Docling

Docling parses PDF, DOCX, PPTX, HTML, and other formats into a rich unified representation including document layout, tables etc., making them ready for generative AI workflows like RAG.

This integration provides Docling's capabilities via the DoclingLoader document loader.

Overview

The presented DoclingLoader component enables you to:

• use various document types in your LLM applications with ease and speed, and
• leverage Docling's rich format for advanced, document-native grounding.

DoclingLoader supports two different export modes:

• ExportType.DOC_CHUNKS (default): if you want to have each input document chunked and to then capture each individual chunk as a separate LangChain Document downstream, or
• ExportType.MARKDOWN: if you want to capture each input document as a separate LangChain Document

The example allows exploring both modes via parameter EXPORT_TYPE; depending on the value set, the example pipeline is then set up accordingly.

Setup

%pip install -qU langchain-docling

Note: you may need to restart the kernel to use updated packages.

For best conversion speed, use GPU acceleration whenever available; e.g. if running on Colab, use a GPU-enabled runtime.

Initialization

Basic initialization looks as follows:

from langchain_docling import DoclingLoader

FILE_PATH = "https://arxiv.org/pdf/2408.09869"

loader = DoclingLoader(file_path=FILE_PATH)

For advanced usage, DoclingLoader has the following parameters:

• file_path: source as single str (URL or local file) or iterable thereof
• converter (optional): any specific Docling converter instance to use
• convert_kwargs (optional): any specific kwargs for conversion execution
• export_type (optional): export mode to use: ExportType.DOC_CHUNKS (default) or ExportType.MARKDOWN
• md_export_kwargs (optional): any specific Markdown export kwargs (for Markdown mode)
• chunker (optional): any specific Docling chunker instance to use (for doc-chunk mode)
• meta_extractor (optional): any specific metadata extractor to use

Load

docs = loader.load()

Token indices sequence length is longer than the specified maximum sequence length for this model (1041 > 512). Running this sequence through the model will result in indexing errors

Note: a message saying "Token indices sequence length is longer than the specified maximum sequence length..." can be ignored in this case — more details here.

Inspecting some sample docs:

for d in docs[:3]:
    print(f"- {d.page_content=}")

- d.page_content='arXiv:2408.09869v5  [cs.CL]  9 Dec 2024'
- d.page_content='Docling Technical Report\nVersion 1.0\nChristoph Auer Maksym Lysak Ahmed Nassar Michele Dolfi Nikolaos Livathinos Panos Vagenas Cesar Berrospi Ramis Matteo Omenetti Fabian Lindlbauer Kasper Dinkla Lokesh Mishra Yusik Kim Shubham Gupta Rafael Teixeira de Lima Valery Weber Lucas Morin Ingmar Meijer Viktor Kuropiatnyk Peter W. J. Staar\nAI4K Group, IBM Research R¨uschlikon, Switzerland'
- d.page_content='Abstract\nThis technical report introduces Docling , an easy to use, self-contained, MITlicensed open-source package for PDF document conversion. It is powered by state-of-the-art specialized AI models for layout analysis (DocLayNet) and table structure recognition (TableFormer), and runs efficiently on commodity hardware in a small resource budget. The code interface allows for easy extensibility and addition of new features and models.'

Lazy Load

Documents can also be loaded in a lazy fashion:

doc_iter = loader.lazy_load()
for doc in doc_iter:
    pass  # you can operate on `doc` here

End-to-end Example

import os

# https://github.com/huggingface/transformers/issues/5486:
os.environ["TOKENIZERS_PARALLELISM"] = "false"

• The following example pipeline uses HuggingFace's Inference API; for increased LLM quota, token can be provided via env var HF_TOKEN.
• Dependencies for this pipeline can be installed as shown below (--no-warn-conflicts meant for Colab's pre-populated Python env; feel free to remove for stricter usage):

%pip install -q --progress-bar off --no-warn-conflicts langchain-core langchain-huggingface langchain_milvus langchain python-dotenv

Note: you may need to restart the kernel to use updated packages.

