Open-source document intelligence

Your documents are more than text.

Docling converts messy documents — PDFs, Office files, HTML, images and audio — into structured data, detecting tables, formulas, reading order, OCR and much more.

Install it locally and try it on your own documents. Move the same workflow to a service or private on-prem capacity when you are ready.


                                    pip install docling

Run Docling locally See real output

document.md

## 5 EXPERIMENTS

The primary goal of DocLayNet is to obtain
high-quality ML models capable of accurate
document-layout analysis [...]

|                | human | MRCNN | FRCNN | YOLO |
|----------------|-------|-------|-------|------|
| Caption        | 84-89 |  68.4 |  70.1 | 77.7 |
| Table          | 77-81 |  82.2 |  82.2 | 86.3 |
| All            | 82-83 |  72.4 |  73.4 | 76.8 |

Source page
Layout detection
Reading order
Structured output

Quickstart

Start in under a minute

Full documentation

No account, no upload, no proprietary SDK. Install the package and convert a document on the machine in front of you.


                                    pip install docling

Requires Python 3.10 or newer. Models are downloaded once and cached, after which conversion runs fully offline.

convert.py

from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("report.pdf")

print(result.document.export_to_markdown())

pip install docling Python API reference

terminal

# Convert a local file or a URL straight from the shell
docling report.pdf --to md --to json

# Point it at a whole directory
docling ./corpus --to md --output ./out

pip install docling CLI reference

docling-serve

# Run the same pipeline behind an HTTP endpoint
docker run -p 5001:5001 quay.io/docling-project/docling-serve

curl -X POST http://localhost:5001/v1/convert/source \
  -H 'Content-Type: application/json' \
  -d '{"sources": [{"kind": "http", "url": "https://arxiv.org/pdf/2206.01062"}]}'

Same document model, different transport. Docling Serve

mcp.json

{
  "mcpServers": {
    "docling": {
      "command": "uvx",
      "args": ["--from", "docling-mcp", "docling-mcp-server"]
    }
  }
}

Give an agent the ability to read documents properly. Docling MCP

See for yourself

Structure, carried through conversion

All supported formats

One page of a real academic paper: two columns, a complex table, a figure with its caption, and a running page header. Switch layers to see what Docling recovers — and what it would cost you to lose.

Reading order, not page order

A two-column page does not read top-to-bottom. Docling resolves the sequence a human would follow and stores components in that order, so text never interleaves across columns.

Markdown

## 5 EXPERIMENTS

The primary goal of DocLayNet is to obtain high-quality ML models
capable of accurate document-layout analysis on a wide variety of
challenging layouts. [...]

## Baselines for Object Detection

In Table 2, we present baseline experiments (given in mAP) on Mask
R-CNN, Faster R-CNN, and YOLOv5. [...]

Documentation

Table structure survives

Rows, columns and multi-level headers are recovered as structure — not flattened into a run of whitespace-separated numbers. Cells keep their spans, and the caption stays attached to the table it describes.

Markdown

**Table 2**: Prediction performance (mAP@0.5-0.95) of object
detection networks on DocLayNet test set.

|                | human | MRCNN R50 | MRCNN R101 | FRCNN R101 | YOLO v5x6 |
|----------------|-------|-----------|------------|------------|-----------|
| Caption        | 84-89 |      68.4 |       71.5 |       70.1 |      77.7 |
| Footnote       | 83-91 |      70.9 |       71.8 |       73.7 |      77.2 |
| Formula        | 83-85 |      60.1 |       63.4 |       63.5 |      66.2 |
| List-item      | 87-88 |      81.2 |       80.8 |       81.0 |      86.2 |
| ...            |       |           |            |            |           |
| All            | 82-83 |      72.4 |       73.5 |       73.4 |      76.8 |

Documentation

Figures, captions and classification

Pictures are extracted as image data, kept with their caption, and classified by content — so a chart is retrievable as a chart rather than lost as a gap in the text.

JSON

{
  "pictures": [
    {
      "self_ref": "#/pictures/0",
      "label": "picture",
      "annotations": [
        { "kind": "classification", "predicted_class": "line_chart" }
      ],
      "captions": [ { "$ref": "#/texts/14" } ],
      "prov": [
        { "page_no": 6, "bbox": { "l": 475, "t": 120, "r": 840, "b": 400 } }
      ]
    }
  ]
}

Documentation

Mathematics as LaTeX

Inline and display equations are detected and converted to LaTeX, so the notation stays searchable and renderable instead of degrading into broken glyph sequences.

