docling-project/docling-graph

What it solves

Docling Graph solves the problem of extracting structured, high-precision data from complex documents (PDFs, images, Office files) and converting it into a directed knowledge graph. Unlike standard RAG systems that rely on approximate text embeddings, this tool ensures exact semantic relationships between entities—such as chemical compounds and reactions or financial instruments and their dependencies—making it essential for specialized domains like chemistry, finance, and law.

How it works

The system uses a config-driven pipeline that transforms documents into validated Pydantic objects and then into NetworkX directed graphs. It supports two primary extraction paths: local VLM (Vision Language Model) extraction via Docling or LLM-based extraction routed through LiteLLM (supporting providers like OpenAI, Gemini, and vLLM). Users define extraction schemas using Pydantic templates, which can be written manually, induced from example documents, or compiled from existing ontologies (OWL/RDFS).

Who it’s for

It is designed for developers and data scientists working in high-precision domains (legal, financial, chemical research) who need to turn unstructured documents into machine-readable knowledge graphs with strict validation and data grounding.

Highlights

• Template Generation: Automatically induces Pydantic templates from example documents or compiles them from formal ontologies.
• Data Grounding: Provides a deterministic provenance ledger with bounding-box geometry to track exactly where data was extracted from.
• Flexible Backends: Supports both VLM and LLM extraction with compatibility for local runtimes (Ollama, vLLM) and remote APIs.
• Graph Fusion: Allows the merging of multiple knowledge graphs into a single, audited, deterministic graph without requiring additional LLM calls.
• Broad Format Support: Handles PDF, images, markdown, and Office files via Docling integration.