Creator Assistant Hub
| 开发者 | alessandrolioce |
|---|---|
| 更新时间 | 2026年10月11日 18:01 |
| PHP版本: | 7.4 及以上 |
| WordPress版本: | 7.1 |
| 版权: | GPL-2.0+ |
| 版权网址: | 版权信息 |
详情介绍:
Creator Assistant Hub is the core infrastructure plugin of the Creator Assistant ecosystem. It connects your WordPress site to OpenAI and Qdrant to automatically generate and store vector embeddings for your content, enabling semantic search and AI-powered features.
How it works
The Hub is the shared infrastructure layer for the Creator Assistant ecosystem. When a compatible add-on is active and activates a vectorizer, the plugin generates vector embeddings via the OpenAI API and stores them in your Qdrant vector database cluster whenever a post is published or updated. This makes your content available for semantic search and retrieval by Creator Assistant add-ons. Without an active add-on, the Hub runs in standby mode: no content is sent to external services.
Key features
- Automatic post vectorization: when an add-on activates a vectorizer, embeddings are generated and stored whenever a post is published or updated. Without an active add-on, no vectorization takes place.
- Chunk-based vectorization: splits post content into word-level segments and stores each chunk as a separate Qdrant point, enabling fine-grained retrieval (activated by add-ons via filter).
- Summary-based vectorization: generates a short AI summary of each post via the OpenAI Responses API, then embeds and stores a single point per post for high-level semantic search (activated by add-ons via filter).
- Background batch vectorization: an hourly cron job automatically picks up any existing posts that have not yet been vectorized, so your entire content library is indexed without manual intervention.
- OpenAI integration: uses the
text-embedding-3-smallmodel (1536 dimensions) via the OpenAI Embeddings API for all vectorizers; the Summary vectorizer additionally calls the OpenAI Responses API (gpt-4.1-nano) to generate post summaries. - Qdrant integration: stores and retrieves vector embeddings from your Qdrant Cloud cluster. Two collections are used:
post_chunkfor chunk-level points andpost_summaryfor post-level summary points. - Block editor panel: shows the vectorization status of each post directly in the block editor sidebar (document settings panel).
- Post list column: adds a "Vector Status" column to the posts and pages admin list, showing the vectorization status of each entry at a glance.
- Settings page: configure OpenAI API key, Qdrant endpoint and API key, and cache expiration from a dedicated admin page.
- API response caching: configurable cache duration to reduce API usage and latency.
- Site Health check: reports the Hub service status in the WordPress Site Health screen.
- Qdrant keep-alive: sends a daily lightweight ping to prevent free-tier Qdrant Cloud clusters from being suspended.
- WPML compatible: detects the post language via WPML hooks (when WPML is active) and includes it in the Qdrant payload for multilingual filtering.
- Extensible: exposes actions and filters so add-ons can register additional post types, override vectorizer behaviour, customise Qdrant payloads, and add settings tabs.
- An OpenAI account with a valid API key.
- A Qdrant Cloud account with an active cluster (free tier is supported).
- WordPress 6.8 or higher.
- PHP 7.4 or higher.
安装:
- Upload the
creator-assistant-hubfolder to the/wp-content/plugins/directory, or install the plugin directly from the WordPress plugin screen. - Activate the plugin through the Plugins screen in WordPress.
- Go to Creator Assistant → Hub in the WordPress admin menu.
- Accept the Terms of Service and Privacy Policy.
- Enter your OpenAI API key in the OpenAI section.
- Enter your Qdrant endpoint URL and Qdrant API key in the Qdrant section.
- Save the settings. Green status notices will confirm that both services are connected.
- Install a compatible Creator Assistant add-on (such as Creator Assistant Search) to activate vectorization. Once an add-on is active, new and updated posts are vectorized automatically on save, and existing posts in your library are picked up by the built-in hourly background batch. More add-ons are coming soon.
屏幕截图:
常见问题:
Do I need a paid OpenAI plan?
Yes. The OpenAI API is a pay-as-you-go service: a valid payment method and prepaid credits are required. There is no free tier for API usage. The text-embedding-3-small model used by the Hub is very affordable: generating embeddings for a typical blog post costs a fraction of a cent. Check OpenAI pricing for details.
Does the plugin work with the free tier of Qdrant Cloud?
Yes. The free tier of Qdrant Cloud provides a single cluster that may be suspended after a period of inactivity. The built-in keep-alive feature sends a daily ping to prevent this.
Which post types are vectorized?
By default, posts and pages are vectorized. Add-ons can register additional post types via the available hooks.
Where can I see the vectorization status of a post?
In two places:
- Block editor: open any post or page in the editor. A Creator Assistant panel in the document settings sidebar (top-right area) shows the vectorization status for each active vectorizer.
