0.24.0: drop PyMuPDF and xgen-contextifier — xgen-pdf + vendored raw layer

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "xgen-edit2docs"
-version = "0.23.0"
+version = "0.24.0"
 description = "AI-agent-native document engine: generate and chat-edit DOCX, XLSX and PPTX as a Python library, agent tool set, MCP server or hosted service. English-first with first-class Korean support. Sister project of edit2ppt."
 readme = "README.md"
 requires-python = ">=3.12"
@@ -31,24 +31,21 @@ dependencies = [
     # Core engine dependencies (from ppt-master)
     "python-pptx>=0.6.21",
     "python-docx>=1.1.0",
-    "PyMuPDF>=1.23.0",
+    # PDF engine: pdfium-based xgen-pdf (BSD/Apache). PyMuPDF (AGPL) is gone.
+    "xgen-pdf @ https://github.com/PlateerLab/xgen-pdf/releases/download/v0.1.1/xgen_pdf-0.1.1-py3-none-any.whl",
     "mammoth>=1.6.0",
     "openpyxl>=3.1.0",
-    # Lossless raw OOXML layer (xgen_contextifier.open_raw): surgical xlsx/docx
-    # edits and PPTX native-content preservation that keep charts,
-    # sparklines, custom XML, cached formula values and native
-    # charts/tables byte-identical through an edit.
-    # ⚠ 내부 패키지는 **URL 직접참조**로 고정한다. 우리 패키지는 어느 인덱스에도
-    # 없으므로 이름으로만 요구하면 `pip install -e .` 이 해석에 실패한다 —
-    # 실제로 이 저장소는 개발 설치조차 되지 않는 상태였다.
-    "xgen-contextifier @ https://github.com/PlateerLab/xgen-contextifier/releases/download/v0.9.0/xgen_contextifier-0.9.0-py3-none-any.whl",
+    # Lossless raw OOXML layer: vendored in xgen_edit2docs.raw (from xgen-contextifier 0.9.0),
+    # so no dependency on that package (which would pull in PyMuPDF).
     # Native chart export (data-pptx-native markers) generates the chart's
     # embedded XLSX workbook with XlsxWriter (upstream 53c6cad5).
     "XlsxWriter>=3.0",
     "nbconvert>=7.0.0",
     "markdownify>=0.11.6",
     "ebooklib>=0.18",
     "beautifulsoup4>=4.12.0",
+    # HWP/OLE containers (was an indirect dependency via xgen-contextifier).
+    "olefile>=0.47",
     "requests>=2.31.0",
     "Pillow>=9.0.0",
     "numpy>=1.20.0",

src/xgen_edit2docs/__init__.py

@@ -19,7 +19,7 @@
 import importlib
 from typing import Any
 
-__version__ = "0.18.0"
+__version__ = "0.24.0"
 
 _LAZY: dict[str, str] = {
     # Unified, extension-dispatched verbs (docx / xlsx / pptx)

src/xgen_edit2docs/core/docs/conversion.md

@@ -33,7 +33,7 @@ Prefer MinerU or another OCR/layout tool when:
 Dependency:
 
 ```bash
-pip install PyMuPDF
+pip install xgen-pdf
 ```
 
 ## `source_to_md/doc_to_md.py`

src/xgen_edit2docs/core/docs/troubleshooting.md

@@ -76,5 +76,5 @@ Important optional packages:
 - `edge-tts` for `notes_to_audio.py` recorded narration audio
 - `Pillow` for image utilities
 - `numpy` for watermark removal
-- `PyMuPDF` for PDF conversion
+- `xgen-pdf` for PDF conversion
 - `google-genai` / `openai` for image generation backends

src/xgen_edit2docs/core/source_to_md/pdf_to_md.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 """
 PDF to Markdown Converter
-Uses PyMuPDF to extract PDF text content and convert to Markdown format.
+Uses xgen-pdf (pdfium) to extract PDF text content and convert to Markdown format.
 Supports heading levels, bold, italic, and list detection.
 """
 
@@ -14,9 +14,9 @@
 from collections import Counter
 
 try:
-    import fitz  # PyMuPDF
+    import xgen_pdf as fitz  # pdfium-based engine (xgen-pdf)
 except ImportError:
-    print("[ERROR] PyMuPDF not installed. Run: pip install PyMuPDF", file=sys.stderr)
+    print("[ERROR] xgen-pdf not installed. Run: pip install xgen-pdf", file=sys.stderr)
     sys.exit(1)
 
 FONT_BODY_SIZE = 12
@@ -356,7 +356,7 @@ def should_keep_image(
     """Filter out small, decorative, or duplicate images.
 
