Files
mnemosyne/mnemosyne/library/services/svg_raster.py
Robert Helewka 224541a4ce 🐾 fix(parsers): rasterize SVG instead of failing ingest
"svg" was in both IMAGE_EXTENSIONS and PYMUPDF_EXTENSIONS, and the image
check ran first, so every SVG reached PIL.Image.open — which cannot decode
vector XML. Ingest raised UnidentifiedImageError, logged "Failed to read
image file_type=svg", and counted documents_parsed_total{status="error"}.
SVG ingest has never worked.

Rasterize to PNG instead, via PyMuPDF (already a dependency). This also
unblocks the vision stage downstream: it sends images to a vision LLM as a
data: URI, which cannot carry image/svg+xml either, so the description and
OCR fields were unreachable for SVG regardless of the parse fix.

Route SVG explicitly before both extension sets. Falling through to
PYMUPDF_EXTENSIONS would parse but not split, and a multi-page Write note
rendered as one document is either blank (no root width/height, so MuPDF
falls back to US-Letter and emits the top-left corner) or an illegible tall
strip (root sized across every stacked page). Splitting per write-page
element is correct for both formats, and since Write never rewrites
existing files both persist indefinitely.

svg_raster is a deliberate twin of daedalus's extraction/svg.py — the repos
share no common package, so the duplication is noted in both docstrings and
fixes belong in both.

lxml is a new dependency: recover=True is needed for the real files that
aren't well-formed XML, and stdlib ElementTree has no equivalent.
2026-07-26 07:41:20 -04:00

