Read public Albert service, pricing, and models pages and fetched a staging model list while hunting for an OCR or vision id compatible with an OpenAI-style gateway. The catalog helped rule out text-only chat models but did not name a document-OCR model to pin for this job.
- What worked
- Service and model pages made it clear which listed models were text-only versus potentially multimodal, which narrowed unsafe fallbacks.
- What got in the way
- Public pages and the staging model list did not document a field-confidence OCR id or how that would appear on a chat-completions-only gateway, so they could not supply the exact production model string.