A few weeks ago I wrote about building my own local AI system — a RAG pipeline that runs on my MacBook Pro and powers the knowledge base for my senior care business. That article was about the joy of building something yourself, of owning your own data, of running a real AI system for under two hundred dollars a year.
This article is about what happened when I tried to take that system to the cloud.
I wanted to go all-in on Google. That was the plan. I already use Google Workspace for everything — Gmail, Drive, Docs, Sheets. My entire business runs on it. My resident files, my compliance documents, my financial records, all of it lives on Google Drive. So when I discovered that Google Cloud had a product called Vertex AI RAG Engine — a managed search pipeline that could index my Drive files and let me query them with AI — I thought I had found the answer. One ecosystem. One vendor. One bill.
I set it up. Connected my Drive. Indexed my Golden Pines folder — about thirty-four hundred files covering resident records, care plans, medication logs, legal documents, incident reports. The system came online. I ran my first real test.
I asked it for a resident's state ID number. A simple question. The answer was sitting right there in a scanned PDF on my Drive — an image of the resident's state identification card. I could see it with my own eyes when I opened the file.
The system couldn't find it.
So I tried my local system — the one I built on my laptop with Qdrant and Apple Vision OCR. Same question. Same file. Also couldn't find it.
Two systems, one question, both failed.
Now here is where it gets interesting. When I manually ran Apple Vision — the OCR engine that ships free on every Mac — directly on that same PDF, it extracted the ID number perfectly. The technology worked. The pipeline didn't.
That sent me down a rabbit hole, and what I found was not a technical problem. It was an organizational one. Google's.
Let me walk you through what Google Cloud actually offers, because it is genuinely impressive when you look at the pieces individually.
Google has Document AI — one of the best optical character recognition systems in the world. It can read scanned documents, handwritten notes, structured forms. It is used by banks and hospitals and government agencies.
Google has text-embedding-005 — an embedding model that can convert not just text but also images and video into searchable vectors. Multimodal, they call it. Meaning you could theoretically take a photo of an ID card and the AI would understand what it is looking at without needing to convert it to text first.
Google has Vertex AI RAG Engine — a managed pipeline that takes your files, chunks them, embeds them, stores them in a vector database, and lets you query them.
Google has Agent Builder — another managed search product that does essentially the same thing as the RAG Engine but is a completely separate product with a completely separate interface.
Here is the problem. None of these talk to each other.
The RAG Engine does not use Document AI for OCR. It has its own parser that missed the embedded image in my resident's ID file. The multimodal embedding model can understand images, but the RAG Engine never sends it images — it only sends text. Agent Builder and RAG Engine are separate products that overlap almost entirely but share no infrastructure. I almost paid for both before I realized they were redundant.
So Google has the best OCR in the world, and their search pipeline doesn't use it. They have a multimodal embedding model, and their search pipeline only uses the text mode. They have two search products that do the same thing, built by what I can only assume are two teams that never had lunch together.
This is not a technology problem. Google has arguably the best AI technology on the planet. This is an integration problem. A coordination problem. Maybe even a culture problem.
And I think it matters far beyond my little senior care operation, because Google is not a small company that just hasn't gotten around to connecting the dots yet. This is one of the largest and most resourced technology companies in the history of civilization, and their own products do not work together.
I do not say this with anger. I say it with genuine dismay, because I wanted it to work. I wanted to write the article that said I went all-in on Google Cloud and it was seamless and everything just connected. That is not the article I am writing.
What I ended up building instead is a hybrid system. I use my own custom pipeline for the hard parts — Apple Vision for OCR, a local language model for metadata enrichment, my own chunking logic — and then I store the results in Qdrant Cloud, a managed vector database that is always on, accessible from anywhere, and happens to be HIPAA compliant. The embedding model I use for my healthcare data is Google's own text-embedding-005, called through their API, covered under their own HIPAA business associate agreement. So Google still gets my money for embeddings. They just do not get to run the pipeline, because their pipeline does not work.
The irony is that every piece of technology I needed existed inside Google's ecosystem. The OCR was there. The multimodal embeddings were there. The vector database was there. If someone at Google had simply wired Document AI into the RAG Engine, and enabled multimodal embedding on ingestion, I would be a pure Google Cloud customer right now. I would be writing a very different article.
Instead, I am a small business owner in Troy, Michigan, duct-taping together five different services from four different vendors because the biggest technology company in the world cannot connect its own products.
I think this is a pattern that goes beyond AI. Google has a history of building extraordinary things and then letting them exist in isolation. Google Chat and Google Meet and Google Spaces are three different communication tools that overlap and confuse. Google Drive and Google Docs have different sharing models. Google Keep and Google Tasks are two note-taking systems that serve almost identical purposes. The list goes on, and anyone who has used Google products for long enough has felt this fragmentation.
For a company that built its reputation on organizing the world's information, there is a deep irony in not being able to organize its own product suite.
I remain a fan of Google's technology. I use their workspace. I use their embedding model. I will probably use more of their services as they mature. But I am no longer waiting for them to assemble the pieces for me. I have learned, the hard way and over several very long late-night sessions, that the assembly is my job.
And maybe that is the real lesson here, for any entrepreneur trying to build with AI in this era. The technology is extraordinary. The platforms are not. The models can see and read and understand at a level that would have been science fiction five years ago. But the products wrapped around those models — the managed services, the pipelines, the one-click solutions — are still catching up. And until they do, the advantage goes to the people willing to build their own.
So that is what I will keep doing. Building. Connecting. Wiring things together that should have been wired together from the start. Because the residents in my care homes deserve a system that actually works, not one that almost works because the pieces were built by teams that never talked to each other.
