Docling Document Intelligence

Docling Document Intelligence Skill

Use this skill to parse, convert, chunk, and analyze document structure and quality with Docling. It handles both local file paths and URLs, and outputs either Markdown or structured JSON (DoclingDocument).

Conversion uses the docling CLI (installed via uv tool install docling). The Python API is used only for features the CLI does not expose (chunking, VLM remote-API endpoint configuration, hybrid force_backend_text mode).

Scope

Step-by-Step Instructions

1. Resolve the input

Determine whether the user supplied a local path or a URL. The docling CLI accepts both directly.

docling path/to/file.pdf
docling https://example.com/a.pdf

2. Choose a pipeline

Docling has two pipeline families. Use this table to pick the right one:

Standard pipeline — no GPU needed; fast; OCR enabled automatically for scanned pages. VLM pipeline — needs GPU; slower; higher fidelity on complex layouts and handwriting.

See pipelines.md for the full decision matrix, OCR engine table (EasyOCR, RapidOCR, Tesseract, macOS), and VLM model presets.

3. Convert the document

CLI (preferred for straightforward conversions)

# --- Output format ---
docling report.pdf --output /tmp/                          # Markdown (default)
docling report.pdf --to json --output /tmp/               # JSON (structured)

# --- Pipeline selection ---
docling report.pdf --pipeline vlm --output /tmp/           # VLM for complex layouts
docling report.pdf --pipeline vlm --vlm-model granite_docling --output /tmp/  # VLM with specific model

# --- OCR options ---
docling report.pdf --ocr-engine tesserocr --output /tmp/  # Custom OCR engine
docling report.pdf --no-ocr --output /tmp/                # Disable OCR (speed)

# --- Speed / feature toggles ---
docling report.pdf --no-tables --output /tmp/             # Skip table extraction

# --- Remote VLM ---
docling report.pdf --pipeline vlm --enable-remote-services --output /tmp/

The CLI writes output files to the --output directory, named after the input file (e.g. report.pdf → report.md or report.json).

CLI reference: https://docling-project.github.io/docling/reference/cli/

Python API (for advanced features)

Use the Python API when you need features the CLI does not expose: chunking, VLM remote-API endpoint configuration, or hybrid force_backend_text mode.

Docling 2.81+ API note: DocumentConverter(format_options=...) expects dict[InputFormat, FormatOption] (e.g. InputFormat.PDF → PdfFormatOption). Using string keys like {"pdf": PdfPipelineOptions(...)} fails at runtime with AttributeError: 'PdfPipelineOptions' object has no attribute 'backend'.

Standard pipeline (default):

from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions

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

converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_options=PdfPipelineOptions(do_ocr=True, do_table_structure=True),
        ),
    }
)
result = converter.convert("report.pdf")

VLM pipeline — local (GraniteDocling via HF Transformers):

from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import VlmPipelineOptions
from docling.datamodel import vlm_model_specs
from docling.pipeline.vlm_pipeline import VlmPipeline

pipeline_options = VlmPipelineOptions(
    vlm_options=vlm_model_specs.GRANITEDOCLING_TRANSFORMERS,
    generate_page_images=True,
)
converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_cls=VlmPipeline,
            pipeline_options=pipeline_options,
        )
    }
)
result = converter.convert("report.pdf")

VLM pipeline — remote API (vLLM / LM Studio / Ollama):

This is only available via the Python API; the CLI does not expose endpoint URL, model name, or API key configuration.

from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import VlmPipelineOptions
from docling.datamodel.pipeline_options_vlm_model import ApiVlmOptions, ResponseFormat
from docling.pipeline.vlm_pipeline import VlmPipeline

vlm_opts = ApiVlmOptions(
    url="http://localhost:8000/v1/chat/completions",
    params=dict(model="ibm-granite/granite-docling-258M", max_tokens=4096),
    prompt="Convert this page to docling.",
    response_format=ResponseFormat.DOCTAGS,
    timeout=120,
)
pipeline_options = VlmPipelineOptions(
    vlm_options=vlm_opts,
    generate_page_images=True,
    enable_remote_services=True,  # required — gates all outbound HTTP
)
converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            pipeline_cls=VlmPipeline,
            pipeline_options=pipeline_options,
        )
    }
)
result = converter.convert("report.pdf")

Hybrid mode (force_backend_text) — Python API only:

Uses deterministic PDF text extraction for text regions while routing images and tables through the VLM. Reduces hallucination on text-heavy pages.

pipeline_options = VlmPipelineOptions(
    vlm_options=vlm_model_specs.GRANITEDOCLING_TRANSFORMERS,
    force_backend_text=True,
    generate_page_images=True,
)

result.document is a DoclingDocument object in all cases.

4. Choose output format

Markdown (default, human-readable):

docling report.pdf --to md --output /tmp/

Or via Python: result.document.export_to_markdown()

JSON / DoclingDocument (structured, lossless):

docling report.pdf --to json --output /tmp/

Or via Python: result.document.export_to_dict()

If the user does not specify a format, ask: "Should I output Markdown or structured JSON (DoclingDocument)?"