Markdown

The estimator is defined as

$$\hat{\mu} = \frac{1}{N} \sum_{i=1}^{N} f(x_i)$$

with variance

$$\sigma^2 = \mathbb{E}[f(x)^2] - \mu^2$$

Documentation

Chunks that respect structure

Docling partitions the document into contiguous, structure-aware chunks ready for embedding — split on real boundaries such as sections and tables rather than on a fixed character count.

Python

from docling.chunking import HybridChunker

chunker = HybridChunker(tokenizer="BAAI/bge-small-en-v1.5")

for chunk in chunker.chunk(doc):
    print(chunk.text)
    print(chunk.meta.headings)     # section path
    print(chunk.meta.doc_items)    # back-references into the document

Documentation

Every element knows where it came from

Each component keeps its page number and bounding box. An answer built from a Docling document can be traced back to the exact region of the exact page it came from.

JSON

{
  "self_ref": "#/texts/22",
  "label": "section_header",
  "text": "5 EXPERIMENTS",
  "prov": [
    {
      "page_no": 6,
      "bbox": { "l": 70, "t": 875, "r": 230, "b": 900 },
      "charspan": [0, 13]
    }
  ]
}

Documentation

Source: DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation
Regenerate: docling https://arxiv.org/pdf/2206.01062 --to md --to json

Formats & portability

Convert once, reuse forever

All formats

PDFs run the world. They also lock away your knowledge in a printer language. Layout, tables, reading order and scanned pages — Docling gets those right where most tools give up and hand back a wall of text.

It also reads websites, Office files, images and audio the same way. Everything lands in DocLang, one open, high-fidelity document format, and the detail that lives in each original file — a Word file's section structure or review comments, a diagram in your Excel spreadsheet — comes across intact instead of being flattened or dropped.

Any format in

PDFs, websites, Office formats, scanned images and audio — including all the hard parts, not just the easy ones.

PDF
DOCX
PPTX
XLSX
HTML
Images
Audio

DocLang

One open, standardized form for every document. Everything the conversion found — structure, tables, reading order and each format's own detail — kept together in one place.

doclang.ai →

Any format out, anytime

Turn a document into Markdown, HTML, JSON or context-informed chunks whenever a later step needs it — produced from DocLang, without converting the original file again.

Markdown
HTML
JSON
Text

Why Docling

Built for structure, not just text

Structure, not a text dump

Hierarchy, layout, reading order, tables, formulas, pictures and provenance all survive conversion. What the page meant is still recoverable after parsing.

See what survives

One document model

Many input formats converge on the same expressive DoclingDocument, serializable to Markdown, JSON, HTML and DocLang. One shape to write your application against.

DoclingDocument

Private by design

Everything runs on your machine by default. No account, no upload, no outbound call — then the same pipeline moves to a private environment or managed infrastructure on your terms.

Deployment options

Ready for AI applications

Structure-aware chunking, exports and an MCP server connect Docling to RAG, agents, extraction and automation without a bespoke adapter.

Real workflows

Different jobs, familiar API

One document or a million, in the shape you need

Whether it's one document right now or all of them at once, a standard format or fields extracted into your own schema — Docling is ready for it.

Conversion for search

Extraction for databases

One at a time instant result

Interactive conversion

A user uploads a document and expects structured Markdown back while they wait. Content-preserving, single-document, latency-bound.


                                converter.convert(source)

On-demand extraction

Pull a defined set of fields out of one PDF or image against a schema, in time to answer a request or drive an agent step.


                                extractor.extract(source, template=Invoice)

In bulk high throughput

Corpus ingestion

Convert an archive once, chunk it, embed it, and keep provenance so retrieved passages trace back to the exact region on the page.


                                converter.convert_all(sources)

Batch structuring

Run a fixed schema across a population of PDFs or images and land typed rows in a warehouse rather than blobs of text.


                                for r in extractor.extract_all(sources, template=Invoice): ...

Local to production

From laptop to production, without a rewrite

Compare deployments

Prove the pipeline locally, keep the same document model and configuration, then change where the processing runs.

01

Local library

Prove extraction quality and learn the document model.

Python package, CLI, examples and notebooks with reproducible output.
Install locally
```
converter = DocumentConverter()
result = converter.convert(source)
```
02

Docling Serve

Wrap the pipeline in a service and test team workflows.

Containers, an HTTP API and an optional UI, running inside your network.
Deploy a local service
```
client = DoclingServiceClient(url='http://docling.internal:5001')
result = client.convert(source)
```
03

Managed & private — commercial

Add managed capacity or run privately, operated for you.

