fix: stop designed / multi-column resumes silently dropping sections in PDF extraction

pdf.py

@@ -49,9 +49,15 @@ def extract_text_from_pdf(self, pdf_path: str) -> Optional[str]:
 
             with pymupdf.open(pdf_path) as doc:
                 pages = range(doc.page_count)
+                # Resumes are text documents: decorative vector graphics used by
+                # designed templates (Canva, Adobe Express, ...) must not suppress
+                # overlapping text. Without ignore_graphics, such panels are treated
+                # as significant images and their text is dropped, silently losing
+                # whole sections (skills, education, contact).
                 resume_text = to_markdown(
                     doc,
                     pages=pages,
+                    ignore_graphics=True,
                 )
                 logger.debug(
                     f"Extracted text from PDF: {len(resume_text) if resume_text else 0} characters"

tests/pdf_extraction_test.py

@@ -0,0 +1,123 @@
+"""Smoke test: designed / multi-column resume layouts must not silently drop text.
+
+Regression guard for passing ``ignore_graphics=True`` to ``pymupdf_rag.to_markdown``
+in ``pdf.PDFHandler.extract_text_from_pdf``.
+
+Designed resume templates (Canva, Adobe Express, ...) paint sidebar/section content
+on decorative vector panels. Without ``ignore_graphics`` those panels are treated as
+significant images and every text line drawn over them is dropped -- so whole
+sections (skills, education, contact) vanish before the LLM ever sees them.
+
+The fixture recreates the ``resume/sample.pdf`` persona (Barack Obama) in a two-column
+layout with a coloured sidebar, then asserts the sidebar text survives extraction.
+No API key / LLM required -- this exercises only the PDF -> Markdown stage.
+"""
+
+import os
+import sys
+import tempfile
+
+import pymupdf
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+import pdf  # noqa: E402
+
+
+def _build_canva_resume(path):
+    """Two-column 'Canva-style' resume: a coloured sidebar panel (Contact / Skills /
+    Education) beside a plain main column (name / work / awards)."""
+    navy, teal, grey, white, dark = (
+        (0.13, 0.19, 0.28),
+        (0.09, 0.46, 0.49),
+        (0.85, 0.87, 0.90),
+        (1, 1, 1),
+        (0.12, 0.12, 0.14),
+    )
+    doc = pymupdf.open()
+    page = doc.new_page(width=595, height=842)
+    sb = 205
+    # decorative vector graphics (varied shapes => flagged "significant")
+    page.draw_rect(pymupdf.Rect(0, 0, sb, 840), color=None, fill=navy)
+    page.draw_circle(pymupdf.Point(sb / 2, 92), 58, color=white, fill=grey, width=3)
+    page.draw_rect(pymupdf.Rect(24, 300, 182, 304), color=None, fill=teal)
+    page.draw_rect(pymupdf.Rect(24, 470, 182, 474), color=None, fill=teal)
+
+    def txt(x, y, s, size=10, color=dark, bold=False):
+        page.insert_text(
+            pymupdf.Point(x, y),
+            s,
+            fontsize=size,
+            fontname="hebo" if bold else "helv",
+            color=color,
+        )
+
+    # main column (plain white) -- survives extraction either way
+    txt(sb + 22, 62, "Barack Obama", size=26, color=navy, bold=True)
+    txt(sb + 22, 90, "44th President of the United States", size=11)
+    y = 175
+    txt(sb + 22, y, "WORK EXPERIENCE", size=13, color=teal, bold=True)
+    y += 26
+    txt(sb + 22, y, "President of the United States", size=11, bold=True)
+    y += 16
+    txt(sb + 22, y, "The White House  |  January 2009 - January 2017", size=9)
+    y += 20
+    for b in (
+        "Served as the 44th President of the United States.",
+        "Represented the United States in international diplomacy.",
+    ):
+        txt(sb + 22, y, "- " + b, size=9)
+        y += 16
+
+    # sidebar (drawn over the decorative panel) -- this is what the bug drops
+    y = 175
+    txt(28, y, "CONTACT", size=12, color=white, bold=True)
+    y += 22
+    for line in ("Phone: (555) 123-4567", "barack.obama@example.com", "Washington, DC"):
+        txt(28, y, line, size=9, color=white)
+        y += 16
+    y = 325
+    txt(28, y, "SKILLS", size=12, color=white, bold=True)
+    y += 22
+    for line in (
+        "Executive Leadership",
+        "Public Speaking",
+        "Strategic Planning",
+        "International Relations",
+    ):
+        txt(28, y, "- " + line, size=9, color=white)
+        y += 17
+    y = 495
+    txt(28, y, "EDUCATION", size=12, color=white, bold=True)
+    y += 22
+    for line in ("Juris Doctor (J.D.)", "Harvard Law School", "1991"):
+        txt(28, y, line, size=9, color=white)
+        y += 16
+
+    doc.save(path)
+
+
+def test_designed_layout_sidebar_is_not_dropped():
+    # extract_text_from_pdf uses no model; stub the provider so PDFHandler() needs
+    # no API key.
+    pdf.initialize_llm_provider = lambda *a, **k: None
+    handler = pdf.PDFHandler()
+    with tempfile.TemporaryDirectory() as d:
+        fixture = os.path.join(d, "canva_resume.pdf")
+        _build_canva_resume(fixture)
+        text = handler.extract_text_from_pdf(fixture) or ""
+
+    for needed in (
+        "Juris Doctor",
+        "Harvard Law School",
+        "Executive Leadership",
+        "barack.obama@example.com",
+    ):
+        assert needed in text, (
+            f"designed-layout extraction dropped {needed!r} "
+            "(to_markdown(ignore_graphics=True) regression?)"
+        )
+
+
+if __name__ == "__main__":
+    test_designed_layout_sidebar_is_not_dropped()
+    print("OK: designed-layout sidebar preserved")