IBM Open Sources Docling-Graph: Converting PDFs to Knowledge Graphs
aigleeson · x · 2026-08-17
IBM has open-sourced Docling-Graph, a tool that converts PDFs into validated knowledge graphs instead of vector embeddings. The pitch is that for chemistry, finance, and legal work, embeddings lose exact entity relationships, whereas knowledge graphs keep them explicit and queryable.
Pipeline:
- Parse: Docling parses inputs (PDF, image, DocLang, markdown, Office).
- Define: A Pydantic template defines the schema and graph structure in one file.
- Extract: A local VLM or LLM (routed through LiteLLM) extracts entities and edges.
- Output: Produces a NetworkX directed graph with stable IDs.
Key Features:
- Every node carries a provenance attribute by default, pointing back to the source chunk, page, and bounding box geometry, without extra LLM calls.
- Templates do not need to be hand-written: pointing it at 2-3 example PDFs induces a Pydantic class with fields and edges.
More from Infra
- Omarchy: DHH-endorsed, agent-first Linux operating system — vista8 · 2026-08-17
- Llama-3.0-Flash Leaked to Run End-to-End on a Single DGX Spark — Affectionate-File-26 · 2026-08-17
- AWS Trainium 4 Projected to Deploy 5M Units by 2H27, 12M by 2028 — zephyr_z9 · 2026-08-17
- Open-source ETL Duckle loads 20M rows in 15.69s, beats Airbyte — JafarNajafov · 2026-08-17
- How to choose an LLM API? Developers weigh cost, speed, and reliability — Puzzleheaded-Fig1589 · 2026-08-17
- llama.cpp Adds Support for Ling 3.0 Models — parepeg · 2026-08-17