Process DOCX, PPTX, and HTML documents through Docling's unified pipeline — lab audio overview

Host: Welcome back. You're in GenAI Data Pipelines — a course about building the plumbing behind production AI systems. This chapter is all about document ingestion, which is the unglamorous but absolutely critical work of turning messy real-world files into clean text that a language model can actually reason over. And this specific skill — handling more than one file format through one consistent system — is something every data team hits the moment they move past the demo stage. Expert: Picture this. You're on a small data team at a mid-sized law firm, maybe fifteen engineers supporting two hundred attorneys. Leadership wants an AI assistant that can answer questions about any document the firm has ever produced. You go digging through the shared drives and you find the reality: contracts as PDFs, internal memos as Word documents, client pitch decks as PowerPoint slides, and knowledge base articles saved as web pages. Four different formats, millions of files, and no one is going to convert them by hand. If your ingestion pipeline only handles one format, the project dies. If you build four separate pipelines — one per format — you've now got four codebases to maintain, four sets of bugs, and four different output shapes that your search system has to juggle. That's the pain this exercise teaches you to solve. You'll build one ingestion path that swallows all of them and gives you one clean result. Host: Good. And in the previous exercise, you built the starting point — a working document reader that takes PDFs and turns them into structured text, using a tool called Docling. Quick reminder: Docling is an open-source document parsing library from IBM that acts as a universal reader. Now you're extending that same reader. So — what exactly will we build? Expert: You'll take the PDF reader you already have, and you'll teach it three more tricks. By the end, the same piece of code will accept a Word document, a PowerPoint slide deck, and a web page — that's HTML, which stands for Hypertext Markup Language, the language web pages are written in. Same reader, four formats. And critically, no matter what you feed it, it gives back the same shape of result — a structured representation with a title, a list of sections, tables pulled out cleanly, and the reading order preserved. That's the key idea, and it's worth sitting with. Think of it like a universal power adapter. The plug on the wall is different in every country, but the laptop on the other end only speaks one language. The adapter does the translation. Your ingestion pipeline is that adapter. Upstream, the world is chaotic — Word files with tracked changes, slide decks with text boxes floating everywhere, web pages full of navigation menus and ads. Downstream, your search system, your summarizer, your question-answering model — all of them want one predictable shape. The whole point of a unified pipeline is that the chaos stops at the adapter. You'll also configure each format slightly differently under the hood — a slide deck needs different handling than a Word document — but those tweaks are hidden behind one front door. Host: Makes sense. So before someone starts coding, what's the thing that tends to trip people up here? Expert: The trap is assuming "one pipeline" means "one setting." It doesn't. Docling is smart enough to auto-detect the file type, but each format has its own quirks that you have to think about. A PowerPoint deck has slides, and you need to decide whether each slide becomes its own section or whether the whole deck is one flowing document. Word documents have headers, footers, and tracked changes — do you keep them or strip them? Web pages have tons of junk — navigation bars, cookie banners, sidebar ads — that you almost never want in your final text. The mistake people make is running everything through the default settings and then being surprised when the output is polluted. The tip: before you batch-process anything, run one file of each type through on its own, look at the output with your own eyes, and decide what to keep and what to throw away. Treat the per-format settings as deliberate choices, not defaults. Host: Great tip. So after this exercise, what can the listener actually do, and where does it lead? Expert: After this, you'll be able to take a folder full of mixed document types — Word, PowerPoint, web pages, PDFs — and produce one clean, structured output for every single one, ready to feed into a search index or a language model. That's a real building block. This is the kind of component your team can drop into the foundation of an internal document search system or a retrieval-augmented assistant, and it just works regardless of what the business throws at it. Next, you'll extend this one more step — you'll build a quality evaluation layer that compares how well Docling handles each format, so you know where it shines and where you might need a backup plan. But that's for the next episode. For now, focus on getting one reader to handle all four formats cleanly. Thanks for listening, and good luck.