forked from Archives/langchain
Docugami DataLoader (#4727)
### Adds a document loader for Docugami Specifically: 1. Adds a data loader that talks to the [Docugami](http://docugami.com) API to download processed documents as semantic XML 2. Parses the semantic XML into chunks, with additional metadata capturing chunk semantics 3. Adds a detailed notebook showing how you can use additional metadata returned by Docugami for techniques like the [self-querying retriever](https://python.langchain.com/en/latest/modules/indexes/retrievers/examples/self_query_retriever.html) 4. Adds an integration test, and related documentation Here is an example of a result that is not possible without the capabilities added by Docugami (from the notebook): <img width="1585" alt="image" src="https://github.com/hwchase17/langchain/assets/749277/bb6c1ce3-13dc-4349-a53b-de16681fdd5b"> --------- Co-authored-by: Taqi Jaffri <tjaffri@docugami.com> Co-authored-by: Taqi Jaffri <tjaffri@gmail.com>dynamic_agent_tools
Eugene Yurtsev
1 year ago
committed by
GitHub
parent
c2761aa8f4
commit
3c490b5ba3
@ -0,0 +1,25 @@
|
||||
# Docugami
|
||||
|
||||
This page covers how to use [Docugami](https://docugami.com) within LangChain.
|
||||
|
||||
## What is Docugami?
|
||||
|
||||
Docugami converts business documents into a Document XML Knowledge Graph, generating forests of XML semantic trees representing entire documents. This is a rich representation that includes the semantic and structural characteristics of various chunks in the document as an XML tree.
|
||||
|
||||
## Quick start
|
||||
|
||||
1. Create a Docugami workspace: http://www.docugami.com (free trials available)
|
||||
2. Add your documents (PDF, DOCX or DOC) and allow Docugami to ingest and cluster them into sets of similar documents, e.g. NDAs, Lease Agreements, and Service Agreements. There is no fixed set of document types supported by the system, the clusters created depend on your particular documents, and you can [change the docset assignments](https://help.docugami.com/home/working-with-the-doc-sets-view) later.
|
||||
3. Create an access token via the Developer Playground for your workspace. Detailed instructions: https://help.docugami.com/home/docugami-api
|
||||
4. Explore the Docugami API at https://api-docs.docugami.com/ to get a list of your processed docset IDs, or just the document IDs for a particular docset.
|
||||
6. Use the DocugamiLoader as detailed in [this notebook](../modules/indexes/document_loaders/examples/docugami.ipynb), to get rich semantic chunks for your documents.
|
||||
7. Optionally, build and publish one or more [reports or abstracts](https://help.docugami.com/home/reports). This helps Docugami improve the semantic XML with better tags based on your preferences, which are then added to the DocugamiLoader output as metadata. Use techniques like [self-querying retriever](https://python.langchain.com/en/latest/modules/indexes/retrievers/examples/self_query_retriever.html) to do high accuracy Document QA.
|
||||
|
||||
# Advantages vs Other Chunking Techniques
|
||||
|
||||
Appropriate chunking of your documents is critical for retrieval from documents. Many chunking techniques exist, including simple ones that rely on whitespace and recursive chunk splitting based on character length. Docugami offers a different approach:
|
||||
|
||||
1. **Intelligent Chunking:** Docugami breaks down every document into a hierarchical semantic XML tree of chunks of varying sizes, from single words or numerical values to entire sections. These chunks follow the semantic contours of the document, providing a more meaningful representation than arbitrary length or simple whitespace-based chunking.
|
||||
2. **Structured Representation:** In addition, the XML tree indicates the structural contours of every document, using attributes denoting headings, paragraphs, lists, tables, and other common elements, and does that consistently across all supported document formats, such as scanned PDFs or DOCX files. It appropriately handles long-form document characteristics like page headers/footers or multi-column flows for clean text extraction.
