If you've built RAG over PDFs, you know the problem. You pick a good embedding model, pay for a good LLM, and retrieval still returns garbage because the parser mangled the tables and chunked a footnote into the middle of a paragraph. This talk is about Docling, an open-source Python library for parsing documents into something an LLM can actually use. It handles more than ten file formats, preserves structure like hierarchy and tables, and uses deep learning models to read documents closer to how a person would. The output is a structured format called DoclingDocument, and once your documents are in that shape, a lot of downstream work gets easier. Chunking is the obvious example: instead of splitting on token counts and hoping for the best, you can chunk along the document's actual structure, keeping tables intact and respecting section boundaries. I'll cover how Docling works, show a parsing-to-retrieval pipeline built on structure-aware chunking, and walk through docling-agent, a newer project that skips chunking entirely. Instead of pre-chunking and embedding, the agent navigates the DoclingDocument structure at query time, which is an interesting answer to a problem chunking has always papered over. Docling is MIT-licensed and runs locally, which matters when your documents can't leave your environment or you care about latency. For engineers building RAG, agents over documents, or anything that starts with "get the text out of this PDF." Python familiarity assumed, no ML background needed.
Ming Zhao is an open source developer and Developer Advocate at IBM Research, where he helps IBM leverage open technologies while building impactful tools and growing vibrant open-source communities. He’s passionate about making open tech accessible to all and ensuring developers have the tools they need to succeed in the rapidly developing AI space. Ming now leads community efforts around Docling, IBM’s fastest-growing open source project, recently welcomed into the LF AI & Data Foundation.