5. Chunk for RAG (hybrid strategy)

Chunking is only available via the Python API.

Default: hybrid chunker — splits first by heading hierarchy, then subdivides oversized sections by token count. This preserves semantic boundaries while respecting model context limits.

The tokenizer API changed in docling-core 2.8.0. Pass a BaseTokenizer object, not a raw string:

HuggingFace tokenizer (default):

from docling.chunking import HybridChunker
from docling_core.transforms.chunker.tokenizer.huggingface import HuggingFaceTokenizer

tokenizer = HuggingFaceTokenizer.from_pretrained(
    model_name="sentence-transformers/all-MiniLM-L6-v2",
    max_tokens=512,
)
chunker = HybridChunker(tokenizer=tokenizer, merge_peers=True)
chunks = list(chunker.chunk(result.document))

for chunk in chunks:
    embed_text = chunker.contextualize(chunk)
    print(chunk.meta.headings)  # heading breadcrumb list
    print(chunk.meta.origin.page_no)  # source page number

OpenAI tokenizer (for OpenAI embedding models):

import tiktoken
from docling_core.transforms.chunker.tokenizer.openai import OpenAITokenizer

tokenizer = OpenAITokenizer(
    tokenizer=tiktoken.encoding_for_model("text-embedding-3-small"),
    max_tokens=8192,
)
# Requires: uv add 'docling-core[chunking-openai]'

For chunking strategies and tokenizer details, see the Docling documentation on chunking and HybridChunker.

6. Analyze document structure

Use the DoclingDocument object directly to inspect structure:

doc = result.document

for item, level in doc.iterate_items():
    if hasattr(item, "label") and item.label.name == "SECTION_HEADER":
        print(f"{'#' * level} {item.text}")

for table in doc.tables:
    print(table.export_to_dataframe())  # pandas DataFrame
    print(table.export_to_markdown())

for picture in doc.pictures:
    print(picture.caption_text(doc))  # caption if present

For the full API surface, see Docling's structure and table export docs.

7. Evaluate output and iterate (required for "best effort" conversions)

After every conversion where the user cares about fidelity (not quick previews), run the bundled evaluator on the JSON export, then refine the pipeline if needed. This is how the agent checks its work and improves the run without guessing.

Step A — Produce JSON and optional Markdown

docling "<source>" --to json --output /tmp/
docling "<source>" --to md --output /tmp/

Step B — Evaluate

uv run scripts/docling-evaluate.py /tmp/<filename>.json --markdown /tmp/<filename>.md

If the user expects tables (invoices, spreadsheets in PDF), add --expect-tables. Tighten gates with --fail-on-warn in CI-style checks.

The script prints a JSON report to stdout: status (pass | warn | fail), metrics, issues, and recommended_actions (concrete docling CLI flags to try next).

Step C — Refinement loop (max 3 attempts unless the user says otherwise)

1. If status is warn or fail, apply one primary change from recommended_actions (e.g. switch --pipeline vlm, change --ocr-engine, ensure tables are enabled).
2. Re-convert with docling, re-run uv run scripts/docling-evaluate.py.
3. Stop when status is pass, or after 3 iterations — then summarize what worked and any remaining issues for the user.

Step D — Self-improvement log (skill memory)

After a successful pass or after the final iteration, append one entry to improvement-log.md in this skill directory:

• Source type (e.g. scanned PDF, digital PDF, DOCX)
• First-run problems (from issues)
• Pipeline + flags that fixed or best mitigated them
• Final status and one line of subjective quality notes

This log is optional for the user to git-ignore; it is for local learning so future runs on similar documents start closer to the right pipeline.

8. Agent quality checklist (manual, if script unavailable)

If scripts/docling-evaluate.py cannot run, still verify:

Common Edge Cases

Pipeline reference

Full decision matrix, all OCR engine options, VLM model presets, and API server configuration: pipelines.md

Output conventions

• Always report the number of pages and conversion status.
• When evaluation is in scope, report evaluator status, top issues, and which refinement attempt produced the final output.
• For Markdown output: wrap in a fenced code block only if the user will copy/paste it; otherwise render directly.
• For JSON output: pretty-print with indent=2 unless the user specifies otherwise.
• For chunks: report total chunk count, min/max/avg token counts.
• For structure analysis: summarise heading tree + table count + figure count before going into detail.

Dependencies

# Install the docling CLI as a uv tool
uv tool install docling
# For OpenAI tokenizer support, add to your project:
uv add 'docling-core[chunking-openai]'

Check installed versions (prefer distribution metadata — docling may not set __version__):

from importlib.metadata import version

print(version("docling"), version("docling-core"))

scripts/docling-evaluate.py declares its own dependencies as inline script metadata (PEP 723) — run it directly with uv run, no install step needed:

uv run scripts/docling-evaluate.py /tmp/report.json --markdown /tmp/report.md

Example

For a worked end-to-end walkthrough see EXAMPLE.md.