Just watched the Docling webinar live. Two things worth noting.

Docling Agent - official repo is up (docling-project/docling-agent). Agentic doc operations: writing, editing, extraction. Works with DoclingDocument in/out, runs locally. Still early stage but the direction is clear, Docling is moving beyond conversion.

Chunkless RAG - instead of the classic chunk+embed+cosine pipeline, the idea is to use graph/tree structures that preserve document hierarchy. Sections, tables, figures stay connected. The LLM navigates the structure instead of searching isolated text fragments. Also designed to run locally.

If you've debugged RAG pipelines you know chunking is where most quality issues come from. This basically says stop flattening documents into chunks, use the structure for retrieval instead.

Makes sense given Docling already has the richest document representation out there. Why flatten a perfect tree into text blobs.

Repo for docling-agent is public on github. More details on chunkless RAG probably coming soon.