- Post list: a Vector Status column in the Posts and Pages admin list shows a green checkmark with the date of last vectorization, or a greyed-out label if the post has not yet been vectorized.
How do I check whether the service is properly configured?
Go to Tools → Site Health. A dedicated test named Creator Assistant Hub service reports whether the Terms of Service have been accepted and both OpenAI and Qdrant are connected and ready.
Also verify that WP-Cron is running correctly on your site. Vectorization is processed asynchronously via WP-Cron: if WP-Cron is disabled (e.g. DISABLE_WP_CRON is set to true in wp-config.php), posts will not be vectorized automatically. You can check this under Tools → Site Health → Info → WordPress constants. On some managed hosting plans, a real server-side cron job is used instead; make sure it is configured and firing regularly.
Can I disable API response caching?
Yes. In the Cache tab of the Hub settings page, set the cache expiration to Disabled.
Is my API key stored securely?
API keys are stored in the WordPress options table in the same way as other plugin settings. Make sure your WordPress database and hosting environment are secure. Keys are never exposed in the front end or in REST API responses accessible to non-administrators.
Do I need an active add-on for vectorization to work?
Yes. The Hub is a core infrastructure plugin: its vectorizers are disabled by default and must be activated by a compatible add-on (such as Creator Assistant Search). Without an active add-on, the plugin registers no vectorizers and no content is sent to OpenAI or Qdrant, even after you configure your API credentials.
Does the plugin modify my posts, pages, or any existing content?
No. The Hub only reads your content to generate vector embeddings. It never writes back to the post body, title, status, slug, or any other standard post field. Your content stays exactly as you wrote it. The only things the Hub writes to your WordPress database are:
- A post meta entry on each vectorized post that records when it was indexed (keys:
_creator_assistant_hub_post_vectorizer_post_chunkand_creator_assistant_hub_post_vectorizer_post_summary). This metadata is invisible to visitors, does not appear in SEO tools, and has no effect on how WordPress renders the post. - The plugin settings you enter (API keys, cache TTL, Terms of Service acceptance), stored in the WordPress options table.
Does the background cron job modify my content?
No. The hourly background cron reads post content to generate embeddings but never writes to the post itself. The only write operation it performs is updating the post meta timestamp once vectorization is complete.
What happens to my data if I deactivate or delete the plugin?
Deactivating the plugin stops all vectorization and background processing; your post meta timestamps and settings are preserved so they are ready if you reactivate. Deleting the plugin removes all plugin-created data from your WordPress database (post meta entries and options). Your Qdrant collections are not deleted automatically. You can remove them manually from your Qdrant Cloud dashboard if you no longer need them. Your posts, pages, and all their original content are completely unaffected either way.
What is the difference between Chunk and Summary vectorization?
The Hub provides two vectorization strategies, each stored in a separate Qdrant collection:
- Chunk: the post content is split into word-level segments. Each segment is embedded individually and stored as a separate point in the
post_chunkcollection. This enables precise, paragraph-level retrieval. - Summary: an AI-generated summary of the post is produced via the OpenAI Responses API, then embedded and stored as a single point in the
post_summarycollection. This enables high-level, concept-based semantic search.
creator_assistant_hub_post_vectorizer_post_chunk and creator_assistant_hub_post_vectorizer_post_summary filters.
Does the plugin support multilingual sites?
Yes. When WPML is active, the plugin detects the language of each post and includes it in the Qdrant payload. Add-ons can then filter search results by language. Sites without WPML are fully supported and unaffected.
Can other plugins or add-ons extend the Hub?
Yes. The Hub exposes WordPress actions and filters that add-ons can use to register additional settings sections, tabs, and custom behaviour. Developer documentation will be published on creator-assistant.it.
更新日志:
1.2.0
- New: a single manager for changes to post content, for add-ons (filter
creator_assistant_hub_content_injectors). - New: action
creator_assistant_hub_post_vectorized. - New:
Qdrant_Client::scroll_points()accepts a filter and can skip the vectors. - New:
OpenAi_Client::get_structured_response()accepts the model, reasoning effort and timeout (filtercreator_assistant_hub_openai_structured_options). - New: action
creator_assistant_hub_cluster_batch_received, fired once perPOST .../clustersafter the per-clustercreator_assistant_hub_clusters_receivedactions, witharray( 'collection', 'source_service', 'clusters' )(the raw top-level clusters, children and members included, so a listener can read a hierarchical tree). - New:
GET .../collections/{collection}/pointsacceptsmatch[<field>]=v1,v2query parameters to restrict the points to those whose payload field equals one of the values. - New:
Clients\Clustering_Client(trigger and task status of the clustering Lambda) and theCREATOR_ASSISTANT_HUB_CLUSTERING_LAMBDA_URLconstant, shared by the add-ons that cluster a collection. - Changed: the clustering job identifies itself as
clustering-job; the previous nameplanner-clustering-lambdais still accepted until the deployed image is rebuilt.
