When using langchain-community, some PDF images will report errors during OCR

[wwbrave002]

Checked other resources

• I added a very descriptive title to this issue.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).

Example Code

class PyPDFParser(BaseBlobParser):
"""Load PDF using pypdf"""

def __init__(
    self, password: Optional[Union[str, bytes]] = None, extract_images: bool = False
):
    self.password = password
    self.extract_images = extract_images

def lazy_parse(self, blob: Blob) -> Iterator[Document]:  # type: ignore[valid-type]
    """Lazily parse the blob."""
    import pypdf
    
    self.pdf_blob = blob

    with blob.as_bytes_io() as pdf_file_obj:  # type: ignore[attr-defined]
        pdf_reader = pypdf.PdfReader(pdf_file_obj, password=self.password)
        yield from [
            Document(
                page_content=page.extract_text()
                + self._extract_images_from_page(page),
                metadata={"source": blob.source, "page": page_number},  # type: ignore[attr-defined]
            )
            for page_number, page in enumerate(pdf_reader.pages)
        ]

def _extract_images_from_page(self, page: pypdf._page.PageObject) -> str:
    """Extract images from page and get the text with RapidOCR."""
    if not self.extract_images or "/XObject" not in page["/Resources"].keys():
        return ""

    xObject = page["/Resources"]["/XObject"].get_object()  # type: ignore
    images = []
    for obj in xObject:
        # print(f"obj: {xObject[obj]}")
        if xObject[obj]["/Subtype"] == "/Image":
            if xObject[obj].get("/Filter"):
                if isinstance(xObject[obj]["/Filter"], str):
                    if xObject[obj]["/Filter"][1:] in _PDF_FILTER_WITHOUT_LOSS:
                        height, width = xObject[obj]["/Height"], xObject[obj]["/Width"]
                        # print(xObject[obj].get_data())
                        try:
                            images.append(
                                np.frombuffer(xObject[obj].get_data(), dtype=np.uint8).reshape(
                                    height, width, -1
                                )
                            )
                        except Exception as e:
                            if xObject[obj]["/Filter"][1:] == "CCITTFaxDecode":
                                import fitz
                                with self.pdf_blob.as_bytes_io() as pdf_file_obj:  # type: ignore[attr-defined]
                                    with fitz.open("pdf", pdf_file_obj.read()) as doc:
                                        pix = doc.load_page(page.page_number).get_pixmap(matrix=fitz.Matrix(1,1), colorspace=fitz.csGRAY)
                                        images.append(pix.tobytes())
                            else:
                                warnings.warn(f"Reshape Error: {xObject[obj]}")
                    elif xObject[obj]["/Filter"][1:] in _PDF_FILTER_WITH_LOSS:
                        images.append(xObject[obj].get_data())
                    else:
                        warnings.warn(f"Unknown PDF Filter: {xObject[obj]["/Filter"][1:]}")
                elif isinstance(xObject[obj]["/Filter"], list):
                    for xObject_filter in xObject[obj]["/Filter"]:
                        if xObject_filter[1:] in _PDF_FILTER_WITHOUT_LOSS:
                            height, width = xObject[obj]["/Height"], xObject[obj]["/Width"]
                            # print(xObject[obj].get_data())
                            try:
                                images.append(
                                    np.frombuffer(xObject[obj].get_data(), dtype=np.uint8).reshape(
                                        height, width, -1
                                    )
                                )
                            except Exception as e:
                                if xObject[obj]["/Filter"][1:] == "CCITTFaxDecode":
                                    import fitz
                                    with self.pdf_blob.as_bytes_io() as pdf_file_obj:  # type: ignore[attr-defined]
                                        with fitz.open("pdf", pdf_file_obj.read()) as doc:
                                            pix = doc.load_page(page.number).get_pixmap(matrix=fitz.Matrix(1,1), colorspace=fitz.csGRAY)
                                            images.append(pix.tobytes())
                                else:
                                    warnings.warn(f"Reshape Error: {xObject[obj]}")
                            break
                        elif xObject_filter[1:] in _PDF_FILTER_WITH_LOSS:
                            images.append(xObject[obj].get_data())
                            break
                        else:
                            warnings.warn(f"Unknown PDF Filter: {xObject_filter[1:]}")
            else:
                warnings.warn("Can Not Find PDF Filter!")
    return extract_from_images_with_rapidocr(images)

Error Message and Stack Trace (if applicable)

No response

Description

When I use langchain-community, some PDF images will report errors during OCR. I tried to add some processing based on the source code PyPDFParser class, which temporarily solved the problem. Administrators can check whether to add this part of code in the new version. The complete PyPDFParser class is shown in Example Code.

System Info

langchain-community==0.2.4