The model dropdown under Sohay → AI Settings lists the models Sohay will call for your selected provider. The list changes as providers release and retire models; what follows is how to choose from whatever is in it.
Start on the default
The defaults are the small, fast, inexpensive models: gpt-4o-mini on
OpenAI, gemini-2.5-flash on Google.
This is the right starting point, and for most support sites it is also the finishing point. The chatbot’s job here is not to reason from scratch — it is to read the articles retrieved from your Knowledge Base and answer from them. That is a task small models are genuinely good at, and one where a larger model mostly buys you a bigger bill.
When to move up a tier
Move up when you can point at a specific answer that was not good enough, and you have already ruled out the more likely cause. In rough order of likelihood, a bad answer is usually:
- A Knowledge Base gap. No article covered it. Writing the article fixes this; a bigger model does not.
- A retrieval miss. The article exists but did not match the question’s wording. Rewriting the article’s title and opening paragraph to sound like the question fixes this.
- A genuine reasoning failure. The right article was in front of the model and it still answered badly. This is the one a larger model fixes.
Reaching for a larger model before checking the first two is the expensive way to not fix your problem.
What a bigger model costs
Providers charge per token, and the rate scales with the tier. Moving from a mini model to a flagship can be a large multiple on your per-reply cost — and because retrieved articles are sent along with each question, support replies carry more input tokens than a bare chat would.
If you do move up, set a daily cap first. See Daily token caps and rate limits.
Restricting the list in code
Two filters set which model slugs the plugin is allowed to call:
add_filter( 'sohaychat_allowed_openai_models', function ( $models ) {
return array( 'gpt-4o-mini' );
} );
sohaychat_allowed_google_models does the same for Google. Narrowing the
allowlist is a hard ceiling on spend that no admin-screen mistake can undo —
useful on a site where several people can reach AI Settings.
sohaychat_default_ai_provider sets which provider answers when no choice
has been saved in the admin.
Where to go next
What Sohay costs to run for the arithmetic, and Daily token caps and rate limits for the ceilings.
