extract-python 0.11.3__tar.gz → 0.12.0__tar.gz

extract_python-0.12.0/.python-version

@@ -0,0 +1 @@
+3.13

{extract_python-0.11.3 → extract_python-0.12.0}/PKG-INFO

@@ -1,13 +1,13 @@
 Metadata-Version: 2.5
 Name: extract-python
-Version: 0.11.3
+Version: 0.12.0
 Summary: Structured content extraction
 Project-URL: Homepage, https://github.com/ICIJ/extract-python
 Project-URL: Repository, https://github.com/ICIJ/extract-python
 Project-URL: Issues, https://github.com/ICIJ/extract-python/issues
 Author-email: Clément Doumouro <cdoumouro@icij.org>
 Requires-Python: <3.15,>=3.13
-Requires-Dist: extract-core~=0.11.3
+Requires-Dist: extract-core~=0.12.1
 Requires-Dist: icij-common~=0.8.2
 Requires-Dist: pympler~=1.1
 Provides-Extra: benches
@@ -21,6 +21,11 @@ Provides-Extra: docling
 Requires-Dist: docling-slim[feat-ocr-easyocr,feat-ocr-mac,feat-ocr-rapidocr,feat-ocr-tesserocr,standard]~=2.127; extra == 'docling'
 Provides-Extra: docling-gpu
 Requires-Dist: docling-slim[feat-ocr-nemotron,models-onnxruntime,models-vlm-inline]~=2.127; extra == 'docling-gpu'
+Provides-Extra: liteparse
+Requires-Dist: liteparse~=2.15; extra == 'liteparse'
+Requires-Dist: pillow~=11.0; extra == 'liteparse'
+Requires-Dist: pypdfium2~=4.0; extra == 'liteparse'
+Requires-Dist: tesserocr~=2.7; extra == 'liteparse'
 Provides-Extra: marker
 Requires-Dist: marker-pdf~=1.10; extra == 'marker'
 Provides-Extra: mineru

{extract_python-0.11.3 → extract_python-0.12.0}/extract_python/__init__.py

@@ -16,15 +16,21 @@ except ModuleNotFoundError:
 
 
 try:
-    from .miner_u import MinerUPipeline
+    from .mineru_ import MinerUPipeline
 except ModuleNotFoundError:
     logger.warning("mineru is not available, might be optional...")
     MinerUPipeline = None
 
+try:
+    from .liteparse_ import LiteParsePipeline
+except ModuleNotFoundError:
+    logger.warning("liteparse is not available, might be optional...")
+    LiteParsePipeline = None
 
 __all__ = [
     "DoclingPipeline",
     "DOCLING_DEFAULT_ARTIFACTS_PATH",
     "MarkerPipeline",
     "MinerUPipeline",
+    "LiteParsePipeline",
 ]

{extract_python-0.11.3 → extract_python-0.12.0}/extract_python/docling_.py

@@ -101,8 +101,8 @@ class DoclingPipeline(Pipeline):
                     if res is sentinel:
                         break
                     pages = next(processed)
-                    buffer.add(pages, res)
-                    if buffer.is_complete(pages.doc_idx):
+                    is_complete = buffer.add(pages, res)
+                    if is_complete:
                         doc_pages = buffer.pop_complete(pages.doc_idx)
                         yield _to_result(
                             doc_pages, pages.doc, output_format, output_path=output_path