Defining the pipeline parameters:

from pathlib import Path
from tempfile import mkdtemp

from dotenv import load_dotenv
from langchain_core.prompts import PromptTemplate
from langchain_docling.loader import ExportType


def _get_env_from_colab_or_os(key):
    try:
        from google.colab import userdata

        try:
            return userdata.get(key)
        except userdata.SecretNotFoundError:
            pass
    except ImportError:
        pass
    return os.getenv(key)


load_dotenv()

HF_TOKEN = _get_env_from_colab_or_os("HF_TOKEN")
FILE_PATH = ["https://arxiv.org/pdf/2408.09869"]  # Docling Technical Report
EMBED_MODEL_ID = "sentence-transformers/all-MiniLM-L6-v2"
GEN_MODEL_ID = "mistralai/Mixtral-8x7B-Instruct-v0.1"
EXPORT_TYPE = ExportType.DOC_CHUNKS
QUESTION = "Which are the main AI models in Docling?"
PROMPT = PromptTemplate.from_template(
    "Context information is below.\n---------------------\n{context}\n---------------------\nGiven the context information and not prior knowledge, answer the query.\nQuery: {input}\nAnswer:\n",
)
TOP_K = 3
MILVUS_URI = str(Path(mkdtemp()) / "docling.db")

API Reference:PromptTemplate

Now we can instantiate our loader and load documents:

from docling.chunking import HybridChunker
from langchain_docling import DoclingLoader

loader = DoclingLoader(
    file_path=FILE_PATH,
    export_type=EXPORT_TYPE,
    chunker=HybridChunker(tokenizer=EMBED_MODEL_ID),
)

docs = loader.load()

Token indices sequence length is longer than the specified maximum sequence length for this model (1041 > 512). Running this sequence through the model will result in indexing errors

Determining the splits:

if EXPORT_TYPE == ExportType.DOC_CHUNKS:
    splits = docs
elif EXPORT_TYPE == ExportType.MARKDOWN:
    from langchain_text_splitters import MarkdownHeaderTextSplitter

    splitter = MarkdownHeaderTextSplitter(
        headers_to_split_on=[
            ("#", "Header_1"),
            ("##", "Header_2"),
            ("###", "Header_3"),
        ],
    )
    splits = [split for doc in docs for split in splitter.split_text(doc.page_content)]
else:
    raise ValueError(f"Unexpected export type: {EXPORT_TYPE}")

API Reference:MarkdownHeaderTextSplitter

Inspecting some sample splits:

for d in splits[:3]:
    print(f"- {d.page_content=}")
print("...")

- d.page_content='arXiv:2408.09869v5  [cs.CL]  9 Dec 2024'
- d.page_content='Docling Technical Report\nVersion 1.0\nChristoph Auer Maksym Lysak Ahmed Nassar Michele Dolfi Nikolaos Livathinos Panos Vagenas Cesar Berrospi Ramis Matteo Omenetti Fabian Lindlbauer Kasper Dinkla Lokesh Mishra Yusik Kim Shubham Gupta Rafael Teixeira de Lima Valery Weber Lucas Morin Ingmar Meijer Viktor Kuropiatnyk Peter W. J. Staar\nAI4K Group, IBM Research R¨uschlikon, Switzerland'
- d.page_content='Abstract\nThis technical report introduces Docling , an easy to use, self-contained, MITlicensed open-source package for PDF document conversion. It is powered by state-of-the-art specialized AI models for layout analysis (DocLayNet) and table structure recognition (TableFormer), and runs efficiently on commodity hardware in a small resource budget. The code interface allows for easy extensibility and addition of new features and models.'
...

Ingestion

import json
from pathlib import Path
from tempfile import mkdtemp

from langchain_huggingface.embeddings import HuggingFaceEmbeddings
from langchain_milvus import Milvus

embedding = HuggingFaceEmbeddings(model_name=EMBED_MODEL_ID)

milvus_uri = str(Path(mkdtemp()) / "docling.db")  # or set as needed
vectorstore = Milvus.from_documents(
    documents=splits,
    embedding=embedding,
    collection_name="docling_demo",
    connection_args={"uri": milvus_uri},
    index_params={"index_type": "FLAT"},
    drop_old=True,
)

API Reference:HuggingFaceEmbeddings | Milvus

RAG

from langchain.chains import create_retrieval_chain
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain_huggingface import HuggingFaceEndpoint

retriever = vectorstore.as_retriever(search_kwargs={"k": TOP_K})
llm = HuggingFaceEndpoint(
    repo_id=GEN_MODEL_ID,
    huggingfacehub_api_token=HF_TOKEN,
)

API Reference:create_retrieval_chain | create_stuff_documents_chain | HuggingFaceEndpoint

def clip_text(text, threshold=100):
    return f"{text[:threshold]}..." if len(text) > threshold else text

question_answer_chain = create_stuff_documents_chain(llm, PROMPT)
rag_chain = create_retrieval_chain(retriever, question_answer_chain)
resp_dict = rag_chain.invoke({"input": QUESTION})

clipped_answer = clip_text(resp_dict["answer"], threshold=350)
print(f"Question:\n{resp_dict['input']}\n\nAnswer:\n{clipped_answer}")
for i, doc in enumerate(resp_dict["context"]):
    print()
    print(f"Source {i+1}:")
    print(f"  text: {json.dumps(clip_text(doc.page_content, threshold=350))}")
    for key in doc.metadata:
        if key != "pk":
            val = doc.metadata.get(key)
            clipped_val = clip_text(val) if isinstance(val, str) else val
            print(f"  {key}: {clipped_val}")

Question:
Which are the main AI models in Docling?

