mnemosyne/library/services/parsers.py

@@ -31,8 +31,11 @@ PYMUPDF_EXTENSIONS = {
 # Plain text extensions — read directly, no PyMuPDF needed
 PLAINTEXT_EXTENSIONS = {"txt", "md", "csv", "tsv", "log", "json", "yaml", "yml", "xml"}
 
-# Image extensions — store as Image nodes directly
+# Image extensions — store as Image nodes directly.
-IMAGE_EXTENSIONS = {"jpg", "jpeg", "png", "gif", "bmp", "tiff", "tif", "webp", "svg"}
+# SVG is deliberately absent: it is vector XML that Pillow cannot decode and
+# that the vision stage cannot send as a data URI, so it gets rasterized by
+# _parse_svg_file instead.
+IMAGE_EXTENSIONS = {"jpg", "jpeg", "png", "gif", "bmp", "tiff", "tif", "webp"}
 
 # Minimum image dimensions to extract (skip tiny icons/bullets)
 MIN_IMAGE_WIDTH = 50
@@ -98,6 +101,12 @@ class DocumentParser:
         if file_type in PLAINTEXT_EXTENSIONS:
             return self._parse_plaintext(file_path, file_type)
 
+        # Checked before PYMUPDF_EXTENSIONS: PyMuPDF can open an SVG, but
+        # rendering a multi-page Write note as one document yields a blank or
+        # illegible image (see svg_raster), so it needs page-aware handling.
+        if file_type == "svg":
+            return self._parse_svg_file(file_path, file_type)
+
         if file_type in IMAGE_EXTENSIONS:
             return self._parse_image_file(file_path, file_type)
 
@@ -309,6 +318,61 @@ class DocumentParser:
                 file_type=file_type,
             )
 
+    def _parse_svg_file(self, file_path: str, file_type: str) -> ParseResult:
+        """
+        Rasterize an SVG into one ExtractedImage per page.
+
+        SVG is vector XML: Pillow cannot decode it and the vision stage cannot
+        put it in a data URI, so it is rendered to PNG here. Multi-page Write
+        notes become one image per page so each page reaches vision/OCR at a
+        legible size.
+
+        :param file_path: Path to the SVG file.
+        :param file_type: Normalized file extension ("svg").
+        :returns: ParseResult with one image per rendered page.
+        """
+        with DOCUMENT_PARSE_DURATION.labels(file_type=file_type).time():
+            try:
+                from library.services.svg_raster import render_svg_pages
+
+                with open(file_path, "rb") as f:
+                    data = f.read()
+
+                pages = render_svg_pages(data)
+            except Exception as exc:
+                DOCUMENTS_PARSED_TOTAL.labels(file_type=file_type, status="error").inc()
+                logger.error("Failed to rasterize SVG file_type=%s: %s", file_type, exc)
+                raise
+
+            images = [
+                ExtractedImage(
+                    data=png,
+                    ext="png",
+                    width=width,
+                    height=height,
+                    source_page=index,
+                    source_index=0,
+                )
+                for index, (png, width, height) in enumerate(pages)
+            ]
+
+            DOCUMENTS_PARSED_TOTAL.labels(file_type=file_type, status="success").inc()
+            IMAGES_EXTRACTED_TOTAL.labels(file_type=file_type).inc(len(images))
+
+            logger.info(
+                "Parsed SVG file_type=%s pages=%d bytes=%d",
+                file_type,
+                len(images),
+                len(data),
+            )
+
+            return ParseResult(
+                text_blocks=[],
+                images=images,
+                metadata={"page_count": len(images)},
+                file_type=file_type,
+            )
+
     def _parse_image_file(self, file_path: str, file_type: str) -> ParseResult:
         """
         Handle a standalone image file — store as a single ExtractedImage.