extract_python-0.12.0/extract_python/liteparse_.py

@@ -0,0 +1,493 @@
+import asyncio
+import logging
+import os
+from collections import defaultdict
+from collections.abc import AsyncIterable, Callable, Generator, Iterable, Sequence
+from contextlib import contextmanager
+from dataclasses import dataclass, field
+from functools import cache
+from itertools import batched
+from pathlib import Path
+from typing import TYPE_CHECKING, Literal, Self, cast
+
+from extract_core import (
+    ConversionOutput,
+    Error,
+    InputDoc,
+    OutputFormat,
+    Pipeline,
+    PipelineType,
+    Result,
+    Status,
+    SupportedExt,
+)
+from pydantic import TypeAdapter
+
+from .constants import ARTIFACTS, DEFAULT_MD_PAGE_SEP
+from .utils import (
+    ProcessedPages,
+    Range,
+    ResultBuffer,
+    UsageTracker,
+    chunk_docs,
+    path_to_artifacts_dirname,
+    publish_and_consume,
+    write_pages,
+)
+
+logger = logging.getLogger(__name__)
+if TYPE_CHECKING:
+    from liteparse import LiteParse
+    from liteparse.types import (
+        ExtractedImage,
+        PageComplexityStats,
+        PageError,
+        ParsedPage,
+        ParseResult,
+    )
+    from PIL import Image
+    from tesserocr import PyTessBaseAPI
+
+_LITEPARSE_CONVERSION_ERRORS = tuple()
+
+
+OrientedBatch = tuple[int | None, tuple[ProcessedPages]]
+LiteParseOutputFormat = Literal["text", "json", "markdown"]
+
+
+@cache
+def get_parse_result_ta() -> TypeAdapter:
+    from liteparse.types import ParseResult  # noqa: PLC0415
+
+    return TypeAdapter(ParseResult)
+
+
+@dataclass
+class LiteParsePageResult:
+    pages: list["ParsedPage"] = field(default_factory=list)
+    images: list["ExtractedImage"] = field(default_factory=list)
+    errors: list["PageError"] = field(default_factory=list)
+    _by_ix: (
+        dict[int, tuple["ParsedPage", list["ExtractedImage"], list["PageError"]]] | None
+    ) = None
+
+    def filter(self, page: ProcessedPages) -> Self:
+        if self._by_ix is None:
+            self._by_ix = self._results_by_index()
+        pages, images, errors = [], [], []
+        for p_i in range(page.page_range[0], page.page_range[1]):
+            p, ims, errors = self._by_ix[p_i]
+            pages.append(p)
+            images.extend(ims)
+            errors.extend(errors)
+        return LiteParsePageResult(pages=pages, images=images, errors=errors)
+
+    def _results_by_index(
+        self,
+    ) -> dict[int, tuple["ParsedPage", list["ExtractedImage"], list["PageError"]]]:
+        by_ix = dict()
+        for p in self.pages:
+            rng_ix = p.page_num - 1
+            ims = [i for i in self.images if i.page == p.page_num]
+            errors = [e for e in self.errors if e.page_num == p.page_num]
+            by_ix[rng_ix] = (p, ims, errors)
+        return by_ix
+
+
+ResultAggregator = Callable[[InputDoc, list[LiteParsePageResult], Path], Result]
+
+
+@Pipeline.register(PipelineType.LITEPARSE)
+class LiteParsePipeline(Pipeline):
+    async def extract_content(
+        self, docs: Iterable[InputDoc], output_format: OutputFormat, output_path: Path
+    ) -> AsyncIterable[Result]:
+        tessdata_path = self._config.tessdata_path
+        if tessdata_path is None:
+            tessdata_path = os.getenv("TESSDATA_PREFIX")
+        if tessdata_path is None:
+            msg = "couldn't find tessdata path in config nor TESSDATA_PREFIX env var"
+            raise RuntimeError(msg)
+        self._tessdata_path = str(tessdata_path)
+        if self._config.detect_page_orientation:
+            logger.info("reading docs page orientation detection...")
+            batches = await self._batch_per_orientation(docs)
+            logger.info("done with page orientation detection !")
+        else:
+            batches = chunk_docs(docs, self._config.page_batch_size)
+            batches = ((None, (c,)) for c in batches)
+        async for res in self._parse(batches, output_format, output_path):
+            yield res
+
+    async def _batch_per_orientation(
+        self, docs: Iterable[InputDoc]
+    ) -> list[OrientedBatch]:
+        result_queue = asyncio.Queue()
+        publisher = self._batch_publisher(docs, result_queue)
+        consumer = self._batch_consumer(result_queue)
+        batches = [b async for b in publish_and_consume(publisher, consumer=consumer)]
+        return batches
+
+    async def _batch_publisher(
+        self, docs: Iterable[InputDoc], queue: asyncio.Queue
+    ) -> None:
+        sem = asyncio.Semaphore(self._config.n_cpus - 1)
+        logger.debug("creating page complexity parser...")
+        parser = self._make_complexity_parser()
