ai-workflows 2.1.1 → 2.3.0

package/examples/02-content-moderation-cascade.ts

@@ -0,0 +1,454 @@
+/**
+ * Example: AI Content Moderation with Human Escalation
+ *
+ * This example demonstrates the CascadeExecutor pattern for content moderation.
+ * Content is processed through escalating tiers:
+ *
+ * 1. Code (rules) - Fast keyword/pattern matching
+ * 2. Generative AI - LLM-based analysis for nuanced content
+ * 3. Agentic AI - Multi-step reasoning with tools
+ * 4. Human - Expert review for edge cases
+ *
+ * Key concepts demonstrated:
+ * - CascadeExecutor for tiered escalation
+ * - Per-tier timeouts and retry configuration
+ * - Skip conditions for tier bypassing
+ * - 5W+H audit trail for compliance
+ * - Integration with workflow events
+ *
+ * @example
+ * ```bash
+ * npx tsx examples/02-content-moderation-cascade.ts
+ * ```
+ */
+
+import {
+  Workflow,
+  CascadeExecutor,
+  type TierContext,
+  type WorkflowContext,
+  TIER_ORDER,
+} from '../dist/index.js'
+
+// ============================================================================
+// Type Definitions
+// ============================================================================
+
+interface ContentItem {
+  id: string
+  type: 'text' | 'image' | 'video'
+  content: string
+  metadata: {
+    userId: string
+    platform: string
+    timestamp: number
+  }
+}
+
+interface ModerationResult {
+  action: 'approve' | 'reject' | 'flag'
+  confidence: number
+  reason: string
+  categories?: string[]
+  requiresReview?: boolean
+}
+
+interface HumanReviewTask {
+  contentId: string
+  content: string
+  aiAnalysis: string
+  priority: 'low' | 'medium' | 'high'
+  assignedTo?: string
+}
+
+// ============================================================================
+// Mock Services
+// ============================================================================
+
+// Blocked keywords for rule-based tier
+const BLOCKED_KEYWORDS = ['spam', 'scam', 'hack', 'illegal', 'violence']
+const SUSPICIOUS_PATTERNS = [
+  /\b(buy|sell)\s+followers\b/i,
+  /\b(free|win)\s+\$?\d+/i,
+  /click\s+here\s+now/i,
+]
+
+// Mock AI service
+const aiService = {
+  async analyzeContent(content: string): Promise<{
+    isSafe: boolean
+    confidence: number
+    categories: string[]
+    explanation: string
+  }> {
+    // Simulate AI analysis with some latency
+    await new Promise((resolve) => setTimeout(resolve, 50))
+
+    // Mock logic based on content length and keywords
+    const hasIssues = content.toLowerCase().includes('suspicious') || content.length > 500
+    const confidence = hasIssues ? 0.65 : 0.95
+
+    return {
+      isSafe: !hasIssues,
+      confidence,
+      categories: hasIssues ? ['potentially_harmful'] : [],
+      explanation: hasIssues
+        ? 'Content requires further analysis due to suspicious patterns'
+        : 'Content appears safe based on AI analysis',
+    }
+  },
+}
+
+// Mock agentic AI with tool access
+const agenticService = {
+  async deepAnalysis(
+    content: string,
+    context: TierContext
+  ): Promise<{
+    isSafe: boolean
+    confidence: number
+    reasoning: string[]
+  }> {
+    // Simulate multi-step reasoning
+    await new Promise((resolve) => setTimeout(resolve, 100))
+
+    const reasoning = [
+      'Step 1: Analyzed semantic meaning of content',
+      'Step 2: Checked against known harmful patterns',
+      'Step 3: Verified user history and context',
+      'Step 4: Applied platform-specific policies',
+    ]
+
+    const isSafe = !content.toLowerCase().includes('escalate')
+    const confidence = isSafe ? 0.92 : 0.45
+
+    return { isSafe, confidence, reasoning }
+  },
+}
+
+// Mock human review queue
+const humanReviewQueue: HumanReviewTask[] = []
+
+// ============================================================================
+// Cascade Executor Configuration
+// ============================================================================
+
+function createModerationCascade() {
+  const events: Array<{ timestamp: number; event: string; data: unknown }> = []
+
+  const cascade = new CascadeExecutor<ModerationResult>({
+    cascadeName: 'content-moderation',
+    actor: 'moderation-system',
+
+    tiers: {
+      // Tier 1: Rule-based (fastest, deterministic)
+      code: {
+        name: 'rule-based-filter',
+        execute: async (input: unknown): Promise<ModerationResult> => {
+          const content = (input as ContentItem).content.toLowerCase()
+
+          // Check blocked keywords
+          for (const keyword of BLOCKED_KEYWORDS) {
+            if (content.includes(keyword)) {
+              return {
+                action: 'reject',
+                confidence: 1.0,
+                reason: `Contains blocked keyword: ${keyword}`,
+                categories: ['blocked_content'],
+              }