     Args:
-        block: Image block extracted from PyMuPDF.
+        block: Image block from ``page.get_text("dict")``.
         page_rect: Current page rectangle.
         seen_hashes: Optional set used to deduplicate image payloads.
 
@@ -618,7 +618,7 @@ def clean_text(text: str) -> str:
 def merge_adjacent_formatting(text: str) -> str:
     """Merge adjacent same-style formatted spans split across PDF tokens.
 
-    PyMuPDF often emits a phrase as several spans, so per-span wrapping in
+    The text engine may emit a phrase as several spans, so per-span wrapping in
     ``format_span_text`` produces ``**X****Y**`` (bold) or ``***X******Y***``
     (bold-italic) where one phrase is intended. Collapse the abutting markers
     so the run reads as a single phrase: ``**X Y**`` / ``***X Y***``.

src/xgen_edit2docs/documents/arrange.py

@@ -42,7 +42,7 @@ def apply_arrange(
     per op ``{op, target, to?, name?, status, message}``; ``warnings`` is
     a list of ``{code, message}`` (e.g. a rename that leaves formula
     references dangling)."""
-    from xgen_contextifier import open_raw
+    from xgen_edit2docs.raw import open_raw
 
     results: list[dict] = []
     warnings: list[dict] = []

src/xgen_edit2docs/documents/chart_edit.py

@@ -73,7 +73,7 @@ def list_charts(content: bytes, fmt: str) -> list[dict]:
     the address source for :func:`apply_chart_edits`.
     """
     try:
-        from xgen_contextifier import open_raw
+        from xgen_edit2docs.raw import open_raw
 
         raw = open_raw(content, extension=fmt)
         out: list[dict] = []
@@ -110,8 +110,8 @@ def apply_chart_edits(
     deterministic editors. The package is only re-serialized when at least
     one edit applied; untouched parts stay byte-identical.
     """
-    from xgen_contextifier import open_raw
-    from xgen_contextifier.raw.opc import RawUnsupportedError
+    from xgen_edit2docs.raw import open_raw
+    from xgen_edit2docs.raw.opc import RawUnsupportedError
 
     raw = open_raw(content, extension=fmt)
     charts = _charts_of(raw, fmt)

src/xgen_edit2docs/documents/docx_engine.py

@@ -318,7 +318,7 @@ def _chart_outline(content: bytes) -> list[dict]:
     """Read-only chart summaries via xgen_contextifier (best-effort: outline
     must never fail because a chart part is exotic)."""
     try:
-        from xgen_contextifier import open_raw
+        from xgen_edit2docs.raw import open_raw
 
         raw = open_raw(content, extension="docx")
         return [
@@ -379,7 +379,7 @@ def apply_docx_edits(content: bytes, edits: Iterable[DocxEdit]) -> tuple[bytes,
     Per-edit soft failures (like the PPTX text editor): ``old_text``
     guards replaces with a whitespace-normalized comparison.
     """
-    from xgen_contextifier import open_raw
+    from xgen_edit2docs.raw import open_raw
 
     try:
         raw = open_raw(content, extension="docx")
@@ -477,7 +477,7 @@ def _fragment_blocks(fragment: bytes) -> list:
     grafting this replaces)."""
     import zipfile
 
-    from xgen_contextifier.raw import qn
+    from xgen_edit2docs.raw import qn
     from lxml import etree
 
     with zipfile.ZipFile(io.BytesIO(fragment)) as zf:

src/xgen_edit2docs/documents/docx_pages.py

@@ -5,7 +5,7 @@
 line wrap (``xgen_edit2docs.render.fonts`` — the same fonts resvg
 rasterizes with), lays out tables/images/headers/footers, and emits
 one self-contained SVG per page. ``render_doc`` feeds these to the
-resvg/PyMuPDF raster layer for PNG/PDF — the piece LibreOffice used to
+resvg/xgen-pdf raster layer for PNG/PDF — the piece LibreOffice used to
 provide.
 