+        logger.debug("creating page complexity parser...")
+        parser.warm_up()
+        with parser:
+            async with asyncio.TaskGroup() as tg:
+                for doc in docs:
+                    if doc.ext == SupportedExt.PDF:
+                        task = self._detect_complexity_and_queue(
+                            parser, doc, queue, sem
+                        )
+                        tg.create_task(task)
+                    else:
+                        queue.put_nowait((doc, None))
+        queue.put_nowait(None)
+
+    async def _batch_consumer(
+        self, queue: asyncio.Queue
+    ) -> AsyncIterable[OrientedBatch]:
+        import pypdfium2 as pdfium  # noqa: PLC0415
+        import tesserocr  # noqa: PLC0415
+        from liteparse.types import PageComplexityStats  # noqa: PLC0415
+
+        batch_size = self._config.page_batch_size
+        osd_reader = tesserocr.PyTessBaseAPI(
+            self._tessdata_path,
+            lang="osd",
+            oem=tesserocr.OEM.DEFAULT,
+            psm=tesserocr.PSM.OSD_ONLY,
+        )
+        pages_per_orientation = defaultdict(list)
+        doc_ix = -1
+        with osd_reader:
+            while (res := await queue.get()) is not None:
+                doc_ix += 1
+                doc, pdf_pages = res
+                if pdf_pages is None:
+                    for batch in chunk_docs([doc], batch_size):
+                        rng = _liteparse_range(batch.page_range)
+                        pages_per_orientation[(rng, None)].append(batch)
+                    continue
+                buffer = defaultdict(list)
+                try:
+                    pdf_doc = None
+                    pages = cast(list[PageComplexityStats], pdf_pages)
+                    for page in pages:
+                        # We don't use orientation for regular page, we reserve osd for
+                        # pages were OCR is required
+                        if not _should_detect_orientation(page):
+                            buffer[None].append(page)
+                            continue
+                        if pdf_doc is None:
+                            pdf_doc = pdfium.PdfDocument(doc.path)
+                        page_im = (
+                            pdf_doc[page.page_number - 1].render(scale=3.0).to_pil()
+                        )
+                        orientation = _detect_orientation(osd_reader, page_im)
+                        buffer[orientation].append(page)
+                finally:
+                    if pdf_doc is not None:
+                        pdf_doc.close()
+                for orientation, pages in buffer.items():
+                    for batch in batched(pages, batch_size, strict=False):
+                        batch = _pages_from_complexities(doc, doc_ix, *batch)  # noqa: PLW2901
+                        rng = _liteparse_range_from_pages(*batch)
+                        pages_per_orientation[(rng, orientation)].append(batch)
+        # We yield pages per orientation only here to make sure to yield
+        # (orientation, range) in the most regular way and avoid reload
+        for (_, orientation), batches in pages_per_orientation.items():
+            for batch in batches:
+                yield orientation, batch
+
+    async def _parse(
+        self,
+        batches: Iterable[OrientedBatch],
+        output_format: OutputFormat,
+        output_path: Path,
+    ) -> AsyncIterable[Result]:
+        match output_format:
+            case OutputFormat.MARKDOWN:
+                liteparse_o_fmt, postprocess_fn = "markdown", aggregate_pages_md
+            case _:
+                fmt = (
+                    f"unsupported output format for {LiteParsePipeline.__name__}"
+                    f" {output_format}"
+                )
+                raise NotImplementedError(fmt)
+        result_queue = asyncio.Queue()
+        with self._result_buffer() as buffer:
+            publisher = self._parse_publisher(batches, liteparse_o_fmt, result_queue)
+            consumer = self._parse_consumer(
+                result_queue, buffer, postprocess_fn, output_path
+            )
+            async for res in publish_and_consume(publisher, consumer=consumer):
+                yield res
+
+    async def _parse_publisher(
+        self,
+        batches: Iterable[OrientedBatch],
+        output_format: LiteParseOutputFormat,
+        queue: asyncio.Queue,
+    ) -> None:
+        parser = None
+        # Buffer partial conversion to filesystem as large docs are chunked into pieces
+        # Prevent more than n_cpu concurrent processing
+        sem = asyncio.Semaphore(self._config.n_cpus - 1)