|
||||
3. **Semantic Annotations:** Chunks are annotated with semantic tags that are coherent across the document set, facilitating consistent hierarchical queries across multiple documents, even if they are written and formatted differently. For example, in set of lease agreements, you can easily identify key provisions like the Landlord, Tenant, or Renewal Date, as well as more complex information such as the wording of any sub-lease provision or whether a specific jurisdiction has an exception section within a Termination Clause.
|
||||
4. **Additional Metadata:** Chunks are also annotated with additional metadata, if a user has been using Docugami. This additional metadata can be used for high-accuracy Document QA without context window restrictions. See detailed code walk-through in [this notebook](../modules/indexes/document_loaders/examples/docugami.ipynb).
|
||||
@ -0,0 +1,343 @@
|
||||
"""Loader that loads processed documents from Docugami."""
|
||||
|
||||
import io
|
||||
import logging
|
||||
import os
|
||||
import re
|
||||
from pathlib import Path
|
||||
from typing import Any, Dict, List, Mapping, Optional, Sequence
|
||||
|
||||
import requests
|
||||
from pydantic import BaseModel, root_validator
|
||||
|
||||
from langchain.docstore.document import Document
|
||||
from langchain.document_loaders.base import BaseLoader
|
||||
|
||||
TD_NAME = "{http://www.w3.org/1999/xhtml}td"
|
||||
TABLE_NAME = "{http://www.w3.org/1999/xhtml}table"
|
||||
|
||||
XPATH_KEY = "xpath"
|
||||
DOCUMENT_ID_KEY = "id"
|
||||
DOCUMENT_NAME_KEY = "name"
|
||||
STRUCTURE_KEY = "structure"
|
||||
TAG_KEY = "tag"
|
||||
PROJECTS_KEY = "projects"
|
||||
|
||||
DEFAULT_API_ENDPOINT = "https://api.docugami.com/v1preview1"
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class DocugamiLoader(BaseLoader, BaseModel):
|
||||
"""Loader that loads processed docs from Docugami.
|
||||
|
||||
To use, you should have the ``lxml`` python package installed.
|
||||
"""
|
||||
|
||||
api: str = DEFAULT_API_ENDPOINT
|
||||
|
||||
access_token: Optional[str] = os.environ.get("DOCUGAMI_API_KEY")
|
||||
docset_id: Optional[str]
|
||||
document_ids: Optional[Sequence[str]]
|
||||
file_paths: Optional[Sequence[Path]]
|
||||
min_chunk_size: int = 32 # appended to the next chunk to avoid over-chunking
|
||||
|
||||
@root_validator
|
||||
def validate_local_or_remote(cls, values: Dict[str, Any]) -> Dict[str, Any]:
|
||||
"""Validate that either local file paths are given, or remote API docset ID."""
|
||||
if values.get("file_paths") and values.get("docset_id"):
|
||||
raise ValueError("Cannot specify both file_paths and remote API docset_id")
|
||||
|
||||
if not values.get("file_paths") and not values.get("docset_id"):
|
||||
raise ValueError("Must specify either file_paths or remote API docset_id")
|
||||
|
||||
if values.get("docset_id") and not values.get("access_token"):
|
||||
raise ValueError("Must specify access token if using remote API docset_id")
|
||||
|
||||
return values
|
||||
|
||||
def _parse_dgml(
|
||||
self, document: Mapping, content: bytes, doc_metadata: Optional[Mapping] = None
|
||||
) -> List[Document]:
|
||||
"""Parse a single DGML document into a list of Documents."""
|
||||
try:
|
||||
from lxml import etree
|
||||
except ImportError:
|
||||
raise ValueError(
|
||||
"Could not import lxml python package. "
|
||||
"Please install it with `pip install lxml`."
|
||||
)
|
||||
|
||||
# helpers
|
||||
def _xpath_qname_for_chunk(chunk: Any) -> str:
|
||||
"""Get the xpath qname for a chunk."""
