How to perform custom key-value extraction for pdf (similar to Azure Document Intelligence)?

[ranjith-3330]

Hi team,

I am trying to build a custom document extraction pipeline using Docling and would appreciate your guidance.

Currently, I am using Azure Document Intelligence (Custom Extraction). In that workflow, we upload documents, define the required fields, and annotate values using bounding boxes. After training, the model is able to extract only the specific key-value pairs we need, while ignoring irrelevant text.

I would like to achieve a similar result using Docling, but I’m not sure where to start or which approach to follow.

Could you please suggest:

The right way to implement this in Docling.

Thanks in advance!

[dosubot[bot]]

Hi @ranjith-3330! I'm Dosu and I’m helping the docling team.

Docling doesn't have a train-and-annotate workflow like Azure Document Intelligence, but it does have something quite close for schema-based key-value extraction: the ExtractionVlmPipeline, which uses the NuExtract VLM to extract structured fields from documents given a schema you define [1].

The user-facing API is DocumentExtractor. You define your desired fields as a JSON string, Python dictionary, or Pydantic model, and the pipeline extracts those fields from each page. Here's a quick example using a Pydantic schema:

from pydantic import BaseModel, Field
from docling.extraction.document_extractor import DocumentExtractor

class InvoiceSchema(BaseModel):
    bill_no: str = Field(examples=["A123", "5414"])
    total: float = Field(default=10.0, examples=[20.0])
    vendor_name: str = Field(examples=["Acme Corp"])

extractor = DocumentExtractor()
result = extractor.extract("your_document.pdf", schema=InvoiceSchema)

You can also use a simple dict or JSON string for the schema:

schema = {"bill_no": "string", "total": "number", "vendor_name": "string"}

The pipeline works by converting each page to an image, serializing your schema template, and then running the NuExtract-2B VLM to extract the structured data [2]. It supports hierarchical/nested Pydantic models for more complex extraction schemas as well [3].

There's a detailed example notebook at docs/examples/extraction.ipynb in the Docling repo that walks through all three template formats (string, dict, Pydantic) — that's the best place to start.

Key configuration options include model selection (repo_id), max_new_tokens, temperature, image scale factor, and quantization (load_in_8bit) via VlmExtractionPipelineOptions [4].

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.