+        # Results may return in arbitrary order, we need to wait until parser is not
+        # used any longer before
+        tracker = UsageTracker()
+        current_key = None
+        with _close_parser(parser):
+            async with asyncio.TaskGroup() as tg:
+                for orientation, batch in batches:
+                    # We avoid recreating the parser at each iteration since it creates
+                    # a worker pool, instead process in (orientation, page range) order
+                    # and cache the parser
+                    page_range = _liteparse_range_from_pages(*batch)
+                    n_pages = sum(p.n_pages for p in batch)
+                    key = (page_range, orientation)
+                    if key != current_key:
+                        if parser is not None:
+                            await tracker.wait_until_free()
+                            parser.__exit__()
+                        logger.debug("loading parser for %s", key)
+                        parser = self._make_parser(output_format, key, n_pages=n_pages)
+                        parser.__enter__().warm_up()
+                        current_key = key
+                    await sem.acquire(), await tracker.acquire()
+                    task = self._convert_and_queue(parser, batch, queue, sem, tracker)
+                    tg.create_task(task)
+        queue.put_nowait(None)
+
+    async def _parse_consumer(
+        self,
+        queue: asyncio.Queue,
+        buffer: ResultBuffer,
+        result_aggregator: ResultAggregator,
+        output_path: Path,
+    ) -> AsyncIterable[Result]:
+        while (queued := await queue.get()) is not None:
+            pages, res = queued
+            for p in pages:
+                filtered = res.filter(p)
+                buffer.add(p, filtered)
+            doc_ix = pages[0].doc_idx
+            doc = pages[0].doc
+            if buffer.is_complete(doc_ix):
+                res = result_aggregator(doc, buffer.pop_complete(doc_ix), output_path)
+                yield res
+        logger.info("done consuming !")
+
+    async def _detect_complexity_and_queue(
+        self,
+        parser: "LiteParse",
+        doc: InputDoc,
+        queue: asyncio.Queue,
+        sem: asyncio.Semaphore,
+    ) -> None:
+        pages = await asyncio.to_thread(parser.is_complex, doc.path)
+        sem.release()
+        queue.put_nowait((doc, pages))
+
+    async def _convert_and_queue(
+        self,
+        parser: "LiteParse",
+        pages: tuple[ProcessedPages],
+        queue: asyncio.Queue,
+        sem: asyncio.Semaphore,
+        tracker: UsageTracker,
+    ) -> None:
+        res = await asyncio.to_thread(parser.parse, pages[0].doc.path)
+        sem.release(), await tracker.release()
+        res = LiteParsePageResult(
+            pages=res.pages, errors=res.page_errors, images=res.images
+        )
+        queue.put_nowait((pages, res))
+
+    def _make_parser(
+        self,
+        output_format: LiteParseOutputFormat,
+        key: tuple[str, int | None],
+        n_pages: int,
+    ) -> "LiteParse":
+        from liteparse.parser import LiteParse  # noqa: PLC0415
+
+        page_batch_size = self._config.page_batch_size
+        parse_timeout = self._config.per_page_timeout
+        if parse_timeout is not None:
+            parse_timeout = parse_timeout * page_batch_size
+        target_pages, orientation = key
+        if orientation is not None:
+            orientation = [(p + 1, orientation) for p in range(n_pages)]
+        pool_size = self._config.n_cpus - 1
+        return LiteParse(
+            ocr_enabled=True,
+            extract_links=True,
+            keep_headers_footers=True,
+            extract_structure_tree=True,
+            extract_blocks=True,
+            ocr_failure_fatal=True,
+            # Corrections are per docs we would have to create 1 parser per doc
+            page_orientation_corrections=orientation,
+            pool_size=pool_size,
+            parse_timeout=parse_timeout,
+            target_pages=target_pages,
+            continue_on_page_error=True,
+            output_format=output_format,
+            extract_images=True,
+            extract_screenshots=False,
+            include_complexity=True,
+            image_mode="placeholder",
+            num_workers=1,
+            image_output_dir=None,
+            **self._config.config.model_dump(),
+        )
+
+    def _make_complexity_parser(self) -> "LiteParse":
+        from liteparse.parser import LiteParse  # noqa: PLC0415
+
+        return LiteParse(ocr_enabled=False, pool_size=self._config.n_cpus - 1)
+
+    def _result_buffer(self) -> ResultBuffer:
