/test_pdf_bookmark_extraction.py
+++ b/tests/parse/test_pdf_bookmark_extraction.py
@@ -7,8 +7,11 @@ Verifies that _extract_bookmarks correctly extracts bookmark entries
and that _convert_local injects them as markdown headings.
"""
+import time
+from concurrent.futures import ThreadPoolExecutor
from contextlib import nullcontext
from pathlib import Path
+from threading import Barrier, Lock
from types import SimpleNamespace
from unittest.mock import MagicMock, patch
@@ -488,17 +491,37 @@ class TestExtractImages:
def setup_method(self):
self.parser = PDFParser()
- def test_extract_image_renders_cropped_bbox_as_png(self):
- page = _FakeImagePage(image_bytes=b"rendered-png")
+ def test_extract_image_renders_cropped_bbox_as_png(self, monkeypatch):
+ pages = [_FakeImagePage(image_bytes=b"rendered-png") for _ in range(2)]
+ parsers = [self.parser, PDFParser()]
+ start = Barrier(2)
+ rendering = Lock()
+ original_render = _FakeCroppedPage.to_image
+
+ def render(page, resolution):
+ assert rendering.acquire(blocking=False), "Concurrent PDFium calls"
+ try:
+ # Native rendering releases the GIL; allow the other thread to run.
+ time.sleep(0.05)
+ return original_render(page, resolution)
+ finally:
+ rendering.release()
+
+ monkeypatch.setattr(_FakeCroppedPage, "to_image", render)
+
+ def extract(index):
+ start.wait(timeout=5)
+ return parsers[index]._extract_image_from_page(
+ pages[index],
+ {"x0": 10, "top": 20, "x1": 40, "bottom": 60},
+ )
- image_data = self.parser._extract_image_from_page(
- page,
- {"x0": 10, "top": 20, "x1": 40, "bottom": 60},
- )
+ with ThreadPoolExecutor(max_workers=2) as executor:
+ assert list(executor.map(extract, range(2))) == [b"rendered-png"] * 2
- assert image_data == b"rendered-png"
- assert page.crop_calls == [(10, 20, 40, 60)]
- assert page.cropped_page.resolution_calls == [self.parser.config.image_resolution]
+ for parser, page in zip(parsers, pages, strict=True):
+ assert page.crop_calls == [(10, 20, 40, 60)]
+ assert page.cropped_page.resolution_calls == [parser.config.image_resolution]
def test_extract_image_clamps_bbox_to_page_bounds(self):
page = _FakeImagePage()