+            }
+          }
+
+          // Check suspicious patterns
+          for (const pattern of SUSPICIOUS_PATTERNS) {
+            if (pattern.test(content)) {
+              return {
+                action: 'reject',
+                confidence: 0.95,
+                reason: 'Matches suspicious pattern',
+                categories: ['spam_pattern'],
+              }
+            }
+          }
+
+          // Quick approve obviously safe content
+          if (content.length < 50 && /^[a-z\s.,!?]+$/.test(content)) {
+            return {
+              action: 'approve',
+              confidence: 0.99,
+              reason: 'Simple safe content',
+            }
+          }
+
+          // Escalate to AI for more complex content
+          throw new Error('Content requires AI analysis')
+        },
+      },
+
+      // Tier 2: Generative AI (nuanced analysis)
+      generative: {
+        name: 'ai-content-analysis',
+        execute: async (input: unknown): Promise<ModerationResult> => {
+          const contentItem = input as ContentItem
+          const analysis = await aiService.analyzeContent(contentItem.content)
+
+          // High confidence safe - approve
+          if (analysis.isSafe && analysis.confidence > 0.9) {
+            return {
+              action: 'approve',
+              confidence: analysis.confidence,
+              reason: analysis.explanation,
+              categories: analysis.categories,
+            }
+          }
+
+          // High confidence unsafe - reject
+          if (!analysis.isSafe && analysis.confidence > 0.9) {
+            return {
+              action: 'reject',
+              confidence: analysis.confidence,
+              reason: analysis.explanation,
+              categories: analysis.categories,
+            }
+          }
+
+          // Low confidence - escalate for deeper analysis
+          throw new Error(`Confidence too low (${analysis.confidence}) - needs deeper analysis`)
+        },
+      },
+
+      // Tier 3: Agentic AI (multi-step reasoning)
+      agentic: {
+        name: 'agentic-deep-analysis',
+        execute: async (input: unknown, context: TierContext): Promise<ModerationResult> => {
+          const contentItem = input as ContentItem
+          const analysis = await agenticService.deepAnalysis(contentItem.content, context)
+
+          console.log('  [Agentic] Reasoning steps:', analysis.reasoning)
+
+          // High confidence from agent
+          if (analysis.confidence > 0.85) {
+            return {
+              action: analysis.isSafe ? 'approve' : 'reject',
+              confidence: analysis.confidence,
+              reason: analysis.reasoning.join(' -> '),
+            }
+          }
+
+          // Still uncertain - needs human
+          throw new Error('Agent uncertain - requires human review')
+        },
+      },
+
+      // Tier 4: Human review (final decision)
+      human: {
+        name: 'human-review',
+        execute: async (input: unknown): Promise<ModerationResult> => {
+          const contentItem = input as ContentItem
+
+          // Create review task
+          const task: HumanReviewTask = {
+            contentId: contentItem.id,
+            content: contentItem.content,
+            aiAnalysis: 'AI was unable to reach high confidence decision',
+            priority: 'high',
+          }
+
+          humanReviewQueue.push(task)
+          console.log(`  [Human] Created review task for content ${contentItem.id}`)
+
+          // For demo, simulate immediate human decision
+          return {
+            action: 'flag',
+            confidence: 1.0,
+            reason: 'Flagged for human review - decision pending',
+            requiresReview: true,
+          }
+        },
+      },
+    },
+
+    // Timeouts per tier
+    timeouts: {
+      code: 100, // 100ms for rules
+      generative: 5000, // 5s for AI
+      agentic: 30000, // 30s for agent
+      human: 3600000, // 1 hour for human
+    },
+
+    // Retry configuration
+    retryConfig: {
+      generative: { maxRetries: 2, baseDelay: 100, multiplier: 2 },
+      agentic: { maxRetries: 1, baseDelay: 500 },
+    },
+
+    // Skip conditions
+    skipConditions: {
+      // Skip agentic for very short content
+      agentic: (input) => (input as ContentItem).content.length < 20,
+    },
+
+    // Event callback for audit trail
+    onEvent: (event) => {
+      events.push({
+        timestamp: event.when,
+        event: event.what,
+        data: {
+          status: event.how.status,
+          duration: event.how.duration,
+          ...(event.why && { reason: event.why }),
+        },
+      })
+    },
+  })
+
+  return { cascade, events, humanReviewQueue }
+}
+
+// ============================================================================
