Documents to knowledge graphs, with receipts

What it does

Docling Graph ingests documents—PDFs, images, Office files, markdown—and extracts structured entities and relationships into Pydantic models. Those models become NetworkX directed graphs with stable IDs, edge metadata, and deterministic provenance linking every node back to its source chunk and page coordinates. The toolkit routes extraction through either Docling’s local VLM or LLM backends via LiteLLM, covering local runtimes and remote APIs.

The interesting bit

The project treats schema and graph structure as the same thing: you define a Pydantic template once, and it drives both the extraction prompt and the resulting graph shape. It can even generate those templates from example documents or existing ontologies, then deterministically merge multiple graphs without calling an LLM again.

Key highlights

• Provenance is default, not optional: every extracted node carries bounding-box geometry and source chunk references with no extra LLM inference.
• Supports deterministic graph fusion—merge multiple knowledge graphs into one audited structure without generative drift.
• Template generation works from documents (with structured-data induction and deterministic rendering) or from OWL/RDFS/SKOS/LinkML ontologies without any LLM involvement.
• Extraction backends span local VLMs, vLLM, Ollama, and commercial APIs through a single LiteLLM abstraction.
• Exports to CSV, Cypher, and other knowledge-graph formats, plus interactive HTML visualizations.

Verdict

Worth a look if you are building retrieval or analysis pipelines where “close enough” vector similarity is not acceptable. Skip it if you just need basic document chunking and embedding.