Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

fix(pdf): stop digits rendering as emoji glyphs in exported PDFs (#5050)

LearningCircuit85daefa43fc5
diff --git a/changelog.d/+pdf-digits-emoji-font.bugfix.md b/changelog.d/+pdf-digits-emoji-font.bugfix.md
new file mode 100644
index 0000000000..e5ffd6decb
--- /dev/null
+++ b/changelog.d/+pdf-digits-emoji-font.bugfix.md
@@ -0,0 +1 @@
+Fixed numbers in exported PDFs rendering as wide, spaced-out emoji glyphs ("2 0 2 6" instead of "2026"). Digits 0-9 carry the Unicode Emoji property, so listing "Noto Color Emoji" in the PDF font stacks (#4730) made Pango draw every digit with the emoji font. Emoji families are removed from the stacks; emoji still render via fontconfig per-character fallback to the installed emoji font.
diff --git a/src/local_deep_research/web/services/pdf_service.py b/src/local_deep_research/web/services/pdf_service.py
index c73bcbfaec..e197ca21ea 100644
--- a/src/local_deep_research/web/services/pdf_service.py
+++ b/src/local_deep_research/web/services/pdf_service.py
@@ -140,17 +140,17 @@ def get_weasyprint_install_instructions() -> str:
# Glyphs still require the corresponding system font (e.g.
# fonts-noto-cjk) to actually be installed.
#
-# The emoji families at the tail of each stack cover emoji
-# codepoints that Arial and the CJK fonts lack entirely — without
-# them, emojis in the markdown render as "tofu" boxes in the PDF
-# even though they look fine in the HTML view. Color first
-# (preferred when WeasyPrint can render the CBDT/CBLC bitmap
-# tables — the format Debian's fonts-noto-color-emoji ships),
-# then a monochrome fallback for renderers without CBDT support.
-# Note: "Noto Color Emoji" is provided by fonts-noto-color-emoji;
-# the monochrome "Noto Emoji" is a separate package
-# (fonts-noto-emoji / fonts-noto-extra) and is a best-effort
-# fallback for hosts that happen to have it installed.
+# Emoji are deliberately NOT listed in these stacks. They render
+# through Pango/fontconfig per-character fallback to whichever emoji
+# font is installed (fonts-noto-color-emoji is bundled in the official
+# Docker image; docs/faq.md covers other platforms). Listing an emoji
+# family explicitly (as #4730 did) makes Pango route every digit 0-9
+# plus '#' and '*' — codepoints that carry the Unicode Emoji property
+# and have glyphs in Noto Color Emoji — to the emoji font, even though
+# earlier families in the stack cover them. The result: all numbers in
+# exported reports rendered as wide, square emoji glyphs ("2 0 2 6"
+# instead of "2026"). Regression test:
+# test_minimal_css_excludes_emoji_font_families.
MINIMAL_CSS = """
@page {
size: A4;
@@ -163,9 +163,7 @@ body {
"PingFang SC", "PingFang TC", "Hiragino Sans",
"Hiragino Kaku Gothic ProN", "Apple SD Gothic Neo",
"Microsoft YaHei", "Microsoft JhengHei",
- "Yu Gothic", "Malgun Gothic", "SimSun",
- "Noto Color Emoji", "Noto Emoji", "Segoe UI Emoji",
- "Apple Color Emoji", sans-serif;
+ "Yu Gothic", "Malgun Gothic", "SimSun", sans-serif;
font-size: 10pt;
line-height: 1.4;
}
@@ -198,9 +196,7 @@ code, pre {
"Noto Sans Mono CJK TC", "Noto Sans Mono CJK JP",
"Noto Sans Mono CJK KR", "Noto Sans CJK SC",
"PingFang SC", "Hiragino Sans", "Apple SD Gothic Neo",
- "Microsoft YaHei", "SimSun",
- "Noto Color Emoji", "Noto Emoji", "Segoe UI Emoji",
- "Apple Color Emoji";
+ "Microsoft YaHei", "SimSun";
background-color: #f5f5f5;
}
@@ -246,7 +242,7 @@ class PDFService:
metadata: Optional metadata dict (author, date, etc.)
custom_css: Optional CSS string layered on top of the default
stylesheet. Rules here win on equal specificity via the
- cascade, but the default's CJK/emoji font fallbacks and
+ cascade, but the default's CJK font fallbacks and
page setup are always applied.
Returns:
@@ -272,9 +268,9 @@ class PDFService:
# Always apply the default stylesheet first, then layer any
# caller-provided custom_css on top. WeasyPrint resolves
# conflicts by cascade order, so later stylesheets win on
- # equal specificity — this preserves the CJK and emoji font
- # fallbacks in MINIMAL_CSS even when a caller supplies their
- # own CSS, while still letting them override any default.