- New:
creator_assistant_hub_menu_positionsfilter (menu slug => integer position, lower first) to order the entries of the Creator Assistant admin menu. Each add-on declares the position of its own entry and the Hub sorts the menu; the Hub page always stays first, so the top-level menu keeps opening it, and entries without a position go last, in the order they were added. Before, the order was the order of registration, and a post type added under the menu came even before the Hub page. - Changed: the text passed to the vectorizers, and to the
creator_assistant_hub_post_vectorizer_textfilter, is now lightweight Markdown: headings, paragraphs, lists, quotes, code blocks and tables keep their structure instead of being flattened into one line. Classic-editor content (no<p>tags) keeps its paragraphs too. Parts returned by that filter are joined as separate paragraphs, soDocument::normalize_array()now returns separate paragraphs instead of one line. New helpersDocument::html_to_markdown()andDocument::normalize_markdown();Document::html_to_text()is unchanged. - Changed: post summaries are now 3-5 sentences of plain prose, stating the topic, angle and subtopics of the post, without lists, Markdown or opening phrases like "This article". The instructions are exposed as
OpenAi_Client::SUMMARIZE_INSTRUCTIONS, and the summary cache key includes them, so a prompt change never returns a stale summary. A truncated response (statusincomplete) is still returned, but is not cached. - Note: after updating, every collection must be re-indexed with the "force reindex" action of the Tools tab, because both the vectorized text and the summaries change. Summaries are longer, so the OpenAI output cost per post roughly doubles. Product summaries of creator-assistant-search-pro go through the new prompt too. If you use creator-assistant-planner, re-tune its similarity thresholds after re-indexing.
- Fixed: a post with a code example containing HTML (for instance
<div>in a code block) is no longer left without chunks. The chunk that contains markup now has its tags stripped before it is indexed, instead of the whole post being discarded and retried by the cron on every run. - Fixed: post summaries were never generated when the configured chat model is a reasoning model (for example gpt-5-nano). The request allowed 300 output tokens, reasoning included, and the model spent them all thinking and returned no text. It now allows 2000.
- Changed: every option, transient, admin-post action, query argument, CSS class and Qdrant collection now carries the plugin's full prefix (
creator_assistant_hub_*,creator-assistant-hub-*) instead ofcah_*,ca-*or no prefix. Custom CSS targeting the old.ca-*classes needs the new names. - New: several WordPress sites can share one Qdrant cluster. Every collection name now carries a random identifier of the site (
creator_assistant_<id>_hub_post_chunk), created the first time it is needed and stored in an option, so it does not change when the domain does and two sites never write into the same collection. It can be overridden with thecreator_assistant_hub_qdrant_namespacefilter, and a new one can be generated from the Tools tab, which a copy of a site (staging) should do. Thecreator_assistant_hub_qdrant_namespace_changedaction fires when it changes. - Changed: Qdrant cannot rename a collection, so on the first admin page load after updating the old ones (
post_chunk,post_summary) are deleted and every post is indexed again into the new ones by the hourly cron; the cost is one more round of OpenAI embeddings. - New: REST endpoints
GET /wp-json/creator-assistant-hub/v1/collections/{collection}/pointsandPOST /wp-json/creator-assistant-hub/v1/clusters, both authenticated via an RS256 JWT bearer token. Intended for a future external clustering service; not yet part of the public plugin, as no service is authorized to call them until its key pair is generated and registered. - New: filter
creator_assistant_hub_authorized_cluster_services(service name => PEM public key) to add or remove the external services allowed to call those two endpoints. A service signs its JWT with its own private key, with its name asissandcreator-assistant-hubasaud. - New: extensibility hook
creator_assistant_hub_clusters_received, fired once per top-level cluster afterPOST .../clustersstores it, with payloadarray( 'id', 'collection', 'source_service', 'centroid_point_id', 'aggregate', 'core_points' ), forwarded as the clustering job sent it (Hub does no Qdrant reads here: resolving a label or the text of the core points is up to the listener). - Changed:
labelin that payload is no longer an OpenAI-generated paraphrase of the whole cluster, and the hook no longer carrieslabelormember_queriesat all - saves one OpenAI call per cluster.OpenAi_Client::generate_label()has been removed as a result. - New: the "Google Connection" settings section now shows a plain-language description of what connecting a Google account is for. New filter
creator_assistant_hub_google_connected_services(default: empty array) lets add-ons that use the Google connection list, in their own words, why the site owner should enable it; when the filter returns a non-empty array, an intro sentence and the list are shown above the connect/disconnect button. - Fix: in