264 lines
9.1 KiB
Python

"""SVG rasterization for the ingest pipeline.
The vision stage sends each extracted image to a vision LLM as a ``data:`` URI,
which cannot carry ``image/svg+xml`` — so an SVG has to become raster before it
can be described, OCR'd, or embedded. PyMuPDF (already a dependency for PDF
parsing) renders SVG natively, so this needs no cairo/rsvg system libraries.
The non-obvious part is page splitting. Handwritten notes from the Write
(Stylus Labs) app are a single SVG document holding one or more
``<svg class="write-page">`` children stacked vertically via x/y offsets. Two
root formats exist in the wild and *both* must be split per page:
- Older files carry no width/height on the root ``<svg>``. Rendering the
document as-is makes PyMuPDF fall back to US-Letter and emit the top-left
corner only — ruled lines and no handwriting, in a perfectly valid PNG.
- Newer files (Write commit eeab021) do carry root width/height, spanning the
full stacked extent. Rendering those as-is yields one very tall strip; capped
to a sane longest edge, a 5-page note squashes to ~209px wide, well past
illegible.
Write never rewrites existing files, so the old format is permanent rather than
a migration window. Page geometry lives on the ``write-page`` element and is
identical across both formats, so splitting ignores the root dimensions
entirely and needs no format detection.
.. note::
Daedalus carries a twin of this module at
``backend/daedalus/extraction/svg.py``, which returns base64 for direct
chat attachment. The two are deliberately duplicated rather than shared —
the repos ship no common package — so fixes belong in both.
"""
from __future__ import annotations
import copy
import logging
import re
from lxml import etree
logger = logging.getLogger(__name__)
SVG_NS = "http://www.w3.org/2000/svg"
XLINK_NS = "http://www.w3.org/1999/xlink"
_SVG = f"{{{SVG_NS}}}"
_PAGE_CLASS = "write-page"
# Write documents render on a grey backdrop; pages themselves are transparent,
# so each rendered page needs an explicit white underlay or strokes land on
# black in the flattened PNG.
_PAGE_BACKGROUND = "#ffffff"
#: Longest edge of a rendered page, in pixels. Enough for a vision model to
#: read handwriting without spending tokens on unusable resolution.
DEFAULT_TARGET_PX = 1568
#: Cap on pages rendered from one document.
DEFAULT_MAX_PAGES = 20
# Unquoted attribute value in the root tag, e.g. Write's `width=auto`. Matches
# only bare alphabetic values so numeric or already-quoted attributes are left
# alone.
_UNQUOTED_ATTR = re.compile(rb"(\s[-\w:]+)=([A-Za-z][-\w]*)(?=[\s>])")
_ROOT_TAG = re.compile(rb"<svg[^>]*>")
class SvgRenderError(Exception):
"""Raised when an SVG cannot be parsed or contains no renderable page."""
def _repair_root_tag(data: bytes) -> bytes:
"""Quote unquoted attribute values in the root ``<svg>`` tag.
Some Write files emit ``width=auto height=auto``, which is not valid XML.
A malformation *inside* the root tag defeats ``recover=True`` differently
from one in the body: rather than dropping a subtree, libxml2 abandons the
whole document and yields a bare root, so every page becomes invisible.
Quoting the values first recovers the full tree.
"""
match = _ROOT_TAG.search(data)
if not match:
return data
repaired = _UNQUOTED_ATTR.sub(rb'\1="\2"', match.group(0))
if repaired == match.group(0):
return data
return data[: match.start()] + repaired + data[match.end() :]
def _parser() -> etree.XMLParser:
"""Build the hardened parser used for all untrusted SVG input.
``resolve_entities=False`` blocks XXE — ingest content is untrusted.
``huge_tree`` is required because handwriting path data runs to megabytes.
``recover`` salvages the handful of Write files that emit unescaped
attribute content and are not well-formed XML.
"""
return etree.XMLParser(
huge_tree=True,
resolve_entities=False,
no_network=True,
recover=True,
)
def _dimension(element: etree._Element, name: str) -> float | None:
"""Read a CSS-pixel dimension attribute, tolerating a ``px`` suffix."""
raw = element.get(name)
if not raw:
return None
try:
return float(raw.strip().removesuffix("px"))
except ValueError:
return None
def _viewbox_size(element: etree._Element) -> tuple[float, float] | None:
"""Derive width/height from a viewBox extent."""
raw = element.get("viewBox")
if not raw:
return None
parts = raw.replace(",", " ").split()
if len(parts) != 4:
return None
try:
width, height = float(parts[2]), float(parts[3])
except ValueError:
return None
return (width, height) if width > 0 and height > 0 else None
def _element_size(element: etree._Element) -> tuple[float, float] | None:
"""Resolve an element's rendered size from width/height, else viewBox."""
width = _dimension(element, "width")
height = _dimension(element, "height")
if width and height:
return width, height
return _viewbox_size(element)
def _standalone_page(
page: etree._Element,
defs: etree._Element | None,
width: float,
height: float,
) -> bytes:
"""Wrap one ``write-page`` element as its own renderable SVG document.
The page is repositioned to the origin (its x/y place it within the stacked
parent) and given an explicit viewBox so the renderer has an unambiguous
size. ``defs`` is copied in because pen and ruling definitions live on the
root and are referenced by page content.
"""
root = etree.Element(f"{_SVG}svg", nsmap={None: SVG_NS, "xlink": XLINK_NS})
root.set("width", str(width))
root.set("height", str(height))
root.set("viewBox", f"0 0 {width} {height}")
background = etree.SubElement(root, f"{_SVG}rect")
background.set("width", "100%")
background.set("height", "100%")
background.set("fill", _PAGE_BACKGROUND)
if defs is not None:
root.append(copy.deepcopy(defs))
element = copy.deepcopy(page)
for positional in ("x", "y"):
element.attrib.pop(positional, None)
element.set("viewBox", f"0 0 {width} {height}")
root.append(element)
return etree.tostring(root)
def split_svg_pages(data: bytes) -> list[bytes]:
"""Split an SVG into one standalone document per renderable page.
Write multi-page notes yield one document per ``write-page`` child. Any
other SVG (a diagram, a logo) yields a single document — the input itself,
which the renderer sizes from its own width/height or viewBox.
:param data: Raw SVG bytes.
:returns: One or more standalone SVG documents, in page order.
:raises SvgRenderError: If the SVG cannot be parsed or has no usable size.
"""
try:
root = etree.fromstring(_repair_root_tag(data), _parser())
except etree.XMLSyntaxError as exc:
raise SvgRenderError(f"Could not parse SVG: {exc}") from exc
if root is None:
raise SvgRenderError("Could not parse SVG: no root element.")
defs = root.find(f"{_SVG}defs")
pages: list[bytes] = []
for element in root.iter(f"{_SVG}svg"):
if element is root:
continue
if _PAGE_CLASS not in (element.get("class") or "").split():
continue
size = _element_size(element)
if not size:
continue
pages.append(_standalone_page(element, defs, *size))
if pages:
return pages
# Generic SVG: render as-is. It must still be sizeable, or the renderer
# would silently substitute a default page box.
if not _element_size(root):
raise SvgRenderError(
"SVG has no width/height or viewBox, so its size is undefined."
)
return [data]
def render_svg_pages(
data: bytes,
max_pages: int = DEFAULT_MAX_PAGES,
target_px: int = DEFAULT_TARGET_PX,
) -> list[tuple[bytes, int, int]]:
"""Rasterize an SVG to one PNG per page.
:param data: Raw SVG bytes.
:param max_pages: Cap on pages rendered.
:param target_px: Longest edge of each rendered page, in pixels.
:returns: One ``(png_bytes, width, height)`` tuple per page, in page order.
:raises SvgRenderError: If nothing could be rendered.
"""
import fitz
pages = split_svg_pages(data)
rendered: list[tuple[bytes, int, int]] = []
for index, page_svg in enumerate(pages[:max_pages]):
try:
with fitz.open(stream=page_svg, filetype="svg") as document:
page = document[0]
longest = max(page.rect.width, page.rect.height)
zoom = target_px / longest if longest else 1.0
pixmap = page.get_pixmap(matrix=fitz.Matrix(zoom, zoom))
rendered.append(
(pixmap.tobytes("png"), pixmap.width, pixmap.height)
)
except Exception as exc:
raise SvgRenderError(
f"Could not render SVG page {index + 1}: {exc}"
) from exc
if not rendered:
raise SvgRenderError("No renderable pages in SVG.")
logger.info(
"Rasterized SVG pages_rendered=%d total_pages=%d target_px=%d",
len(rendered),
len(pages),
target_px,
)
return rendered