|
||||
qname = f"{chunk.prefix}:{chunk.tag.split('}')[-1]}"
|
||||
|
||||
parent = chunk.getparent()
|
||||
if parent is not None:
|
||||
doppelgangers = [x for x in parent if x.tag == chunk.tag]
|
||||
if len(doppelgangers) > 1:
|
||||
idx_of_self = doppelgangers.index(chunk)
|
||||
qname = f"{qname}[{idx_of_self + 1}]"
|
||||
|
||||
return qname
|
||||
|
||||
def _xpath_for_chunk(chunk: Any) -> str:
|
||||
"""Get the xpath for a chunk."""
|
||||
ancestor_chain = chunk.xpath("ancestor-or-self::*")
|
||||
return "/" + "/".join(_xpath_qname_for_chunk(x) for x in ancestor_chain)
|
||||
|
||||
def _structure_value(node: Any) -> str:
|
||||
"""Get the structure value for a node."""
|
||||
structure = (
|
||||
"table"
|
||||
if node.tag == TABLE_NAME
|
||||
else node.attrib["structure"]
|
||||
if "structure" in node.attrib
|
||||
else None
|
||||
)
|
||||
return structure
|
||||
|
||||
def _is_structural(node: Any) -> bool:
|
||||
"""Check if a node is structural."""
|
||||
return _structure_value(node) is not None
|
||||
|
||||
def _is_heading(node: Any) -> bool:
|
||||
"""Check if a node is a heading."""
|
||||
structure = _structure_value(node)
|
||||
return structure is not None and structure.lower().startswith("h")
|
||||
|
||||
def _get_text(node: Any) -> str:
|
||||
"""Get the text of a node."""
|
||||
return " ".join(node.itertext()).strip()
|
||||
|
||||
def _has_structural_descendant(node: Any) -> bool:
|
||||
"""Check if a node has a structural descendant."""
|
||||
for child in node:
|
||||
if _is_structural(child) or _has_structural_descendant(child):
|
||||
return True
|
||||
return False
|
||||
|
||||
def _leaf_structural_nodes(node: Any) -> List:
|
||||
"""Get the leaf structural nodes of a node."""
|
||||
if _is_structural(node) and not _has_structural_descendant(node):
|
||||
return [node]
|
||||
else:
|
||||
leaf_nodes = []
|
||||
for child in node:
|
||||
leaf_nodes.extend(_leaf_structural_nodes(child))
|
||||
return leaf_nodes
|
||||
|
||||
def _create_doc(node: Any, text: str) -> Document:
|
||||
"""Create a Document from a node and text."""
|
||||
metadata = {
|
||||
XPATH_KEY: _xpath_for_chunk(node),
|
||||
DOCUMENT_ID_KEY: document["id"],
|
||||
DOCUMENT_NAME_KEY: document["name"],
|
||||
STRUCTURE_KEY: node.attrib.get("structure", ""),
|
||||
TAG_KEY: re.sub(r"\{.*\}", "", node.tag),
|
||||
}
|
||||
|
||||
if doc_metadata:
|
||||
metadata.update(doc_metadata)
|
||||
|
||||
return Document(
|
||||
page_content=text,
|
||||
metadata=metadata,
|
||||
)
|
||||
|
||||
# parse the tree and return chunks
|
||||
tree = etree.parse(io.BytesIO(content))
|
||||
root = tree.getroot()
|
||||
|
||||
chunks: List[Document] = []
|
||||
prev_small_chunk_text = None
|
||||
for node in _leaf_structural_nodes(root):
|
||||
text = _get_text(node)
|
||||
if prev_small_chunk_text:
|
||||
text = prev_small_chunk_text + " " + text
|
||||
prev_small_chunk_text = None
|
||||
|
||||
if _is_heading(node) or len(text) < self.min_chunk_size:
|
||||
# Save headings or other small chunks to be appended to the next chunk
|
||||
prev_small_chunk_text = text
|
||||
else:
|
||||
chunks.append(_create_doc(node, text))
|
||||
|
||||
if prev_small_chunk_text and len(chunks) > 0:
|
||||
# small chunk at the end left over, just append to last chunk
|
||||
chunks[-1].page_content += " " + prev_small_chunk_text
|
||||
|
||||
return chunks
|
||||
|
||||
def _document_details_for_docset_id(self, docset_id: str) -> List[Dict]:
|
||||
"""Gets all document details for the given docset ID"""
|
||||
url = f"{self.api}/docsets/{docset_id}/documents"
