How Fabric's AI automatic tagging works

When people hear that Fabric has "AI automatic tagging," they usually assume we're doing what most other tools do—scanning your documents and sticking a few keyword labels on them. "Meeting notes," "project update," "client feedback." That sort of thing.

We get why people think this. It's what most tools actually do. But it's not how Fabric works, and this is our attempt to explain it better.

Why most auto-tagging doesn't work

Here's the thing about keyword tagging: it throws away almost everything that makes your content valuable. It's an incredibly lossy representation.

Say you've got a thoughtful research document about remote work trends. A typical auto-tagging system might label it "research" and "remote work." Technically accurate, but completely useless. It tells you nothing about the insights you found, the questions it raised, or how it connects to your other thinking.

It's like describing a movie by saying "it has actors and dialogue." True, but missing the point entirely.

What actually happens in Fabric

When you add something to Fabric, we don't extract keywords. Instead, we create what's called a vector representation—think of it as a very detailed fingerprint of your content's meaning.

Imagine if instead of filing documents in simple folders, you could place each one at exactly the right spot in a space with 12,000 different dimensions. Each dimension captures something different: the topic, sure, but also the tone, the complexity, the writing style, how it relates to other ideas, even subtle things like the underlying assumptions or emotional register.

Documents about similar topics end up near each other in this space, but so do pieces that share a way of thinking, or a particular approach to problems, or even just a similar vibe. The result is that Fabric understands your content in a much richer way than any keyword system could manage.

The hard part: making human-first search actually work

Creating these rich representations is just the beginning. The real challenge is using them well when you search or use the assistant.

When you type something into Fabric's search, we're not just matching words. We're trying to understand what you're actually looking for and find content whose "fingerprint" suggests it'll be useful to you. This is genuinely difficult—we're constantly tuning our algorithms to better interpret your intent and surface the right stuff. Our latest update was just three days ago.

Sometimes you search for "client meeting" but what you really want is that document where you worked through a tricky client situation, even if it never mentions meetings. Sometimes you search for "strategy" but you actually want tactical details from a project that implemented strategic thinking well. Getting this right is an ongoing challenge that requires real engineering effort.

But what about Fabric's manual tags?

You might have noticed that Fabric does have a tagging feature. This is a manual tool—something extra we built for people who like organizing things with their own categories. It's non-hierarchical, so you can tag however makes sense to you.

But here's the important bit: the manual tags don't power Fabric's intelligence. You never have to use them. The AI understanding happens entirely through those vector representations we talked about. The tags are just there if you want them.

Why This Matters

Understanding how Fabric really works explains why it feels different from other tools. You're not managing a filing system or trying to remember how you tagged something six months ago. You're working with software that actually understands your content and can find what you need based on what you're thinking about, not just what words you remember.

So no, Fabric doesn't do "auto-tagging" in the way most people imagine. It does something considerably more interesting.