mnemosyne/library/services/svg_raster.py

@@ -0,0 +1,263 @@
+"""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

mnemosyne/library/tests/test_svg_raster.py

@@ -0,0 +1,195 @@
+"""
+Tests for SVG rasterization in the ingest pipeline.
+
+SVG is vector XML: Pillow cannot decode it and the vision stage cannot send it
+as a data URI, so it is rendered to PNG at parse time. The cases that matter
+are the ones that fail *silently* — Write notes stack their pages inside one
+document, and rendering that document whole yields a plausible-looking PNG that
+is either blank or an illegible tall strip.
+"""
+
+import io
+import os
+import tempfile
+
+from django.test import TestCase
+from PIL import Image
+
+from library.services.parsers import (
+    IMAGE_EXTENSIONS,
+    DocumentParser,
+)
+from library.services.svg_raster import (
+    SvgRenderError,
+    render_svg_pages,
+    split_svg_pages,
+)
+
+# Geometry copied from a real Write note; the page element is identical across
+# both root formats, which is what makes splitting format-agnostic.
+PAGE_WIDTH = 1094
+PAGE_HEIGHT = 1654
+PAGE_PITCH = 1674  # page height + inter-page gap
+
+
+def write_document(pages: int, root_size: bool = False) -> bytes:
+    """
+    Build a Write-style document.
+
+    :param pages: Number of stacked pages.
+    :param root_size: Emit width/height on the root <svg>. False models
+        pre-eeab021 files (the permanent majority), True models newer saves.
+    """
+    root_attrs = ""
+    if root_size:
+        total = 10 + pages * PAGE_PITCH
+        root_attrs = f' width="{PAGE_WIDTH + 20}" height="{total}"'
+
+    body = "".join(
+        f'<svg class="write-page" x="10" y="{10 + i * PAGE_PITCH}" '
+        f'width="{PAGE_WIDTH}px" height="{PAGE_HEIGHT}px" '
+        f'xmlns="http://www.w3.org/2000/svg">'
+        f'<path d="M 100 {100 + i * 40} L {600 + i * 120} {700 + i * 40}" '
+        f'stroke="#000000" stroke-width="12" fill="none"/></svg>'
+        for i in range(pages)
+    )
+    return (
+        f'<svg id="write-document"{root_attrs} '
+        f'xmlns="http://www.w3.org/2000/svg" '
+        f'xmlns:xlink="http://www.w3.org/1999/xlink">'
+        f'<rect id="write-doc-background" width="100%" height="100%" fill="#808080"/>'
+        f'<defs id="write-defs"><style>.write-flat-pen{{fill:none}}</style></defs>'
+        f"{body}</svg>"
+    ).encode()
+
+
+GENERIC_SVG = (
+    b'<svg xmlns="http://www.w3.org/2000/svg" width="400" height="200" '
+    b'viewBox="0 0 400 200">'
+    b'<path d="M 20 20 L 380 180" stroke="#000" stroke-width="10"/></svg>'
+)
+
+
+def ink_fraction(png: bytes) -> float:
+    """Fraction of non-white pixels — the blank-render canary."""
+    grey = Image.open(io.BytesIO(png)).convert("L")
+    histogram = grey.histogram()
+    return sum(histogram[:200]) / sum(histogram)
+
+
+class SvgSplitTests(TestCase):
+    """Splitting a Write document into per-page SVGs."""
+
+    def test_splits_one_document_per_page(self):
+        for pages in (1, 2, 5):
+            for root_size in (False, True):
+                with self.subTest(pages=pages, root_size=root_size):
+                    document = write_document(pages, root_size=root_size)
+                    self.assertEqual(len(split_svg_pages(document)), pages)
+
+    def test_both_root_formats_split_identically(self):
+        """Root width/height (Write eeab021) must not change the outcome.
+
+        Existing notes are never rewritten, so both formats persist
+        indefinitely and have to render the same.
+        """
+        without = split_svg_pages(write_document(3, root_size=False))
+        with_size = split_svg_pages(write_document(3, root_size=True))
+        self.assertEqual(len(without), len(with_size))
+        self.assertEqual(len(without), 3)
+
+    def test_generic_svg_is_a_single_page(self):
+        self.assertEqual(len(split_svg_pages(GENERIC_SVG)), 1)
+
+    def test_unsized_svg_is_rejected_rather_than_guessed(self):
+        unsized = b'<svg xmlns="http://www.w3.org/2000/svg"><rect/></svg>'
+        with self.assertRaises(SvgRenderError):
+            split_svg_pages(unsized)