+// Workflow Integration
+// ============================================================================
+
+async function runModerationWorkflow() {
+  const { cascade, events, humanReviewQueue } = createModerationCascade()
+
+  const workflow = Workflow(($) => {
+    $.on.Content.submitted(async (content: ContentItem, $: WorkflowContext) => {
+      $.log(`Processing content ${content.id}`)
+
+      try {
+        const result = await cascade.execute(content)
+
+        $.log(`Moderation complete: ${result.tier} tier -> ${result.value.action}`)
+
+        // Emit result event
+        $.send('Content.moderated', {
+          contentId: content.id,
+          result: result.value,
+          resolvedBy: result.tier,
+          metrics: result.metrics,
+        })
+
+        // Handle based on result
+        if (result.value.action === 'approve') {
+          $.send('Content.published', { contentId: content.id })
+        } else if (result.value.action === 'reject') {
+          $.send('Content.rejected', {
+            contentId: content.id,
+            reason: result.value.reason,
+          })
+        }
+      } catch (error) {
+        $.log(`Moderation failed: ${error}`)
+        $.send('Content.moderationFailed', {
+          contentId: content.id,
+          error: error instanceof Error ? error.message : String(error),
+        })
+      }
+    })
+
+    $.on.Content.published(async (data: { contentId: string }, $: WorkflowContext) => {
+      $.log(`Content ${data.contentId} published successfully`)
+    })
+
+    $.on.Content.rejected(
+      async (data: { contentId: string; reason: string }, $: WorkflowContext) => {
+        $.log(`Content ${data.contentId} rejected: ${data.reason}`)
+      }
+    )
+  })
+
+  await workflow.start()
+  return { workflow, cascade, events, humanReviewQueue }
+}
+
+// ============================================================================
+// Demo Execution
+// ============================================================================
+
+async function runDemo() {
+  console.log('='.repeat(60))
+  console.log('AI Content Moderation with Human Escalation Demo')
+  console.log('='.repeat(60))
+  console.log()
+
+  const { workflow, events } = await runModerationWorkflow()
+
+  // Test cases demonstrating different tiers
+  const testContent: ContentItem[] = [
+    // Case 1: Blocked keyword - handled by code tier
+    {
+      id: 'content_001',
+      type: 'text',
+      content: 'Check out this amazing spam offer!',
+      metadata: { userId: 'user_1', platform: 'web', timestamp: Date.now() },
+    },
+
+    // Case 2: Simple safe content - handled by code tier
+    {
+      id: 'content_002',
+      type: 'text',
+      content: 'Hello, how are you today?',
+      metadata: { userId: 'user_2', platform: 'mobile', timestamp: Date.now() },
+    },
+
+    // Case 3: Complex content - needs AI analysis
+    {
+      id: 'content_003',
+      type: 'text',
+      content:
+        'This is a longer piece of content that discusses various topics in a nuanced way. ' +
+        'It requires deeper analysis to determine if it meets community guidelines.',
+      metadata: { userId: 'user_3', platform: 'web', timestamp: Date.now() },
+    },
+
+    // Case 4: Suspicious content - may need agent
+    {
+      id: 'content_004',
+      type: 'text',
+      content:
+        'This content has suspicious elements that the AI flagged but cannot ' +
+        'determine with high confidence. It requires multi-step reasoning.',
+      metadata: { userId: 'user_4', platform: 'web', timestamp: Date.now() },
+    },
+
+    // Case 5: Needs escalation to human
+    {
+      id: 'content_005',
+      type: 'text',
+      content:
+        'This complex content contains patterns that need to escalate ' +
+        'all the way to human review for a final decision.',
+      metadata: { userId: 'user_5', platform: 'web', timestamp: Date.now() },
+    },
+  ]
+
+  // Process each content item
+  for (const content of testContent) {
+    console.log('-'.repeat(60))
+    console.log(`Processing: ${content.id}`)
+    console.log(`Content: "${content.content.substring(0, 50)}..."`)
+    console.log()
+
+    await workflow.send('Content.submitted', content)
+
+    // Small delay between items
+    await new Promise((resolve) => setTimeout(resolve, 200))
+  }
+
+  // Display audit trail
+  console.log()
+  console.log('='.repeat(60))
+  console.log('Audit Trail (5W+H Events):')
+  console.log('='.repeat(60))
+  for (const event of events.slice(-10)) {
+    console.log(`  ${new Date(event.timestamp).toISOString()} | ${event.event}`)
+    console.log(`    Data: ${JSON.stringify(event.data)}`)
+  }
+
+  // Clean up
+  await workflow.stop()
+}
+
+// Run if executed directly
+runDemo().catch(console.error)