PishTop Content Suggestion with AI
| 开发者 | micromax2 |
|---|---|
| 更新时间 | 2026年7月13日 00:18 |
| PHP版本: | 7.4 及以上 |
| WordPress版本: | 7.0 |
| 版权: | GPLv2 |
| 版权网址: | 版权信息 |
详情介绍:
PishTop Content Suggestion with AI is a next-generation related posts plugin that uses advanced semantic modeling. Rather than relying on simple keyword matching, it leverages local vector embeddings (generated via OpenRouter.ai) and cosine similarity in PHP to find contextually relevant recommendations.
To achieve maximum precision, the plugin performs local similarity pre-filtering and passes the top candidates to a selected OpenRouter Chat LLM (e.g., Gemini, GPT-4) for a final re-ranking step.
Key Features
- Embedding-Only Mode: Option to completely bypass the OpenRouter LLM re-ranking phase, using raw vector similarity to fetch recommendations locally, cutting API billing and optimizing site performance.
- Configurable Similarity Threshold: Enforce similarity score limits (0-100%) in the embedding-only phase to filter out low-relevance match candidates.
- Local Vector Storage: Custom database tables store text embedding vectors, preventing external API calls on every recommendation render.
- Embedding Model Versioning: Embeddings are versioned with the model ID. Stale embeddings (e.g., if you change models) are automatically filtered out.
- OpenRouter.ai Integration: Fetch high-quality text embeddings and perform LLM re-ranking using state-of-the-art models.
- Cost-Control Quotas: Separate daily limits block embedding and ranking operations once exceeded, automatically falling back to native category recommendations to prevent surprise API charges. Aligned with local WordPress timezone settings.
- Mutex Lock Stampede Protection: Restricts concurrent API calls during content updates, protecting your budgets under traffic surges.
- Display Layout Templates: Create custom wrapper and item markup right inside the settings dashboard. Include WooCommerce price and custom metadata variables instantly.
- Dynamic WooCommerce Contexts: Automatically overrides generic page titles on Cart, Checkout, and Thank You pages with active cart products and purchased items to yield relevant recommendations.
- WooCommerce Out-of-Stock Filter: Respects WooCommerce catalog visibility configurations, automatically hiding out-of-stock items from recommendations.
- User-Isolated Caching Keys: Cache transients on WooCommerce pages vary dynamically using secure cart item hashes and order IDs, preventing cross-user caching leakage.
- Developer Extensibility Filters: Hook into
pishtop_ai_post_textto modify source text dynamically, andpishtop_ai_recommendations_transient_keyto route custom transient cache partitions. - Viewport Lazy Loading: Trigger frontend AJAX suggestion loading via Intersection Observer, saving bandwidth and reducing API costs by querying only when the recommendations block becomes visible.
安装:
- Upload the plugin files to the
/wp-content/plugins/pishtop-content-suggestion-with-aidirectory, or install the plugin through the WordPress plugins screen directly. - Activate the plugin through the 'Plugins' screen in WordPress.
- Navigate to Settings > AI Suggestions to enter your OpenRouter API Key and configure your parameters.
- Use the shortcode
[pishtop_suggestions count="5" template="default_list"]to insert suggestions on your pages/posts.
常见问题:
How do I display suggestions?
You can insert suggestions using the shortcode [pishtop_suggestions count="5" template="default_list"] on your posts.
What API models does it support?
The plugin automatically pulls all available embedding and chat models from the OpenRouter API. You can choose any supported model (e.g., Google Gemini, OpenAI GPT, Cohere) in the Matching Engine settings tab.
How do cost budgets work?
You can set daily quota limits for both embedding generation and LLM re-ranking. Once a limit is reached, the plugin will seamlessly fall back to native WordPress matching to prevent any additional charges. Quotas reset at local midnight.
Can I bypass the LLM re-ranking phase?
Yes! Under the Matching Engine settings tab, you can disable the "Enable LLM Re-ranking" option. In this "Embedding-Only" mode, the plugin recommends items based purely on their local vector similarity, completely bypassing the OpenRouter LLM chat API call to save on cost and improve page load speeds. You can also configure a "Similarity Threshold (%)" so only items above a certain similarity match score are recommended.
Does it support WooCommerce pages?
Yes! When used inside WooCommerce Cart, Checkout, or Order Received (Thank You) pages, the plugin dynamically extracts active cart products or purchased order items to match suggestions, rather than using generic page titles. Out-of-stock items can also be hidden.
How does it handle caching on WooCommerce pages?
To prevent user cache cross-leakage, the caching transient keys are dynamically partitioned per customer session using MD5 hashes of cart items and order IDs.
What happens during a Cache Stampede?
If a popular post cache expires, a mutex lock is set. Parallel visitor hits on the page immediately receive native fallback recommendations while the first request fetches the new OpenRouter results in the background, keeping page loads fast and preventing multiple duplicate API charges.
How do developer filters work?
The plugin provides two filters:
pishtop_ai_post_text(filters post source text before embedding).pishtop_ai_recommendations_transient_key(filters cache keys).
How does the diagnostic log cap work?
The log table is capped at 5,000 rows. Pruning prunes old rows down to a configured ratio (e.g., 90%) to prevent database stress. A warning is logged if high volume causes early truncation.
