personal_asset
A Smarter Model Still Can't Recover a Deleted Image
I hooked a vision model up to my personal knowledge base expecting it to search inside article screenshots too, and the test came back with zero images. The model wasn't the problem; the pipeline that turns a webpage into "clean text" had already stripped the pictures out before the model ever saw them. The ceiling on an AI tool's capability is often set upstream, where you're not looking.
I've been building my own personal knowledge base for a while now — articles I read, links I save, all funneled in so I can search them later. Recently I hooked up a model that can actually "look" at images: Qwen3-VL-32B-Instruct, a multimodal model Qwen released in October 2025. Their own description says it does deep visual understanding of documents, charts, and screenshots.
My plan was simple. A lot of the value in the articles I'd saved sits inside the images — a comparison table, a flowchart, an error screenshot. Text search alone can't reach any of that. Once the vision model was wired in, searching for "that one flowchart" should have pulled the right article straight up.
That night, I had the backend logs open and re-parsed an article that had images in it. When the result came back, the image count read: zero. Not a misread — the article went through the whole pipeline without a single picture ever reaching the model's eyes.
My first guess was a misconfigured model. I spent a while checking parameters — nothing wrong there. So I went back one step earlier, to the stage that turns a raw webpage into plain article text.
What that stage does is "clean up": strip out the ads, the navigation bar, the sidebar — everything that isn't the actual content — and keep only the readable text. It does this job well. The side effect is that it also treats image links as non-text noise and throws them out along with everything else. By the time that clean block of text reaches the vision model, the images are already gone. The model can be as good at seeing as you like — it can only see what's handed to it, and nothing was handed to it.
I paid a per-token bill for a capability that never got a chance to run. It genuinely works for the PDFs I upload directly, the screenshots I paste into a chat — those take a different path, where the images travel with the content and never pass through that "clean it up" stage. The gap isn't in the model. It's in whether the image survived long enough to reach it.
This sent me back through a few other places where I've plugged AI into things, checking for the same pattern: a model that's newer and bigger, fed input that had already been tampered with by some earlier step before the model ever saw it.
I'd bet this isn't the only place I've made this mistake — it just happened to get caught this time because the number came back as zero. Next time I wire up a new model capability, the first thing I'll do isn't test how good the model is. It's print out exactly what's being handed to it, and see what's actually left. The ceiling on a model's capability is often set one layer before the model.