A Hermes model swap can break work that looked solved
A Hermes setup is being used to build an order page on an existing WordPress site for a service that only works in certain locations. Before checkout, the page must check whether an entered address is valid for service.
The address search needs to look through a preloaded list of valid postal addresses, support partial search, show suggestions, tolerate typos, and understand short forms such as “St” for “Street” and “Rd” for “Road.” built the first address search prototype successfully. During the move into the wider order flow, OpenAI rate limits were reached, so the was switched to the paid .
Both models received the same Markdown . The practical problem is that DeepSeek struggled with an address integration that OpenAI handled more easily, leaving the cause unclear: either a model limitation or a prompting problem.
Key points
- Hermes is being used as the agent runner for a WordPress ordering flow.
- The address step must support partial search, , typo tolerance, and abbreviation handling.
- created the first working address search prototype.
- OpenAI rate limits pushed the work toward .
- The same Markdown did not lead to the same result across models.