Latency figures on named hardware, the model comparison that overturned the obvious choice, and the failures that produced both. Written for the engineer deciding whether this belongs in their plant.
The four things that can happen to an uploaded document, and why the answer is usually decided by a contract nobody has re-read rather than by the vendor.
OCR makes the words findable. It does not make the manual answerable, and the gap between those two is where the work is.
What each is genuinely better at, measured numbers on both sides, and the three questions that settle it.
A practical definition, the four reasons manufacturers require it, and what changes architecturally when nothing may leave the building.
Measured latency, memory and throughput. Why concurrency, not corpus size, decides whether you need a GPU.
Why similarity thresholds cannot separate answerable questions from unanswerable ones, and the four controls that can.