|
||||
all_documents = []
|
||||
|
||||
while url:
|
||||
response = requests.get(
|
||||
url,
|
||||
headers={"Authorization": f"Bearer {self.access_token}"},
|
||||
)
|
||||
if response.ok:
|
||||
data = response.json()
|
||||
all_documents.extend(data["documents"])
|
||||
url = data.get("next", None)
|
||||
else:
|
||||
raise Exception(
|
||||
f"Failed to download {url} (status: {response.status_code})"
|
||||
)
|
||||
|
||||
return all_documents
|
||||
|
||||
def _project_details_for_docset_id(self, docset_id: str) -> List[Dict]:
|
||||
"""Gets all project details for the given docset ID"""
|
||||
url = f"{self.api}/projects?docset.id={docset_id}"
|
||||
all_projects = []
|
||||
|
||||
while url:
|
||||
response = requests.request(
|
||||
"GET",
|
||||
url,
|
||||
headers={"Authorization": f"Bearer {self.access_token}"},
|
||||
data={},
|
||||
)
|
||||
if response.ok:
|
||||
data = response.json()
|
||||
all_projects.extend(data["projects"])
|
||||
url = data.get("next", None)
|
||||
else:
|
||||
raise Exception(
|
||||
f"Failed to download {url} (status: {response.status_code})"
|
||||
)
|
||||
|
||||
return all_projects
|
||||
|
||||
def _metadata_for_project(self, project: Dict) -> Dict:
|
||||
"""Gets project metadata for all files"""
|
||||
project_id = project.get("id")
|
||||
|
||||
url = f"{self.api}/projects/{project_id}/artifacts/latest"
|
||||
all_artifacts = []
|
||||
|
||||
while url:
|
||||
response = requests.request(
|
||||
"GET",
|
||||
url,
|
||||
headers={"Authorization": f"Bearer {self.access_token}"},
|
||||
data={},
|
||||
)
|
||||
if response.ok:
|
||||
data = response.json()
|
||||
all_artifacts.extend(data["artifacts"])
|
||||
url = data.get("next", None)
|
||||
else:
|
||||
raise Exception(
|
||||
f"Failed to download {url} (status: {response.status_code})"
|
||||
)
|
||||
|
||||
per_file_metadata = {}
|
||||
for artifact in all_artifacts:
|
||||
artifact_name = artifact.get("name")
|
||||
artifact_url = artifact.get("url")
|
||||
artifact_doc = artifact.get("document")
|
||||
|
||||
if artifact_name == f"{project_id}.xml" and artifact_url and artifact_doc:
|
||||
doc_id = artifact_doc["id"]
|
||||
metadata: Dict = {}
|
||||
|
||||
# the evaluated XML for each document is named after the project
|
||||
response = requests.request(
|
||||
"GET",
|
||||
f"{artifact_url}/content",
|
||||
headers={"Authorization": f"Bearer {self.access_token}"},
|
||||
data={},
|
||||
)
|
||||
|
||||
if response.ok:
|
||||
try:
|
||||
from lxml import etree
|
||||
except ImportError:
|
||||
raise ValueError(
|
||||
"Could not import lxml python package. "
|
||||
"Please install it with `pip install lxml`."
|
||||
)
|
||||
artifact_tree = etree.parse(io.BytesIO(response.content))
|
||||
artifact_root = artifact_tree.getroot()
|
||||
ns = artifact_root.nsmap
|
||||
entries = artifact_root.xpath("//wp:Entry", namespaces=ns)
|
||||
for entry in entries:
|
||||
heading = entry.xpath("./wp:Heading", namespaces=ns)[0].text
|
||||
value = " ".join(
|
||||
entry.xpath("./wp:Value", namespaces=ns)[0].itertext()
|
||||
).strip()
|
||||
metadata[heading] = value
|
||||
per_file_metadata[doc_id] = metadata
|
||||
else:
|
||||
raise Exception(
|
||||
f"Failed to download {artifact_url}/content "
|
||||
+ "(status: {response.status_code})"
|
||||
)
|
||||
|
||||
return per_file_metadata
|
||||
|
||||
def _load_chunks_for_document(
|
||||
self, docset_id: str, document: Dict, doc_metadata: Optional[Dict] = None
|
||||
) -> List[Document]:
|
||||
"""Load chunks for a document."""
