The composition model is the part I would copy even if I never shipped inference to a browser. Each stage reads one named field on a per-page row and writes another: `PdfToImage` writes `image`, `PaddleTextRecognizer` reads it and writes `text`, `GlinerNer` reads that and writes `ner` [7]. That is the same contract the Python ScaleDP library uses on Spark [7], and it gives a debug session stage boundaries. Empty `ner` with populated `text` is a model problem. Empty `text` is a rendering problem.
The byte budget is where the design gets awkward. Six plus 580 is 586 MB on a cold profile [17], and the entity model is roughly 97 times the size of the recogniser [20]. Swapping the default for `gliner-pii-edge` cuts 399 MB, landing near 187 MB, and IndexedDB caching means a browser profile pays once rather than per document [18][15]. The first visit still pays [15]. Downloading 580 MB to skip a form looks like a poor trade until you have waited in the queue for the form [4].
At the quoted hosted rates of $0.50 to $1.50 per thousand pages plus egress [5], the compute spend you displace is not what carries the decision. The author's own example workload, 10,000 pages a night [6], is $5 to $15 a night, or $1,825 to $5,475 a year before egress [19]. That is a rounding error against the artefacts the hosted third box drags into procurement [3]. The interesting saving is administrative, and the post is straight about it: with no third party in the path there is no processor, so the agreement is not merely easier to obtain, it does not apply, and "how long do you keep it" is answered by the tab closing [10].
Treat the 200 ms re-run figure as a measurement of somebody's laptop [12]. For it to transfer, three things have to hold on your users' machines: the weights are already cached, the changed parameter invalidates only downstream stages, and if the change is upstream there is enough local CPU to re-render the PDF at 300 dpi [9]. The library does not assume any of that. `isWebGpuAvailable` and `isCrossOriginIsolated` report what the tab actually got, because WebGPU may be absent and threads depend on response headers the host may never send [16]. Reporting the environment rather than assuming it is the part of this package that reads like it has been run somewhere real.
In my context the trade is worth taking for low-volume sensitive work: a few files per session, a reviewer moving thresholds, data that would otherwise need a signed processor agreement [10]. It is the wrong trade for a nightly batch, and the post says so rather than pretending otherwise [6].