+
+        buffer = ResultBuffer(
+            max_size_bytes=self._config.result_buffer.max_size,
+            root=self._config.result_buffer.root,
+            save_fn=_save_conversion_result,
+            load_fn=_load_conversion_result,
+        )
+        return buffer
+
+
+def aggregate_pages_md(
+    doc: InputDoc,
+    results: Sequence[LiteParsePageResult],
+    output_path: Path,
+    sep: str = DEFAULT_MD_PAGE_SEP,
+) -> Result:
+    md_dir_name = path_to_artifacts_dirname(doc.path)
+    md_dir = output_path / md_dir_name
+    md_dir.mkdir(parents=True, exist_ok=True)
+    im_dir = md_dir / ARTIFACTS
+    im_dir.mkdir(parents=True, exist_ok=True)
+    pages, errors = [], []
+    results = sorted(results, key=lambda p: p.pages[0].page_num)
+    for res in results:
+        if res.errors:
+            errors.extend(Error.from_liteparse(err) for err in res.errors)
+        per_page_md = {p.page_num: md_from_page(p) for p in res.pages}
+        for im in res.images:
+            if im.bytes is None:
+                msg = "images were not extracted, extract_images must be set to True"
+                raise ValueError(msg)
+            im_path = im_dir / im.name
+            im_path.write_bytes(im.bytes)
+            page_md = per_page_md[im.page]
+            page_md = page_md.replace(im.name, f"{ARTIFACTS}/{im.name}")
+            per_page_md[im.page] = page_md
+        for _, p in sorted(per_page_md.items()):
+            pages.append(p)
+    if not errors:
+        status = Status.SUCCESS
+    else:
+        errors_on_all_pages = all(bool(len(p.errors)) for p in pages)
+        status = Status.FAILURE if errors_on_all_pages else Status.FAILURE
+    md_file = (md_dir / md_dir.name).with_suffix(OutputFormat.MARKDOWN)
+    with md_file.open(mode="wb") as f:
+        pages = write_pages(pages, sep, f)
+    output = ConversionOutput(
+        path=md_dir.relative_to(output_path), pages=pages, confidence=None
+    )
+    return Result(input=doc, status=status, output=output, errors=errors)
+
+
+def md_from_page(page: "ParsedPage") -> str:
+    md = page.markdown
+    if "scanned" not in page.complexity.reasons:
+        return md
+    if md.startswith("![]("):
+        md = md.split("\n")[1:]
+        if not md[0]:
+            md = md[1:]
+        md = "\n".join(md)
+    return md
+
+
+def _save_conversion_result(res: "ParseResult", path: Path) -> None:
+    path.write_bytes(get_parse_result_ta().dump_json(res))
+
+
+def _load_conversion_result(path: Path) -> "ParseResult":
+    return get_parse_result_ta().validate_json(path.read_text())
+
+
+@contextmanager
+def _close_parser(parser: "LiteParse | None") -> Generator[None]:
+    try:
+        yield
+    finally:
+        if parser is not None:
+            parser.close()
+
+
+def _liteparse_range(rng: Range) -> str:
+    # page range is a Python range, liteparse indexes from 1 and upper bound
+    # is inclusive
+    return f"{rng[0] + 1}-{rng[1]}"
+
+
+def _liteparse_range_from_pages(*pages: ProcessedPages) -> str:
+    pages = sorted(pages, key=lambda p: p.page_range[0])
+    ranges = (_liteparse_range(p.page_range) for p in pages)
+    return ",".join(ranges)
+
+
+def _pages_from_complexities(
+    doc: InputDoc, doc_ix: int, *complexities: "PageComplexityStats"
+) -> tuple[ProcessedPages]:
+    complexities = sorted(complexities, key=lambda p: p.page_number)
+    pages = []
+    start, end = None, None
+    for p in complexities:
+        if start is None:
+            start = p.page_number - 1
+            end = start + 1
+        if p.page_number != end:
+            processed = ProcessedPages(doc=doc, doc_idx=doc_ix, page_range=(start, end))
+            pages.append(processed)
+            start = p.page_number - 1
+        end = p.page_number
+    processed = ProcessedPages(doc=doc, doc_idx=doc_ix, page_range=(start, end))
+    pages.append(processed)
+    return pages
+
+
+_CLIPPED_ORIENTATIONS = [0, 90, 180, 270]
+
+
+def _should_detect_orientation(page: "PageComplexityStats") -> bool:
+    if not page.needs_ocr:
+        return False
+    reasons = set(page.reasons)
+    if "scanned" in reasons:
+        return True
+    if "garbled" in reasons:
+        return True
+    if "vector-text" in reasons:  # noqa: SIM103
+        return True
+    return False
+
+
+def _detect_orientation(osd_reader: "PyTessBaseAPI", page_im: "Image.Image") -> int:
+    osd_reader.SetImage(page_im)
+    orientation = osd_reader.DetectOrientationScript()["orient_deg"]
+    return _clipped_orientation(orientation)
+
+
+def _clipped_orientation(angle: float) -> int:
+    return min((abs(angle - o) % 360, o) for o in _CLIPPED_ORIENTATIONS)[1]