IBM's commercial offering: managed SaaS or private / on-prem deployment, with the same document model and API.
Docling for IBM watsonx
```
client = DoclingServiceClient(url=MANAGED_ENDPOINT, api_key=KEY)
result = client.convert(source)
```

Real workflows

What people build with it

More use cases

RAG and enterprise search

RAG with visual grounding

The problem

Flat text extraction destroys the two things retrieval depends on: where a passage sits in the document, and where it came from on the page. Chunks split mid-table and answers cannot be traced.

Why Docling

Every element and chunk keeps its page and bounding-box coordinates, so a retrieved passage points at the exact region on the page. You can highlight where an answer came from, not just cite a page number.

What it uses

Reading order
Structure-aware chunks
Tables
Figures
Provenance

Path to production

Batch processing
Concurrency
Managed queues
Private deployment

recipe.py

from docling.document_converter import DocumentConverter
from docling.chunking import HybridChunker

doc = DocumentConverter().convert("report.pdf").document

for chunk in (chunker := HybridChunker()).chunk(doc):
    embed(
        text=chunker.contextualize(chunk),
        metadata={"doc_id": chunk.meta.origin.binary_hash},
        # for resolving chunk.meta.doc_items page/bbox info
    )

RAG examples

Agents and automation

Give an agent a document it can actually read

The problem

An agent handed raw PDF bytes or a wall of stripped text has to guess at structure on every step. Tool calls become unreliable exactly where the document is hardest.

Why Docling

Docling exposes documents to agents through MCP, so reading, converting and extracting are typed tool calls against a stable document model instead of prompt-level parsing.

What it uses

MCP server
Schema-based extraction
Machine-readable output

Path to production

API reliability
Job orchestration
Governance

recipe.json

{
  "mcpServers": {
    "docling": {
      "command": "uvx",
      "args": ["--from", "docling-mcp", "docling-mcp-server"]
    }
  }
}

Docling MCP

Deployment

Choose how you run it

Deployment detail

Python library

Local development and custom pipelines

Install Docling

CLI

Batch conversion and scripts

View CLI

Docling Serve

Internal APIs and production services

Deploy Serve

Managed SaaS

Elastic capacity and managed operations

Scale with Docling

On-prem / private cloud

Security, residency, private infrastructure

Deploy privately

Ecosystem

It already fits your stack

All integrations

Docling is a library, not a platform you have to move into. It plugs into the frameworks, agents, pipelines and infrastructure you are already running.

Frameworks

Drop Docling in as the document loader for an existing AI stack.

Agents and MCP

Expose document reading to agents as typed tool calls.

Data and pipelines

Prepare, label and transform document corpora at volume.

Platforms and infrastructure

Run the same pipeline as a service, on a cluster, or on the edge.

Search and retrieval

Land structured chunks where your application reads them.

Momentum

Latest activities

All posts

Announcement Jun 15, 2026

Give your AI the document structure it is missing.

Start on your laptop with one command. Nothing to sign up for, nothing to upload, no proprietary SDK to adopt before you can judge the output.


                            pip install docling

Run Docling locally Deployment options View on GitHub

Your documents are more than text.

Start in under a minute

Structure, carried through conversion

Reading order, not page order

Table structure survives

Figures, captions and classification

Mathematics as LaTeX

Chunks that respect structure

Every element knows where it came from

Convert once, reuse forever

Any format in

DocLang

Any format out, anytime

Built for structure, not just text

Structure, not a text dump

One document model

Private by design

Ready for AI applications

One document or a million, in the shape you need

Interactive conversion

On-demand extraction

Corpus ingestion

Batch structuring

From laptop to production, without a rewrite

Local library

Docling Serve

Managed & private — commercial

What people build with it

RAG and enterprise search

The problem

Why Docling

What it uses

Path to production

Agents and automation

The problem

Why Docling

What it uses

Path to production

Choose how you run it

Python library

CLI

Docling Serve

Managed SaaS

On-prem / private cloud

It already fits your stack

Frameworks

Agents and MCP

Data and pipelines

Platforms and infrastructure

Search and retrieval

Latest activities

Docling for IBM watsonx: A Managed Service, Built on Open Source

"ChartNet: The Million-Scale Dataset Behind Robust Chart Understanding"

Moving Beyond Sparse Grounding with Complete Screen Parsing Supervision

Give your AI the document structure it is missing.