+ # equal specificity — this preserves the CJK font fallbacks
+ # in MINIMAL_CSS even when a caller supplies their own CSS,
+ # while still letting them override any default.
css_list = [self.minimal_css]
if custom_css:
css_list.append(CSS(string=custom_css))
diff --git a/tests/api_tests_with_login/pytest_tests/test_export_formats.py b/tests/api_tests_with_login/pytest_tests/test_export_formats.py
index 3e654ad2bd..ca98de0a25 100644
--- a/tests/api_tests_with_login/pytest_tests/test_export_formats.py
+++ b/tests/api_tests_with_login/pytest_tests/test_export_formats.py
@@ -96,8 +96,14 @@ def test_export_latex(auth_session, base_url):
print(f" First 200 chars: {content[:200]}...")
-def test_export_pdf_via_javascript(auth_session, base_url):
- """Test PDF export (generated client-side via JavaScript)"""
+def test_export_pdf(auth_session, base_url):
+ """Test exporting research report as PDF via the server-side endpoint.
+
+ This is the endpoint the UI's "Download PDF" button calls
+ (results.js → POST /api/v1/research/<id>/export/pdf), rendered
+ server-side by WeasyPrint. Also guards the digit-capture regression
+ (#5050): digits must be drawn with a text font, not an emoji font.
+ """
session, csrf_token = auth_session
# First, start a simple research
@@ -140,24 +146,60 @@ def test_export_pdf_via_javascript(auth_session, base_url):
# Additional wait to ensure report is saved
time.sleep(2) # allow: unmarked-sleep
- # Get the research report content
- report_response = session.get(f"{base_url}/api/report/{research_id}")
- assert report_response.status_code == 200
+ # Export as PDF through the endpoint the UI download button uses
+ export_response = session.post(
+ f"{base_url}/api/v1/research/{research_id}/export/pdf"
+ )
- report_data = report_response.json()
- assert "content" in report_data, "No content in report"
+ assert export_response.status_code == 200, (
+ f"PDF export failed: {export_response.text}"
+ )
- content = report_data.get("content", "")
- assert len(content) > 100, (
- f"Report content too short for PDF: {len(content)} bytes"
+ content_type = export_response.headers.get("Content-Type", "")
+ assert "application/pdf" in content_type, (
+ f"Unexpected content type: {content_type}"
)
- print(
- f"✓ Report has content suitable for PDF export - {len(content)} bytes"
+ pdf_bytes = export_response.content
+ assert pdf_bytes[:4] == b"%PDF", "Response is not a PDF document"
+ assert len(pdf_bytes) > 1000, f"PDF too small: {len(pdf_bytes)} bytes"
+
+ content_disposition = export_response.headers.get("Content-Disposition", "")
+ assert ".pdf" in content_disposition, (
+ f"Expected .pdf file in Content-Disposition: {content_disposition}"
)
- # Note: PDF generation happens client-side via jsPDF
- # We can't test the actual PDF generation here, but we've verified the content exists
+ # Digit-capture regression (#5050): no digit glyph may come from an
+ # emoji font. Skipped silently when pdfminer isn't installed.
+ try:
+ import io
+
+ from pdfminer.high_level import extract_pages
+ from pdfminer.layout import LTChar, LTTextContainer, LTTextLine
+ except ImportError:
+ print("pdfminer not available — skipping digit font check")
+ else:
+ digit_fonts = set()
+ for page in extract_pages(io.BytesIO(pdf_bytes)):
+ for element in page:
+ if not isinstance(element, LTTextContainer):
+ continue
+ for line in element:
+ if not isinstance(line, LTTextLine):
+ continue
+ for char in line:
+ if (
+ isinstance(char, LTChar)
+ and char.get_text().isdigit()
+ ):
+ digit_fonts.add(char.fontname)
+ emoji_fonts = {f for f in digit_fonts if "emoji" in f.lower()}
+ assert not emoji_fonts, (
+ f"digits in the exported PDF were rendered with an emoji "
+ f"font: {emoji_fonts}"
+ )
+
+ print(f"✓ PDF export successful - {len(pdf_bytes)} bytes")
def test_export_markdown(auth_session, base_url):
diff --git a/tests/web/routes/test_export_pdf_rendering.py b/tests/web/routes/test_export_pdf_rendering.py
new file mode 100644
index 0000000000..f057ee7eba
--- /dev/null
+++ b/tests/web/routes/test_export_pdf_rendering.py
@@ -0,0 +1,133 @@
+"""End-to-end rendering test for the PDF export endpoint.