Answer:
The main AI models in Docling are a layout analysis model, which is an accurate object-detector for page elements, and TableFormer, a state-of-the-art table structure recognition model.

Source 1:
  text: "3.2 AI models\nAs part of Docling, we initially release two highly capable AI models to the open-source community, which have been developed and published recently by our team. The first model is a layout analysis model, an accurate object-detector for page elements [13]. The second model is TableFormer [12, 9], a state-of-the-art table structure re..."
  dl_meta: {'schema_name': 'docling_core.transforms.chunker.DocMeta', 'version': '1.0.0', 'doc_items': [{'self_ref': '#/texts/50', 'parent': {'$ref': '#/body'}, 'children': [], 'label': 'text', 'prov': [{'page_no': 3, 'bbox': {'l': 108.0, 't': 405.1419982910156, 'r': 504.00299072265625, 'b': 330.7799987792969, 'coord_origin': 'BOTTOMLEFT'}, 'charspan': [0, 608]}]}], 'headings': ['3.2 AI models'], 'origin': {'mimetype': 'application/pdf', 'binary_hash': 11465328351749295394, 'filename': '2408.09869v5.pdf'}}
  source: https://arxiv.org/pdf/2408.09869

Source 2:
  text: "3 Processing pipeline\nDocling implements a linear pipeline of operations, which execute sequentially on each given document (see Fig. 1). Each document is first parsed by a PDF backend, which retrieves the programmatic text tokens, consisting of string content and its coordinates on the page, and also renders a bitmap image of each page to support ..."
  dl_meta: {'schema_name': 'docling_core.transforms.chunker.DocMeta', 'version': '1.0.0', 'doc_items': [{'self_ref': '#/texts/26', 'parent': {'$ref': '#/body'}, 'children': [], 'label': 'text', 'prov': [{'page_no': 2, 'bbox': {'l': 108.0, 't': 273.01800537109375, 'r': 504.00299072265625, 'b': 176.83799743652344, 'coord_origin': 'BOTTOMLEFT'}, 'charspan': [0, 796]}]}], 'headings': ['3 Processing pipeline'], 'origin': {'mimetype': 'application/pdf', 'binary_hash': 11465328351749295394, 'filename': '2408.09869v5.pdf'}}
  source: https://arxiv.org/pdf/2408.09869

Source 3:
  text: "6 Future work and contributions\nDocling is designed to allow easy extension of the model library and pipelines. In the future, we plan to extend Docling with several more models, such as a figure-classifier model, an equationrecognition model, a code-recognition model and more. This will help improve the quality of conversion for specific types of ..."
  dl_meta: {'schema_name': 'docling_core.transforms.chunker.DocMeta', 'version': '1.0.0', 'doc_items': [{'self_ref': '#/texts/76', 'parent': {'$ref': '#/body'}, 'children': [], 'label': 'text', 'prov': [{'page_no': 5, 'bbox': {'l': 108.0, 't': 322.468994140625, 'r': 504.00299072265625, 'b': 259.0169982910156, 'coord_origin': 'BOTTOMLEFT'}, 'charspan': [0, 543]}]}, {'self_ref': '#/texts/77', 'parent': {'$ref': '#/body'}, 'children': [], 'label': 'text', 'prov': [{'page_no': 5, 'bbox': {'l': 108.0, 't': 251.6540069580078, 'r': 504.00299072265625, 'b': 198.99200439453125, 'coord_origin': 'BOTTOMLEFT'}, 'charspan': [0, 402]}]}], 'headings': ['6 Future work and contributions'], 'origin': {'mimetype': 'application/pdf', 'binary_hash': 11465328351749295394, 'filename': '2408.09869v5.pdf'}}
  source: https://arxiv.org/pdf/2408.09869

Notice that the sources contain rich grounding information, including the passage headings (i.e. section), page, and precise bounding box.

API reference

• LangChain Docling integration GitHub
• Docling GitHub
• Docling docs

Related

• Document loader conceptual guide
• Document loader how-to guides