SCRIPT_DEBUGmode, the two Sass-authored stylesheets (ai-animated-border,settings-tabs) threw a fatal error instead of loading — their compileddistfile is the only browser-usable form, but the asset loader always looked for a same-named file insrcfirst. It now falls back todistfor them. - New: extensibility hooks
creator_assistant_hub_google_connectedandcreator_assistant_hub_google_disconnected, fired after a Google account is connected or disconnected, so add-ons that cache data read through the connection (property lists, account info, …) can drop that cache instead of serving it stale until its own TTL expires. - Fixed: a payload index declared by an add-on through
creator_assistant_hub_post_chunk_required_payload_indexesorcreator_assistant_hub_post_summary_required_payload_indexes(thelangindex of creator-assistant-hub-pro, the product fields of creator-assistant-search-pro) was created only by the next post vectorization. Activating or updating the add-on on a site already indexed made every filtered search fail with "Index required but not found" until some post was saved. The indexes are now synced on the next admin page, cron run or WP-CLI command, with no Qdrant call while the declared indexes are unchanged. An index Qdrant rejects is no longer marked as created: the next vectorization tries again, and the sync after 15 minutes. New methodPost_Vectorizer::sync_payload_indexes(); a strategy declares its indexes by overridingPost_Abstract::get_required_payload_indexes(). - Changed:
Qdrant_Client::collection_exists()now returnsnullnot only without credentials but also when the request fails (network error, non-2xx status, unreadable body), instead offalse:falsenow always means Qdrant answered that the collection does not exist. Compare the result withtrue/falseexplicitly. NewQdrant_Client::is_configured()tells the twonullcases apart. A failed check no longer makes the payload-index sync, the collection creation before a vectorization or the upgrade of the collection names treat an existing collection as missing. - Fix: a failed Google connection attempt (consent denied, the OAuth proxy unreachable, no refresh token in its response, …) silently bounced back to the same "Connect with Google" button with no indication anything went wrong — the only trace was a
WP_DEBUG-gated log line. The "Google Connection" settings section now shows an error notice in that case.
- New:
CreatorAssistantHub\Utils\Settings_Tabs, a shared renderer for the family's tabbed settings pages. It owns the tab markup, the CSS classes, the tab-switching script and the validation of add-on-supplied tab definitions, so every add-on with its own settings page gets the same behaviour without copying it. EnqueueSettings_Tabs::ASSET_HANDLE(or declare it as a dependency of your own handle) to pull in the stylesheet and script. - Fix: a settings section that no tab claims is now rendered in the first tab instead of being dropped. Dropping it removed its fields from the form entirely, so the value could never be saved and the add-on author got no hint why.
- Change: a tab can declare
hide_submit => trueto hide the Save button while it is active. Replaces the previous hard-coded list of tab slugs in the script.
- Fix: added
Plugin::run_at_init(), a helper that defers a callback to theinithook (or runs it immediately ifinithas already fired). Creator Assistant Search 1.0.0 already calls this method for its theme-aware registration logic (depends onwp_is_block_theme()); without this update, sites running both plugins hit a fatal error.
- Initial release.
- Automatic post and page vectorization using OpenAI
text-embedding-3-small. - Qdrant Cloud integration for vector storage and retrieval.
- Settings page with TOS acceptance, OpenAI, Qdrant, and cache configuration.
- Block editor sidebar panel showing per-post vectorization status.
- Post list column ("Vector Status") with per-vectorizer status indicators.
- Hourly background batch cron to automatically vectorize existing content.
- Site Health integration reporting service status.
- Daily Qdrant keep-alive cron for free-tier clusters.
- Extensibility hooks for add-ons, by category. Settings:
creator_assistant_hub_loaded,creator_assistant_hub_after_register_settings,creator_assistant_hub_settings_tabs,creator_assistant_hub_sanitize_settings; OpenAI models:creator_assistant_hub_openai_embedding_model,creator_assistant_hub_openai_chat_model,creator_assistant_hub_embedding_dimension; vectorizer control:creator_assistant_hub_post_vectorizer_post_chunk,creator_assistant_hub_post_vectorizer_post_summary,creator_assistant_hub_skip_vectorizer,creator_assistant_hub_post_vectorizer_text,creator_assistant_hub_post_vectorizer_chunks; post types:creator_assistant_hub_additional_post_types,creator_assistant_hub_excluded_post_types; Qdrant payloads:creator_assistant_hub_post_vectorizer_post_chunk_payload,creator_assistant_hub_post_vectorizer_post_summary_payload,creator_assistant_hub_post_vectorizer_post_chunk_search_filter; collection lifecycle:creator_assistant_hub_post_vectorizer_post_chunk_collection_created,creator_assistant_hub_post_vectorizer_post_summary_collection_created.