Refactoring PDF loaders: 02 PyMuPDF #29063
Conversation
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎
|
dosubot
bot
added
size:XXL
Add file_path with PurePath Add CloudBlobLoader in __init__ Replace Dict/List to dict/list
pprados
force-pushed
the
pprados/02-pymupdf
branch
from
039819c to
3beda82
Compare
|
@eyurtsev I rebase the code with master ;-) |
libs/community/langchain_community/document_loaders/parsers/pdf.py
Outdated
Show resolved
Hide resolved
libs/community/langchain_community/document_loaders/parsers/pdf.py
Outdated
Show resolved
Hide resolved
libs/community/langchain_community/document_loaders/parsers/pdf.py
Outdated
Show resolved
Hide resolved
libs/community/langchain_community/document_loaders/parsers/pdf.py
Outdated
Show resolved
Hide resolved
libs/community/langchain_community/document_loaders/parsers/pdf.py
Outdated
Show resolved
Hide resolved
| @@ -78,6 +203,192 @@ def extract_from_images_with_rapidocr( | |||
| return text | |||
|
|
|||
|
|
|||
| # Type to change the function to convert images to text. | |||
| CONVERT_IMAGE_TO_TEXT = Optional[Callable[[Iterable[np.ndarray]], Iterator[str]]] | |||
There was a problem hiding this comment.
MAJOR:
Why not use an ImageBlobParser w/ the regular Blob to Document interface. it'll allow reusing the image logic for images that do not originate from pdfs (e.g., to re-use for a web crawler)
A PDF parser doesn't would accept a parser as part of the initializer
class PDFParser(...):
def __Init__(self, ... *, ..., image_blob_parser: Optional[BlobParser] = None):
passIf the image_pdf_parser is provided, then it'll be used for OCR purposes.
There was a problem hiding this comment.
I've done it! There are now 3 ImageBlogParsers!
pprados
force-pushed
the
pprados/02-pymupdf
branch
from
0d99673 to
3fe4ec5
Compare
pprados
force-pushed
the
pprados/02-pymupdf
branch
2 times, most recently
from
4342991 to
760267b
Compare
pprados
force-pushed
the
pprados/02-pymupdf
branch
3 times, most recently
from
9fc89e0 to
d30b26d
Compare
pprados
force-pushed
the
pprados/02-pymupdf
branch
2 times, most recently
from
6765dbf to
df1d4d5
Compare
pprados
force-pushed
the
pprados/02-pymupdf
branch
from
df1d4d5 to
91234f0
Compare
dosubot
bot
added
the
🤖:docs
| all_text = _merge_text_and_extras(extras, text_from_page) | ||
|
|
||
| if not all_text: | ||
| # logger.warning( |
| ) | ||
| except ImportError: | ||
| raise ImportError( | ||
| "`rapidocr-onnxruntime` package not found, please install it with " |
There was a problem hiding this comment.
nit: incorrect import error
| logger.debug("Image text: %s", content.replace("\n", "\\n")) | ||
| yield Document( | ||
| page_content=content, | ||
| metadata={"source": blob.source}, |
There was a problem hiding this comment.
Propagate blob metadata as well?
| ) | ||
|
|
||
|
|
||
| class MultimodalBlobParser(ImageBlobParser): |
There was a problem hiding this comment.
LLMImageBlobParser or something like that?
Goal is to communicate that this is being done by a multi modal llm
| def __init__( | ||
| self, | ||
| *, | ||
| format: Literal["text", "markdown", "html"] = "text", |
There was a problem hiding this comment.
This format is fairly surprising to see as part of the API -- but I think I'm OK with it.
| logger = logging.getLogger(__name__) | ||
|
|
||
|
|
||
| class ImageBlobParser(BaseBlobParser): |
There was a problem hiding this comment.
Could we rename to BaseImageBlobParser or mark as private so it's clear that i's abstract
| return pytesseract.image_to_string(img, lang="+".join(self.langs)).strip() | ||
|
|
||
|
|
||
| _prompt_images_to_description = ( |
There was a problem hiding this comment.
nit: style to follow google conventions a bit more closely
| _prompt_images_to_description = ( | |
| _PROMPT_IMAGES_TO_DESCRIPTION = ( |
| @@ -0,0 +1,149 @@ | |||
| import base64 | |||
There was a problem hiding this comment.
I think this makes sense -- would you be willing to do a documentation pass for the API reference for this file? you could push the entire code through chat gpt and ask for google style doc-strings. It'll probably do a reasonable
| Insert image, if possible, between two paragraphs. | ||
| In this way, a paragraph can be continued on the next page. | ||
| """ | ||
| parser = self.parser |
There was a problem hiding this comment.
the change in load in this file is still a breaking change .. see comment on the left hand side of the PR review
Refactoring PDF loaders step 2: "community: Refactoring PDF loaders to standardize approaches"
Description: Update PyMuPDFParser/Loader
Twitter handle: pprados
This is one part of a larger Pull Request (PR) that is too large to be submitted all at once.
This specific part focuses to prepare the update of all parsers.
For more details, see PR 28970.
@eyurtsev it's the continuation of PDFLoader modifications.