+
+    def test_script_inside_defs_does_not_desync_the_split(self):
+        """Regression: a non-greedy <defs>...</defs> regex mis-parses these."""
+        document = write_document(2).replace(
+            b'<defs id="write-defs">',
+            b'<defs id="write-defs"><script><float value="770" /></script>',
+        )
+        self.assertEqual(len(split_svg_pages(document)), 2)
+
+    def test_unquoted_root_attributes_are_repaired(self):
+        """Write can emit ``width=auto``, which is not valid XML.
+
+        A malformation inside the root tag defeats recovery differently from
+        one in the body: libxml2 abandons the whole document and returns a
+        bare root, so every page silently disappears.
+        """
+        document = write_document(2).replace(
+            b'<svg id="write-document"',
+            b'<svg width=auto height=auto id="write-document"',
+        )
+        self.assertEqual(len(split_svg_pages(document)), 2)
+
+
+class SvgRenderTests(TestCase):
+    """Rasterizing pages to PNG."""
+
+    def test_renders_one_png_per_page(self):
+        pages = render_svg_pages(write_document(3))
+        self.assertEqual(len(pages), 3)
+        for png, width, height in pages:
+            self.assertEqual(Image.open(io.BytesIO(png)).format, "PNG")
+            self.assertEqual(max(width, height), 1568)
+
+    def test_rendered_pages_are_not_blank(self):
+        """The pre-patch trap: a whole-document render emits ruling, no ink."""
+        for root_size in (False, True):
+            with self.subTest(root_size=root_size):
+                for png, _, _ in render_svg_pages(
+                    write_document(3, root_size=root_size)
+                ):
+                    self.assertGreater(ink_fraction(png), 0.001)
+
+    def test_pages_are_page_shaped_not_a_stacked_strip(self):
+        """The post-patch trap: root dimensions span every stacked page.
+
+        Rendering that whole gives one tall strip which, once capped, squashes
+        a 5-page note to ~209px wide.
+        """
+        expected = PAGE_HEIGHT / PAGE_WIDTH
+        for root_size in (False, True):
+            with self.subTest(root_size=root_size):
+                for _, width, height in render_svg_pages(
+                    write_document(5, root_size=root_size)
+                ):
+                    self.assertAlmostEqual(height / width, expected, delta=0.05)
+
+    def test_page_cap_is_respected(self):
+        self.assertEqual(len(render_svg_pages(write_document(8), max_pages=5)), 5)
+
+
+class SvgParserIntegrationTests(TestCase):
+    """The parser dispatch — SVG must not reach the Pillow image path."""
+
+    def setUp(self):
+        self.parser = DocumentParser()
+
+    def test_svg_is_not_in_image_extensions(self):
+        # It was, and PIL.Image.open cannot decode SVG, so ingest always failed.
+        self.assertNotIn("svg", IMAGE_EXTENSIONS)
+
+    def test_parse_multipage_svg_yields_one_image_per_page(self):
+        with tempfile.NamedTemporaryFile(suffix=".svg", delete=False) as f:
+            f.write(write_document(3))
+            f.flush()
+            path = f.name
+        try:
+            result = self.parser.parse(path, "svg")
+        finally:
+            os.unlink(path)
+
+        self.assertEqual(len(result.images), 3)
+        self.assertEqual(result.metadata["page_count"], 3)
+        self.assertEqual(result.text_blocks, [])
+        for index, image in enumerate(result.images):
+            # PNG, not svg — the vision stage needs raster for its data URI.
+            self.assertEqual(image.ext, "png")
+            self.assertEqual(image.source_page, index)
+            self.assertGreater(ink_fraction(image.data), 0.001)

pyproject.toml

@@ -30,6 +30,9 @@ dependencies = [
     "semantic-text-splitter>=0.20,<1.0",
     "tokenizers>=0.20,<1.0",
     "Pillow>=10.0,<12.0",
+    # SVG page splitting — needs recover=True for Write notes that emit
+    # unescaped attribute content and aren't well-formed XML
+    "lxml>=5.3,<7",
     "requests>=2.31,<3.0",
     # Phase 5: MCP Server
     "fastmcp>=2.0,<3.0",