[Question]: I want to get the coordinates using Unstructured as the metadata of nodes

[Hritesh-Hexaware]

Question Validation

• [X] I have searched both the documentation and discord for an answer.

Question

 dir_reader = SimpleDirectoryReader("./data-from-azure", file_extractor={
            ".pdf": UnstructuredReader()
        })

documents = dir_reader.load_data(show_progress=True)

Settings.llm = llm

Settings.embed_model = embed_model

parser = SemanticSplitterNodeParser(
    embed_model=embed_model,
    breakpoint_percentile_threshold=95,
    buffer_size=1,
)

semantic_nodes = parser.build_semantic_nodes_from_documents(documents)

print(semantic_nodes)

I want the coordinates as well with the nodes but the default implementation is skipping the coordinates in llama_index.readers.file.unstructured.base.py file

 if field == "coordinates":
     continue

Is there any way we can get the coordinates as well ?

[dosubot[bot]]

Hey @Hritesh-Hexaware! I'm here to help you with any issues or questions you have. Feel free to ask anything related to the repository.

Yes, you can modify the implementation in llama_index.readers.file.unstructured.base.py to include the coordinates. You need to remove the condition that skips the coordinates and ensure they are included in the metadata. Here is the modified code:

"""Unstructured file reader.

A parser for unstructured text files using Unstructured.io.
Supports .txt, .docx, .pptx, .jpg, .png, .eml, .html, and .pdf documents.

"""

from pathlib import Path
from typing import Any, Dict, List, Optional

from llama_index.core.readers.base import BaseReader
from llama_index.core.schema import Document

class UnstructuredReader(BaseReader):
    """General unstructured text reader for a variety of files."""

    def __init__(self, *args: Any, **kwargs: Any) -> None:
        """Init params."""
        super().__init__(*args)  # not passing kwargs to parent bc it cannot accept it

        self.api = False  # we default to local
        if "url" in kwargs:
            self.server_url = str(kwargs["url"])
            self.api = True  # is url was set, switch to api
        else:
            self.server_url = "http://localhost:8000"

        if "api" in kwargs:
            self.api = kwargs["api"]

        self.api_key = ""
        if "api_key" in kwargs:
            self.api_key = kwargs["api_key"]

        # Prerequisite for Unstructured.io to work
        import nltk

        if not nltk.data.find("tokenizers/punkt"):
            nltk.download("punkt")
        if not nltk.data.find("taggers/averaged_perceptron_tagger"):
            nltk.download("averaged_perceptron_tagger")

    """ Loads data using Unstructured.io py

        Depending on the constructin if url is set or api = True
        it'll parse file using API call, else parse it locally
        extra_info is extended by the returned metadata if
        split_documents is True

        Returns list of documents
    """

    def load_data(
        self,
        file: Path,
        extra_info: Optional[Dict] = None,
        split_documents: Optional[bool] = False,
    ) -> List[Document]:
        """If api is set, parse through api."""
        if self.api:
            from unstructured.partition.api import partition_via_api

            elements = partition_via_api(
                filename=str(file),
                api_key=self.api_key,
                api_url=self.server_url + "/general/v0/general",
            )
        else:
            """Parse file locally"""
            from unstructured.partition.auto import partition

            elements = partition(filename=str(file))

        """ Process elements """
        docs = []
        if split_documents:
            for node in elements:
                metadata = {}
                if hasattr(node, "metadata"):
                    """Load metadata fields"""
                    for field, val in vars(node.metadata).items():
                        if field == "_known_field_names":
                            continue
                        # removing parent_id because it might cause interference
                        if field == "parent_id":
                            continue
                        metadata[field] = val

                if extra_info is not None:
                    metadata.update(extra_info)

                metadata["filename"] = str(file)
                docs.append(Document(text=node.text, extra_info=metadata))

        else:
            text_chunks = [" ".join(str(el).split()) for el in elements]

            metadata = {}

            if extra_info is not None:
                metadata.update(extra_info)

            metadata["filename"] = str(file)
            # Create a single document by joining all the texts
            docs.append(Document(text="\n\n".join(text_chunks), extra_info=metadata))

        return docs

In this modified code, the condition that skips the coordinates (if field == "coordinates": continue) has been removed, allowing the coordinates to be included in the metadata ^[1].

To continue talking to Dosu, mention @dosu.

[Hritesh-Hexaware]

@dosu i cant comment llama-index files

[dosubot[bot]]

To include coordinates in the metadata of nodes without modifying the llama_index.readers.file.unstructured.base.py file directly, you can subclass the UnstructuredReader and override the load_data method. Here is an example:

from pathlib import Path
from typing import Any, Dict, List, Optional
from llama_index.core.schema import Document
from llama_index.readers.file.unstructured.base import UnstructuredReader

class CustomUnstructuredReader(UnstructuredReader):
    def load_data(
        self,
        file: Path,
        extra_info: Optional[Dict] = None,
        split_documents: Optional[bool] = False,
    ) -> List[Document]:
        """If api is set, parse through api."""
        if self.api:
            from unstructured.partition.api import partition_via_api

            elements = partition_via_api(
                filename=str(file),
                api_key=self.api_key,
                api_url=self.server_url + "/general/v0/general",
            )
        else:
            """Parse file locally"""
            from unstructured.partition.auto import partition

            elements = partition(filename=str(file))

        docs = []
        if split_documents:
            for node in elements:
                metadata = {}
                if hasattr(node, "metadata"):
                    """Load metadata fields"""
                    for field, val in vars(node.metadata).items():
                        if field == "_known_field_names":
                            continue
                        # Include coordinates in the metadata
                        metadata[field] = val

                if extra_info is not None:
                    metadata.update(extra_info)

                metadata["filename"] = str(file)
                docs.append(Document(text=node.text, extra_info=metadata))

        else:
            text_chunks = [" ".join(str(el).split()) for el in elements]

            metadata = {}

            if extra_info is not None:
                metadata.update(extra_info)

            metadata["filename"] = str(file)
            # Create a single document by joining all the texts
            docs.append(Document(text="\n\n".join(text_chunks), extra_info=metadata))

        return docs

By using this subclass, you can include the coordinates in the metadata without modifying the original UnstructuredReader class ^[1].

To continue talking to Dosu, mention @dosu.