+
+Drives POST /api/v1/research/<id>/export/pdf — the endpoint the UI's
+"Download PDF" button calls — through the real exporter chain
+(ExporterRegistry → PDFExporter → PDFService → WeasyPrint). Only the
+data layer (DB session, report assembly) is mocked, so the returned
+bytes are a genuine WeasyPrint render.
+
+Regression coverage for the digit-capture bug (#5050): listing an emoji
+family in the PDF font stacks made Pango draw every digit 0-9 with Noto
+Color Emoji, so numbers in exported reports rendered as wide, square
+emoji glyphs ("2 0 2 6" instead of "2026"). The service-level tests in
+tests/web/services/test_pdf_service.py assert the same property on
+PDFService directly; this test pins it to the HTTP surface so a change
+anywhere in the export chain (route, exporter registry, service) that
+reintroduces the bug still fails CI.
+"""
+
+import io
+from contextlib import contextmanager
+from unittest.mock import MagicMock, patch
+
+import pytest
+from flask import Flask
+
+MODULE = "local_deep_research.web.routes.research_routes"
+ASSEMBLY_MOD = "local_deep_research.web.services.report_assembly_service"
+
+# Digit-heavy markdown mirroring a real research report (dates, prices,
+# token counts, citation markers) — the shapes that were unreadable.
+REPORT_MARKDOWN = (
+ "# Pricing report 2026-07-11\n\n"
+ "As of July 11, 2026, plans cost $20/month and allow 44,000 tokens "
+ "per rolling 5-hour window [21][30].\n"
+)
+
+
+def app():
+ flask_app = Flask(__name__)
+ flask_app.secret_key = "test-secret-key"
+ flask_app.config["TESTING"] = True
+
+ from local_deep_research.web.routes.research_routes import research_bp
+
+ flask_app.register_blueprint(research_bp)
+
+ with patch("local_deep_research.web.auth.decorators.db_manager") as mock_db:
+ mock_db.is_user_connected.return_value = True
+ yield flask_app
+
+
+def client(app):
+ return app.test_client()
+
+
[email protected](autouse=True)
+def _inject_session(app):
+ @app.before_request
+ def _set_sess():
+ from flask import session
+
+ session["username"] = "testuser"
+ session["session_id"] = "sid-1"
+
+
+@contextmanager
+def _ctx(session):
+ yield session
+
+
+def _digit_fonts(pdf_bytes):
+ """Return the set of font names used to draw digit glyphs."""
+ from pdfminer.high_level import extract_pages
+ from pdfminer.layout import LTChar, LTTextContainer, LTTextLine
+
+ fonts = set()
+ for page in extract_pages(io.BytesIO(pdf_bytes)):
+ for element in page:
+ if not isinstance(element, LTTextContainer):
+ continue
+ for line in element:
+ if not isinstance(line, LTTextLine):
+ continue
+ for char in line:
+ if isinstance(char, LTChar) and char.get_text().isdigit():
+ fonts.add(char.fontname)
+ return fonts
+
+
+def test_export_pdf_endpoint_renders_digits_with_text_font(client):
+ """The exported PDF must draw digits with a text font, not an emoji font.
+
+ On hosts with no emoji font installed the emoji-font assertion passes
+ trivially, which is correct — there is no emoji font to capture the
+ digits.
+ """
+ from local_deep_research.web.services.pdf_service import (
+ weasyprint_available,
+ )
+
+ if not weasyprint_available():
+ pytest.skip("WeasyPrint system libraries not available")
+ pytest.importorskip("pdfminer.high_level")
+
+ research = MagicMock()
+ research.id = "res-1"
+ research.title = "Pricing report"
+ research.query = "pricing"
+ mock_session = MagicMock()
+ mock_session.query.return_value.filter_by.return_value.first.return_value = research
+
+ with (
+ patch(f"{MODULE}.get_user_db_session", return_value=_ctx(mock_session)),
+ patch(
+ f"{ASSEMBLY_MOD}.assemble_full_report",
+ return_value=REPORT_MARKDOWN,
+ ),
+ ):
+ resp = client.post("/api/v1/research/res-1/export/pdf")
+
+ assert resp.status_code == 200
+ assert resp.content_type == "application/pdf"
+ assert resp.data[:4] == b"%PDF"
+
+ digit_fonts = _digit_fonts(resp.data)
+ assert digit_fonts, "expected the rendered PDF to contain digits"
+ emoji_fonts = {f for f in digit_fonts if "emoji" in f.lower()}
+ assert not emoji_fonts, (
+ f"digits in the exported PDF were rendered with an emoji font: "
+ f"{emoji_fonts}"
+ )
diff --git a/tests/web/services/test_pdf_service.py b/tests/web/services/test_pdf_service.py
index 10d18cfa99..93a66f1fe2 100644
--- a/tests/web/services/test_pdf_service.py
+++ b/tests/web/services/test_pdf_service.py
@@ -33,35 +33,76 @@ class TestPDFServiceInit:
# The CSS string should be accessible
assert service.minimal_css is not None
- def test_minimal_css_includes_emoji_font_families(self):
- """Body and code font stacks must fall back to an emoji font.
