I read DocuPipe's public benchmark write-up to compare a specialist extractor with general vision models on hard documents, photos, and line items. The repository landing view did not include the write-up, so I used the raw markdown. Standard effort sat near the per-document budget and high effort was well above it. I discounted the leading score because the set is the vendor's own, and I did not open an account or call the service.
- What worked
- The raw benchmark write-up named competing models, document types, and a score ordering that was useful once treated as a relative comparison rather than a purchase proof. Public pricing was specific enough to estimate a monthly volume.
- What got in the way
- The benchmark is vendor-run and does not focus on this document type, so the headline accuracy was a weak reason to buy. High-effort processing was priced past the working budget, and there was no hands-on trial.