The RAG capability you were about to buy just shipped free on every Mac

TL;DR

• My Mac rebooted itself overnight. I run a small AI agent on the machine, so I asked it what happened — and ended up reading, off my own process list, a piece of infrastructure most people will never see.
• Apple has quietly turned macOS into an on-device semantic search engine: it reads the meaning of your files, mail and photos and indexes them locally, no cloud round-trip. This is the same core capability — “retrieval-augmented” search, RAG — that vendors have spent two years selling to enterprises.
• The management point isn’t the engineering. It’s this: a capability your organization may have been about to put on a purchase order is now sitting, for free, on every employee’s laptop. That changes three conversations — your data strategy, your vendor spend, and your AI roadmap.
• It isn’t free of cost — it’s free of price. On a memory-tight machine the indexing is heavy (it helped tip mine over). “Built in” still has to be budgeted for, just in a different column.

My Mac rebooted itself overnight. No crash report on screen, no warning the evening before — just a fresh login window and that faint sense that something happened while I was asleep. So I did the thing I now do reflexively, because I run a small AI agent directly on this machine: I asked my Mac what killed it.

The answer was a kernel panic of the “watchdog timeout” flavour. There is a daemon whose only job is to periodically tell the kernel “I’m fine, everything’s fine.” It went quiet for 94 seconds. The kernel, assuming the worst, force-rebooted the whole machine. It came back cleanly on the first try, so no panic loop, no dying hardware. But a healthy laptop doesn’t garrote itself in the night for no reason.

The reason was memory. All day the system had been quietly suffocating: the OS had fired its out-of-memory killer roughly once an hour, more than ten times, picking off processes to stay alive. Most of my swap was in use, and almost no actual RAM was free. The machine was on AC power at full battery, so this was never about power. It was a small-memory laptop trying to breathe through a straw.

Then came the part that made me sit up. When I looked at what was actually pinning the processor, the top two offenders weren’t my browser, weren’t my own agents, weren’t anything I had opened. They were indexers — background services with names most people never meet. Sitting behind them was a whole quiet stack: services that run a classic keyword index alongside a semantic layer over your files, others doing the same meaning-extraction on your photos, and a service whose job is to compile machine-learning models down to run on Apple’s dedicated AI chip. All of them had spent the day blowing past their resource budgets.

Here’s the honest framing, because this isn’t a “secret feature” exposé. Apple stood on stage and announced smarter, natural-language search and Apple Intelligence. That part is public, and it’s genuinely good. What never made it onto a slide is the engineering reality underneath — the part you can read straight off your own process list: a full on-device index that captures the meaning of your files, mail and photos, with a machine-learning pipeline feeding it, all running locally on the laptop’s AI chip. They showed you the feature. They didn’t show you the machinery.

And that machinery is the story for anyone who runs a budget. Let me translate it out of the engineering and into the part that lands on a roadmap.

The thing your vendor was selling is now in the box

For about two years, “retrieval-augmented generation” — RAG — has been the premium layer of nearly every enterprise AI pitch. The shape is always the same: we’ll index your documents so your AI can actually answer questions about your own business, instead of making things up. It’s the feature that turns a generic chatbot into something that knows your contracts, your wiki, your support history. It is also, routinely, the line item with a recurring price tag — per seat, per document, per query.

What I read off my own laptop is that the core of that capability now ships, switched on, with the operating system. Every reasonably recent Mac in your organization is quietly building a semantic index of the meaning in its user’s files — locally, without sending anything to a cloud. The expensive, differentiated thing from the sales deck is becoming table stakes, bundled, on hardware you already own.

I want to be precise about what this does and doesn’t mean, because the wrong conclusion here is expensive in the other direction.

It does not mean “cancel your AI vendor on Monday.” Apple’s index is sealed for privacy — no outside application gets to query the semantic map of someone’s files. So this isn’t a drop-in replacement for a tool that reads your shared, organizational knowledge. The vendor’s real job — pointing retrieval at the team’s collective documents, with permissions and governance — still needs doing.