|
||||
document_id = document["id"]
|
||||
url = f"{self.api}/docsets/{docset_id}/documents/{document_id}/dgml"
|
||||
|
||||
response = requests.request(
|
||||
"GET",
|
||||
url,
|
||||
headers={"Authorization": f"Bearer {self.access_token}"},
|
||||
data={},
|
||||
)
|
||||
|
||||
if response.ok:
|
||||
return self._parse_dgml(document, response.content, doc_metadata)
|
||||
else:
|
||||
raise Exception(
|
||||
f"Failed to download {url} (status: {response.status_code})"
|
||||
)
|
||||
|
||||
def load(self) -> List[Document]:
|
||||
"""Load documents."""
|
||||
chunks: List[Document] = []
|
||||
|
||||
if self.access_token and self.docset_id:
|
||||
# remote mode
|
||||
_document_details = self._document_details_for_docset_id(self.docset_id)
|
||||
if self.document_ids:
|
||||
_document_details = [
|
||||
d for d in _document_details if d["id"] in self.document_ids
|
||||
]
|
||||
|
||||
_project_details = self._project_details_for_docset_id(self.docset_id)
|
||||
combined_project_metadata = {}
|
||||
if _project_details:
|
||||
# if there are any projects for this docset, load project metadata
|
||||
for project in _project_details:
|
||||
metadata = self._metadata_for_project(project)
|
||||
combined_project_metadata.update(metadata)
|
||||
|
||||
for doc in _document_details:
|
||||
doc_metadata = combined_project_metadata.get(doc["id"])
|
||||
chunks += self._load_chunks_for_document(
|
||||
self.docset_id, doc, doc_metadata
|
||||
)
|
||||
elif self.file_paths:
|
||||
# local mode (for integration testing, or pre-downloaded XML)
|
||||
for path in self.file_paths:
|
||||
with open(path, "rb") as file:
|
||||
chunks += self._parse_dgml(
|
||||
{
|
||||
DOCUMENT_ID_KEY: path.name,
|
||||
DOCUMENT_NAME_KEY: path.name,
|
||||
},
|
||||
file.read(),
|
||||
)
|
||||
|
||||
return chunks
|
||||
@ -0,0 +1,28 @@
|
||||
"""Test DocugamiLoader."""
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
|
||||
from langchain.document_loaders import DocugamiLoader
|
||||
|
||||
DOCUGAMI_XML_PATH = Path(__file__).parent / "test_data" / "docugami-example.xml"
|
||||
|
||||
|
||||
@pytest.mark.requires("lxml")
|
||||
def test_docugami_loader_local() -> None:
|
||||
"""Test DocugamiLoader."""
|
||||
loader = DocugamiLoader(file_paths=[DOCUGAMI_XML_PATH])
|
||||
docs = loader.load()
|
||||
|
||||
assert len(docs) == 19
|
||||
|
||||
xpath = docs[0].metadata.get("xpath")
|
||||
assert str(xpath).endswith("/docset:Preamble")
|
||||
assert docs[0].metadata["structure"] == "p"
|
||||
assert docs[0].metadata["tag"] == "Preamble"
|
||||
assert docs[0].page_content.startswith("MUTUAL NON-DISCLOSURE AGREEMENT")
|
||||
|
||||
|
||||
def test_docugami_initialization() -> None:
|
||||
"""Test correct initialization in remote mode."""
|
||||
DocugamiLoader(access_token="test", docset_id="123")
|
||||
Loading…
Reference in New Issue