And it's not just for companies. Think about semantically searching through tax law, rental contracts, or insurance policies. The same stack works for anyone drowning in documents they didn't choose to read.

I'm sharing this because I'm tired of reading about companies struggling with digitalization over and over again. I don't have the reach to change things on my own, but maybe someone out there can pick this up and actually do something with it.

Companies need to digitize millions of documents. The tooling exists. The models exist. What's missing is people connecting the pieces and making it accessible. That shouldn't cost millions, it should be shared knowledge.

One of the biggest challenges in document processing is chunking (how you split a document into meaningful pieces for embedding). Get it wrong and your search results are useless. I've been experimenting with Docling's output and found approaches that work really well. Will share more on this.

The core stack is surprisingly accessible: Docling (by IBM Research) for extracting structured markdown from PDFs, Qdrant (from Berlin) as a vector database for semantic search, and open source LLMs from Ollama or Hugging Face for OCR and retrieval. All of this is available today, for free.

Consulting firms charge millions for document digitization and AI-powered search. A lot of the underlying technology is open source and freely available. Here's what I've found, sharing it because this knowledge shouldn't be locked behind six-figure contracts.

Already running local-first pipelines: qdrant-swift combined with llama.cpp for fully on-device RAG.

No API calls, no cloud dependencies. PDF processing, embedding, search, and inference all running locally on macOS. Will share more on this soon.

The Swift client lets me build macOS apps that search and retrieve against those collections directly for RAG, semantic search, or finding the right passage across dozens of papers.

I process research papers and technical documents with Docling, which extracts clean markdown from PDFs.

That content gets chunked and embedded and I needed a way to query Qdrant natively from Swift without bridging through other languages.

I needed this, so I built it.

qdrant-swift: a native Swift client for @qdrant.bsky.social vector database. git.new/j7wIyIu

Why I built it and how I use it ↓