更新日志:
1.6.7
- Remove all Gutenberg block registration and code references.
- Remove custom style tag allowance in templates sanitization to prevent reviewer rejection.
- Add an on-demand "Update Models" button next to the Embedding Model and LLM Re-ranking Model dropdown selects to clear local transients and fetch the fresh list from OpenRouter.
- Update the default registered settings prompt template defined in the plugin options initialization block inside the admin class.
- Update the "Reset to Default" prompt template inside the javascript settings dashboard to match the new default instructions template.
- Skip clearing and purging the transient recommendation cache when the only setting modified is the
enable_cachetoggle.
- Fix active recommendations count parsing in admin settings controller to support version-prefixed transient keys, resolving false zero indicator readings.
- Fix background ranking worker transient version prefix check and key resolution, resolving worker execution block.
- Enforce that LLM re-ranking (both via cron ranking worker and frontend matching cache misses) is executed only after database vector embedding indexing reaches 100% completion, falling back to local vector similarity matching during the indexing phase.
- Allow real-time frontend matching on cache misses to suggest posts immediately from the subset of already-indexed candidates.
- Implement cache transient key versioning (
pishtop_rec_v{version}_) to guarantee cache invalidation on external object caches (like Redis or Memcached). - Resolve query crashes on MySQL strict
ONLY_FULL_GROUP_BYSQL mode. - Update taxonomy queries to use
tt_ids(term taxonomy IDs) for strict relationship matching. - Remove third-party Google Fonts imports and styling to comply with GDPR privacy guidelines and WordPress.org directory policies.
- Bypass nonce check on the public suggestions AJAX endpoint to prevent page-cached guest visitor requests from breaking.
- Add key character filters on Template ID settings input fields to prevent template mismatch errors.
- Update default LLM re-ranking prompt templates with priority criteria, security rules, and JSON output contracts.
- Add LLM Shortfall Behavior option to let users choose between filling empty slots with similarity results or hiding them.
- Reorganize settings screen layout: rename OpenRouter Integration section to General, move Enable Caching before Cache Expiry (TTL), move Default Thumbnail Size after Final Output Sorting, and configure settings-dependent toggling fields.
- Auto-clear transient recommendations caches when saving settings to immediately show updated recommendations to visitors.
- Clear outdated embedding vector cache instantly on post updates.
- Improve log descriptiveness by adding context such as post ID, embedding model, API codes, and reasons.
- Remove Custom CSS styling from display templates and backend dashboard to comply with WordPress.org security policies regarding arbitrary code/styling insertion.
- Remove the shortcode alias
[ai_related_posts]completely to enforce proper prefixing standards and simplify usage. - Add dedicated External Services section to readme documenting integration, data transfer details, terms, and privacy policy links for OpenRouter.ai API services.
- Introduce Embedding-Only matching phase to completely bypass the OpenRouter LLM re-ranking step, saving on API costs and server overhead.
- Implement a configurable Similarity Threshold (0-100%) setting to filter out low-relevance candidate recommendations.
- Update settings in-app documentation and help tabs to cover the new features.
- Bypass database embedding storage and retrieval for WooCommerce dynamic pages (Cart, Checkout, and Thank You) to prevent different users from receiving mixed-up cart recommendations.
- Wrap all key AJAX endpoints, shortcode rendering, and background cron callbacks in defensive try-catch blocks to prevent HTTP 500 server errors.
- Guard WooCommerce cart access to prevent PHP fatal errors when WooCommerce cart is uninitialized.
- Resolve query crash by checking if taxonomy terms is WP_Error.
- Filter LLM-ranked IDs against candidates to drop prose numbers.
- Query configured post types in sorting logic to support hidden types.
- Prepare transient timeout LIKE query placeholder in database.
- Fix cron reschedule loop during direct hits to wp-cron.php (prevents worker execution blocking).
- Prevent PHP type errors on post saving hook when post is null.
- Add weekly custom cron interval schedule registration.
- Optimize database queries in ranking worker by checking active transients timeout option first.
- Guard against SQL syntax errors in cron queries when allowed post types is empty.
- Clear background indexing cron single events on deactivation and uninstall.
- Add template fields validation: enforces non-empty Wrapper HTML and Item HTML, and verifies Wrapper HTML contains {{items}} prior to submission.
- Add conditional metadata placeholders: supports single or fallback placeholders like {{meta:custom_key | {{title}} }} evaluating recursively from outside-in.
- Remove Native Category/Tag matching from default fallback options.
- Update Recent posts fallback to target the template's Target Post Type Filter.
- Move background ranking fallback checks to the cache miss stage to ensure pre-cached suggestions are correctly served from transients when background worker is active.
- Implement robust multi-stage fallback queries: if the category matching fallback returns fewer posts than requested, the plugin automatically fills up the remaining spots using recent posts of the same post type.
- Bypass transient cache lookups entirely during active background indexing or ranking runs to always serve fresh native fallback posts.
- Enforce native fallback rendering during active background indexing or ranking runs.
- Expand in-app Help and Documentation tab with comprehensive user guidelines.
- Add shortcut Settings link on Plugins list screen.
- Support {{id}} template layout placeholder.
- Initial release. Expose all parameters, timeouts, logging rates, delays, and customization options to the administrator dashboard.