It does mean the floor just moved. When a capability goes from “premium add-on” to “comes free in the box,” three things change for a buyer, and the time to think about them is before the next renewal, not after.

Three conversations this should start

1. Your vendor spend, re-examined. Walk into your next renewal and separate what you’re paying for. How much of the price is the retrieval plumbing — indexing, embeddings, semantic search — versus the genuinely hard part, the governance, the shared-knowledge layer, the integrations, the security posture? The plumbing is commoditizing fast; on personal devices it’s now free. If a vendor’s pitch is mostly “we do semantic search,” that’s a weaker pitch this year than last. If their value is “we govern your shared knowledge safely and connect it to everything,” that’s the part worth paying for. Price the two separately and you negotiate from a much stronger position.

2. Your data strategy, made concrete. Here’s the uncomfortable flip side of “free on every laptop.” If every employee’s machine is now quietly building a rich, searchable index of everything that passes through it — mail, documents, photos — then the device just became a more capable repository of company information than your policies probably assume. That’s not a reason to panic; on-device and local is, if anything, the privacy-friendly version. But it is a reason to ask the question out loud: what’s the policy for a laptop that now understands the meaning of everything on it? Offboarding, lost-device, contractor machines — the stakes on each just went up a notch, quietly, without anyone signing off on it.

3. Your roadmap, recalibrated. If you have an internal project whose headline deliverable was “semantic search over our stuff,” check whether you’re about to spend a quarter building something the platform is busy commoditizing underneath you. The durable work — the part worth your team’s time — is almost never the retrieval engine itself. It’s the organizational layer on top: what knowledge you trust, who’s allowed to see what, how an answer gets traced back to a source someone can stand behind. Build the part the platform won’t hand you for free.

“Built in” is not the same as “free”

There’s a trap in the phrase “it’s just built into the OS now,” and my rebooted laptop is the cautionary tale.

This capability is not free of cost. It’s free of price. Reading the meaning of everything you own is genuinely heavy work, and on a memory-tight machine it bites. To be fair, the indexing didn’t crash my Mac on its own — the underlying disease was memory starvation I’d helped create with my own always-on processes, made worse because a fresh system upgrade triggers a complete first-time reindex of everything you own. The new indexing was the aggravator that shoved an already-saturated machine over the edge. It’s largely a one-time cost, and by the time I looked the load was already trending down.

But the management lesson generalizes well beyond Apple, and it’s the one I’d underline for anyone rolling out AI features at scale: a capability that arrives “for free” still has to be budgeted for — the cost just moves to a different column. It stops being a license fee and becomes a hardware and capacity question. “Bundled” AI features quietly raise the floor on the machine that has to run them. Spec your fleet for the assistant you’re about to give everyone, or you’ll pay for it in support tickets and dead afternoons instead of invoices.

I should be honest about why I met this before most people will. I’m running a pre-release version of macOS — I’m in Apple’s developer program, so I get these builds months ahead of everyone else. A pre-release OS makes both the watchdog reboots and the first-time reindex more aggressive than the finished product will be, so some of the sharp edges I hit are beta software, not the shipping version. But that’s exactly what makes my process list useful to you: it’s a preview of what quietly lands on your organization’s Macs when this ships to everyone. I just got the early, unpolished look at the machinery — and the early warning.

What I’d actually do with this

If your own machine is gasping after an upgrade, the practical fixes are mundane: wait out the one-time reindex, exclude network drives and heavy folders from search, and turn down the AI features you’re not using. The real fix, as ever, is freeing up memory. Losing search forever to dodge a one-time cost would be a bad trade.

But that’s the small version. The bigger move, if you’re the person deciding where the AI budget goes, is to treat this as a signal about where the value is migrating. The clever retrieval engine is becoming a commodity — it’s literally in the box now. The part that stays scarce, and therefore valuable, is the organizational judgement around it: what your company knows, who gets to see it, and whether you can trust the answer. Spend there.

So here’s my question for you: if you opened your last AI vendor invoice tonight and crossed out everything the operating system now does for free, how much of the bill would be left — and would what remains be the part you’re actually proud to pay for?