Add UnstructuredReader with support for various legacy files (.doc, .xls) (#99)

knowledgehub/loaders/__init__.py

@@ -2,6 +2,7 @@ from .base import AutoReader, DirectoryReader
 from .excel_loader import PandasExcelReader
 from .mathpix_loader import MathpixPDFReader
 from .ocr_loader import OCRReader
+from .unstructured_loader import UnstructuredReader
 
 __all__ = [
     "AutoReader",
@@ -9,4 +10,5 @@ __all__ = [
     "MathpixPDFReader",
     "OCRReader",
     "DirectoryReader",
+    "UnstructuredReader",
 ]

knowledgehub/loaders/unstructured_loader.py

@@ -0,0 +1,110 @@
+"""Unstructured file reader.
+
+A parser for unstructured text files using Unstructured.io.
+Supports .txt, .docx, .pptx, .jpg, .png, .eml, .html, and .pdf documents.
+
+To use .doc and .xls parser, install
+
+sudo apt-get install -y libmagic-dev poppler-utils libreoffice
+pip install xlrd
+
+"""
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from llama_index.readers.base import BaseReader
+
+from kotaemon.base 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"]
+
+    """ Loads data using Unstructured.io
+
+        Depending on the construction if url is set or api = True
+        it'll parse file using API call, else parse it locally
+        additional_metadata is extended by the returned metadata if
+        split_documents is True
+
+        Returns list of documents
+    """
+
+    def load_data(
+        self,
+        file: Path,
+        additional_metadata: Optional[Dict] = None,
+        split_documents: Optional[bool] = False,
+        **kwargs,
+    ) -> List[Document]:
+        """If api is set, parse through api"""
+        file_path_str = str(file)
+        if self.api:
+            from unstructured.partition.api import partition_via_api
+
+            elements = partition_via_api(
+                filename=file_path_str,
+                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=file_path_str)
+
+        """ Process elements """
+        docs = []
+        file_name = Path(file).name
+        if split_documents:
+            for node in elements:
+                metadata = {"file_name": file_name}
+                if hasattr(node, "metadata"):
+                    """Load metadata fields"""
+                    for field, val in vars(node.metadata).items():
+                        if field == "_known_field_names":
+                            continue
+                        # removing coordinates because it does not serialize
+                        # and dont want to bother with it
+                        if field == "coordinates":
+                            continue
+                        # removing bc it might cause interference
+                        if field == "parent_id":
+                            continue
+                        metadata[field] = val
+
+                if additional_metadata is not None:
+                    metadata.update(additional_metadata)
+
+                metadata["file_name"] = file_name
+                docs.append(Document(text=node.text, metadata=metadata))
+
+        else:
+            text_chunks = [" ".join(str(el).split()) for el in elements]
+            metadata = {"file_name": file_name}
+
+            if additional_metadata is not None:
+                metadata.update(additional_metadata)
+
+            # Create a single document by joining all the texts
+            docs.append(Document(text="\n\n".join(text_chunks), metadata=metadata))
+
+        return docs

tests/test_reader.py

@@ -4,7 +4,7 @@ from langchain.schema import Document as LangchainDocument
 from llama_index.node_parser import SimpleNodeParser
 
 from kotaemon.base import Document
-from kotaemon.loaders import AutoReader
+from kotaemon.loaders import AutoReader, UnstructuredReader
 
 
 def test_pdf_reader():
@@ -26,3 +26,22 @@ def test_pdf_reader():
     node_parser = SimpleNodeParser.from_defaults(chunk_size=100, chunk_overlap=20)
     nodes = node_parser.get_nodes_from_documents(documents)
     assert len(nodes) > 0
+
+
+def test_unstructured_pdf_reader():
+    reader = UnstructuredReader()
+    dirpath = Path(__file__).parent
+    input_path = dirpath / "resources/dummy.pdf"
+    documents = reader.load_data(input_path)
+
+    # check document reader output
+    assert len(documents) == 1
+
+    first_doc = documents[0]
+    assert isinstance(first_doc, Document)
+    assert first_doc.text.lower().replace(" ", "") == "dummypdffile"
+
+    # split documents mode
+    documents = reader.load_data(input_path, split_documents=True)
+    # check document reader output
+    assert len(documents) == 1