-
- Without an emoji family in the stack, WeasyPrint renders emoji
- codepoints as "tofu" boxes because Arial and the CJK fonts have
- no emoji coverage. Regression test for the emoji font stack.
+ def test_minimal_css_excludes_emoji_font_families(self):
+ """No font-family stack may list an emoji font explicitly.
+
+ Digits 0-9, '#' and '*' carry the Unicode Emoji property and have
+ glyphs in Noto Color Emoji, so Pango routes them to an explicitly
+ listed emoji font even when an earlier family in the stack covers
+ them — every number in an exported report then renders as wide,
+ square emoji glyphs ("2 0 2 6" instead of "2026"). Emoji still
+ render via Pango/fontconfig per-character fallback to an installed
+ emoji font, so the stacks must not name one. Regression test for
+ the digit-capture bug introduced by #4730.
"""
- from local_deep_research.web.services.pdf_service import MINIMAL_CSS
-
- assert "Noto Color Emoji" in MINIMAL_CSS
- assert "Noto Emoji" in MINIMAL_CSS
-
- # Both the body and code/mono stacks should have emoji fallbacks.
- # Pull out just the font-family declarations to make the assertion
- # resilient to other CSS edits.
import re
+ from local_deep_research.web.services.pdf_service import MINIMAL_CSS
+
font_family_blocks = re.findall(r"font-family:\s*([^;]+);", MINIMAL_CSS)
assert len(font_family_blocks) >= 2, (
"expected at least body + code font-family declarations"
)
for block in font_family_blocks:
- assert "Noto Color Emoji" in block, (
- f"font-family block missing color emoji fallback: {block!r}"
- )
- assert "Noto Emoji" in block, (
- f"font-family block missing monochrome emoji fallback: {block!r}"
+ assert "emoji" not in block.lower(), (
+ f"font-family block must not list an emoji font (it would "
+ f"capture digits): {block!r}"
)
+ def test_digits_not_rendered_with_emoji_font(self):
+ """Digits in a rendered PDF must come from a text font, not an emoji font.
+
+ Functional regression test for the digit-capture bug: with an emoji
+ family in the font-family stack, Pango renders 0-9 with Noto Color
+ Emoji (wide square glyphs). Renders a digit-heavy document and
+ asserts no digit glyph is drawn with an emoji font. On hosts with
+ no emoji font installed the assertion passes trivially, which is
+ correct — there is no emoji font to capture the digits.
+ """
+ import io
+
+ pytest.importorskip("pdfminer.high_level")
+ from pdfminer.high_level import extract_pages
+ from pdfminer.layout import LTChar, LTTextContainer, LTTextLine
+
+ from local_deep_research.web.services.pdf_service import PDFService
+
+ service = PDFService()
+ pdf_bytes = service.markdown_to_pdf(
+ "As of July 11, 2026, plans cost $20/month [21][30]."
+ )
+
+ digit_fonts = set()
+ for page in extract_pages(io.BytesIO(pdf_bytes)):
+ for element in page:
+ if not isinstance(element, LTTextContainer):
+ continue
+ for line in element:
+ if not isinstance(line, LTTextLine):
+ continue
+ for char in line:
+ if (
+ isinstance(char, LTChar)
+ and char.get_text().isdigit()
+ ):
+ digit_fonts.add(char.fontname)
+
+ assert digit_fonts, "expected the rendered PDF to contain digits"
+ emoji_fonts = {f for f in digit_fonts if "emoji" in f.lower()}
+ assert not emoji_fonts, (
+ f"digits were rendered with an emoji font: {emoji_fonts}"
+ )
+
class TestMarkdownToHTML:
"""Tests for _markdown_to_html method."""
@@ -233,8 +274,8 @@ class TestMarkdownToPDF:
Regression test: previously markdown_to_pdf used an if/else that
dropped self.minimal_css entirely when custom_css was provided,
- silently losing the CJK and emoji font fallbacks. Verify the
- default stylesheet is always passed to WeasyPrint.
+ silently losing the CJK font fallbacks. Verify the default
+ stylesheet is always passed to WeasyPrint.
"""
from unittest.mock import patch, MagicMock
from local_deep_research.web.services.pdf_service import PDFService