 Fidelity scope (deliberate): body paragraphs (runs with bold/italic/

src/xgen_edit2docs/documents/xlsx_engine.py

@@ -187,7 +187,7 @@ def _chart_outline(content: bytes) -> list[dict]:
     """Read-only chart summaries via xgen_contextifier (best-effort: outline
     must never fail because a chart part is exotic)."""
     try:
-        from xgen_contextifier import open_raw
+        from xgen_edit2docs.raw import open_raw
 
         raw = open_raw(content, extension="xlsx")
         return [
@@ -266,7 +266,7 @@ def apply_xlsx_edits(content: bytes, edits: Iterable[XlsxEdit]) -> tuple[bytes,
     formula values all survive byte-identical (the old openpyxl
     load→save round-trip destroyed every one of those on EVERY edit).
     """
-    from xgen_contextifier import open_raw
+    from xgen_edit2docs.raw import open_raw
 
     try:
         raw = open_raw(content, extension="xlsx")
@@ -369,7 +369,7 @@ def _add_raw_sheet(raw, title: str) -> None:
     rels entry, and the ``<sheet>`` element in ``xl/workbook.xml``.
     Everything else in the package stays byte-identical.
     """
-    from xgen_contextifier.raw import qn
+    from xgen_edit2docs.raw import qn
 
     package = raw.package
     n = 1

src/xgen_edit2docs/documents/xml_edit.py

@@ -51,7 +51,7 @@ class XmlEditResult:
 
 
 def _open_package(content: bytes):
-    from xgen_contextifier.raw.opc import OpcPackage
+    from xgen_edit2docs.raw.opc import OpcPackage
 
     return OpcPackage.open(content)
 

src/xgen_edit2docs/raw/__init__.py

@@ -0,0 +1,93 @@
+# xgen_contextifier/raw
+"""
+Raw document access — the lossless twin of the extraction pipeline.
+
+Contextifier has two ways to look at a document:
+
+* ``DocumentProcessor.extract_text()`` / ``.process()`` — the existing
+  pipeline that renders an **AI-friendly** view (clean text, normalized
+  tables/charts) and throws the rest away.
+* ``open_raw()`` (this package) — a **lossless, addressable, writable**
+  view of the same file. Nothing is discarded: every OPC part stays
+  available, XML is parsed lazily, and edits are *surgical* — when you
+  save, untouched parts are written back **byte-identical** (the
+  byte-preservation contract), so charts, pivot tables, sparklines,
+  custom XML, styles and anything else the higher-level libraries can't
+  model all survive.
+
+Usage::
+
+    from xgen_edit2docs.raw import open_raw
+
+    raw = open_raw("report.xlsx")          # XlsxRawDocument
+    raw.sheets["Sales"].set_cell("B3", 142)
+    raw.charts[0].set_data(categories=["Q1", "Q2"], series=[("Sales", [1, 2])])
+    raw.save("report-edited.xlsx")         # or raw.to_bytes()
+
+    raw = open_raw("deck.pptx")            # PptxRawDocument
+    raw = open_raw("paper.docx")           # DocxRawDocument
+
+Every format model also exposes ``.package`` (the raw
+:class:`~xgen_edit2docs.raw.opc.OpcPackage`) for part-level work, so the
+"easy" interface never locks you out of the full container.
+
+Supported today: the OOXML trio (.xlsx / .docx / .pptx). Other handlers
+raise :class:`RawUnsupportedError`.
+"""
+
+from __future__ import annotations
+
+from xgen_edit2docs.raw.opc import OpcPackage, OpcPart, RawUnsupportedError
+from xgen_edit2docs.raw.xmlpart import NS, XmlPart, qn
+
+__all__ = [
+    "OpcPackage",
+    "OpcPart",
+    "RawUnsupportedError",
+    "XmlPart",
+    "NS",
+    "qn",
+    "open_raw",
+]
+
+
+def open_raw(source, *, extension: str | None = None):
+    """Open a document for lossless, writable access.
+
+    Args:
+        source: path (str/Path), bytes, or a binary file object.
+        extension: override the format sniff (e.g. ``"xlsx"``); by default
+            the file extension (for paths) or the package content is used.
+
+    Returns:
+        ``XlsxRawDocument`` / ``DocxRawDocument`` / ``PptxRawDocument``.
+
+    Raises:
+        RawUnsupportedError: format has no raw model yet.
+    """
+    from pathlib import Path
+
+    ext = (extension or "").lower().lstrip(".")
+    if not ext and isinstance(source, (str, Path)):
+        ext = Path(source).suffix.lower().lstrip(".")
+
+    package = OpcPackage.open(source)
+    if not ext:
+        ext = package.sniff_format() or ""
+
+    if ext == "xlsx":
+        from xgen_edit2docs.raw.xlsx import XlsxRawDocument
+
+        return XlsxRawDocument(package)
+    if ext == "docx":
+        from xgen_edit2docs.raw.docx import DocxRawDocument
+
+        return DocxRawDocument(package)
+    if ext == "pptx":
+        from xgen_edit2docs.raw.pptx import PptxRawDocument
+
+        return PptxRawDocument(package)
+    raise RawUnsupportedError(
+        f"No raw model for {ext or 'unknown format'!r} yet "
+        "(supported: xlsx, docx, pptx)"
+    )