Ask AI: Context Source, Custom Instructions, and Provider Selection
Three settings under Settings → AI control how Ask AI decides what to search, how it behaves, and which model answers.
Context source: Local vs Remote#
Under Settings → AI, the Context source setting decides where Ask AI searches for content when answering a reader’s question.
Local (WordPress). Searches docs directly from the WordPress database. This is the default, and it works without any remote storage configuration.
Remote (index). Fetches the doc index from your configured storage provider (Cloudflare R2 or Bunny.net) and searches that pre-built index instead, with no WordPress database queries during search. The label shows which provider is active, e.g. “Remote (Cloudflare R2 index)” or “Remote (Bunny.net index)”.
When to use Remote#
Use Remote if:
- Your site gets high AI-search traffic and you want to keep that load off the WordPress database.
- You’ve already set up index storage (Cloudflare R2 or Bunny.net) under Settings → Integrations → Remote Access for the Remote Access API.
The Remote option is disabled (greyed out) until storage is configured. If you select it before setup, the settings page tells you to configure storage under Settings → Integrations → Remote Access first.
If you’re not sure which to pick, stay on Local. It requires no setup and reads live from your published docs, which is right for most sites.
Custom instructions#

Also under Settings → AI, in the Custom instructions section, there’s a System prompt additions textarea (monospaced, pre-filled with a sensible default).
Whatever is in this box is appended to the AI’s system prompt on every Ask AI request. Use it to:
- Tune tone (more formal, more casual, brand voice).
- Restrict topics the assistant should or shouldn’t touch.
- Add product-specific context that doesn’t live in a doc.
The default instructions that ship in the box already cover:
- Search behavior: expanding a question into related concepts before searching, trying alternate wording if the first search comes up empty, and not giving up after a single failed search.
- Answering style: leading with the direct answer, using the documentation’s own terminology, and clearly separating documented fact from reasonable inference.
- Handling missing information: saying “I could not find documentation about this” rather than claiming a feature doesn’t exist, and answering the part of a question that can be answered even if part can’t.
A Reset to default button restores this text at any time (it asks for confirmation first, since it overwrites whatever you’ve typed).
You don’t need to touch this at all for Ask AI to work well. It’s there for teams who want to steer tone or scope beyond what the defaults already do.
Provider and model selector#
Also under Settings → AI, in the AI Provider section (labeled Override Provider (optional) if WordPress’s native AI is detected).
Why this changed#
Previously, if WordPress 7’s native AI was available, the provider and model dropdowns were hidden entirely, and you had no way to pick a specific model without disconnecting from native AI first. They’re now always visible, regardless of whether native AI is available.
Priority order#
- Your own API key, if one is saved, is always used, even on WordPress 7 with native AI available.
- WordPress native AI, if no key is saved and WordPress 7’s AI client is detected.
- If neither is available, AI features are unavailable and the plugin says so rather than failing silently.
In short: adding your own key always takes priority. Leave the key field blank to fall back to native AI where available.
What you can configure#

- Provider: OpenAI (ChatGPT), Anthropic (Claude), or Google Gemini.
- API Key: a password-style field. Once saved, it shows a masked placeholder and a confirmation that a key is stored; paste a new key to replace it.
- Model: a per-provider dropdown (e.g. Claude Haiku 4.5, GPT-4o mini, Gemini 2.0 Flash), each with a recommended default balancing speed, cost, and quality.
This is what lets you pick something like Claude Haiku specifically, instead of whatever model native AI happens to use.
Was this helpful?
Thanks for your feedback!
✓ Feedback received. Thank you!