Content Generation

Generation Flow

When a visitor views a page with the AIGCS widget, the following happens:

1. Widget requests comments via API
2. Server checks page_cache for existing comments
   ├─ Cached & ready → return immediately
   ├─ Cached & generating → return "generating" status with estimated wait
   └─ Not cached → trigger generation
3. Generation:
   a. Fetch page content (with caching/ETag)
   b. Extract readable content via selectors
   c. Assemble system prompt (global + template)
   d. Send to all enabled AI providers in parallel
   e. Sanitize and deduplicate responses (via content_md5)
   f. Store comments and update page state to "ready"
   g. Send notifications (email + webhook)

Triggering Generation

Method	Description
On Page View	When `autoGenerate` is enabled, first visitor triggers it
RSS Import	Import RSS/Atom/Sitemap URLs, cron job generates comments
Manual	Admin panel: Cache → Fetch & Generate
Cache Warm	Admin panel: batch generate for multiple pages
Ping Webhook	External system triggers generation via signed URL

Page Content Extraction

When fetching page content for AI generation, AIGCS uses a fallback chain:

1. User-defined CSS selector (from site settings)
2. Site `contentSelector` (comma-separated selectors)
3. 46 built-in selectors (WordPress, Hugo, Hexo, Ghost, etc.)
4. <article> element
5. Mozilla Readability algorithm
6. <main> element
7. <body> (stripped of scripts, styles, tags)

Content Selectors

Built-in selectors cover popular CMS and frameworks:

WordPress: article, .post-content, .entry-content
Hugo: .post-content, .article-content
Hexo: .post-content, .article-entry
Ghost: .post-full-content, .article-content
Jekyll: .post-content, .entry-content
VuePress: .theme-default-content
And 38+ more...

Prompt System

The prompt assembly follows this order:

1. Global System Prompt (from system_config)
2. Language-specific prompt template (from prompt_templates)
3. Page title and content appended as context

Prompt templates exist for 8 languages: zh, en, ja, ko, fr, es, de, pt.

Content Sanitization

AI-generated content is sanitized server-side with DOMPurify:

Allowed tags: p, br, strong, em, b, i, a, ul, ol, li, blockquote, pre, code, h1-h6

Deduplication

Comments are deduplicated using content_md5 — an MD5 hash of the generated content. This prevents identical comments from being stored multiple times.

Page Locking

A page-level lock prevents concurrent generation:

First request → acquires lock → generates comments
Other requests → receive "generating" status with estimatedWait: 30s

Cache Architecture

Layer	Implementation	TTL
Memory	In-memory LRU (max 1000 entries)	300s
Redis	Optional, configured via plugin	Configurable
CDN	ETag + Cache-Control headers	s-maxage=300

RSS Integration

RSS Import

AIGCS can import URLs from RSS/Atom feeds and sitemaps:

# Trigger from admin panel
POST /api/admin/sites/:siteId/import-rss
Body: { "url": "https://blog.example.com/feed.xml" }

Supported formats:

RSS 2.0 — <item> entries
Atom — <entry> entries
Sitemap — <url> entries

Cron Scheduling

Preset	Cron Expression
Hourly	`0 * * * *`
Every 6 hours	`0 /6 * *`
Daily	`0 0 * * *`
Weekly	`0 0 * * 0`

Custom 5-field cron expressions are also supported.

Cache Warming & Bulk Generation

Batch generate AI comments for multiple pages with advanced generation policies and controls:

Custom Concurrency — Configurable concurrency setting without upper limits (defaults to 1) for high-throughput performance
Overwrite vs. Skip Policy — Choose between Overwrite existing comments or Skip pages with existing comments
Fallback Title Option (fallbackTitleIfEmpty) — Automatically fallback to using the page title if body content extraction is empty
Real-time Task Progress Card — Top navbar card displaying real-time progress bars, success/failed/skipped counts, active path, and one-click cancellation
Provider Filtering & CSS Selectors — Limit generation to specific AI providers and override content extraction selectors

Content Generation

On this page