Skip to main content
YBH Pulse Content uses multiple AI providers for different content generation tasks. Google Gemini is the primary provider, with Anthropic Claude as a fallback.

​
Supported providers

​
Google Gemini (primary)

Model: gemini-2.0-flash-exp Used for:
  • PRF (Podcast Repurposing Framework) generation
  • Viral hooks extraction
  • Episode metadata generation
  • LinkedIn and Instagram post creation
  • Visual suggestions (infographics, quote cards)
  • Fact-checking and content validation
  • Spell checking
Capabilities:
  • Agentic tool calling
  • Structured output (JSON mode)
  • Long context window (over 1M tokens)
  • Fast response times
  • Cost-effective for high-volume generation
Get your API key: https://aistudio.google.com/apikey

​
Anthropic Claude (fallback)

Model: claude-3-5-sonnet-20241022 Used for:
  • All generation tasks when Google key is not available
  • Same capabilities as Google Gemini
Format: sk-ant-... Get your API key: https://console.anthropic.com/

​
OpenRouter (embeddings)

Model: text-embedding-3-small via OpenRouter Used for:
  • Vector embeddings for Pinecone RAG
  • Semantic search across brand guidelines
  • Infographic design database retrieval
Format: sk-or-... Get your API key: https://openrouter.ai/keys

​
Configuration

1

Add API keys to .dev.vars

# Primary AI provider
GOOGLE_GENERATIVE_AI_API_KEY=your-google-api-key-here

# Fallback AI provider
ANTHROPIC_API_KEY=sk-ant-your-key-here

# Embeddings provider
OPENROUTER_API_KEY=sk-or-your-key-here
2

Set production secrets

wrangler secret put GOOGLE_GENERATIVE_AI_API_KEY
wrangler secret put ANTHROPIC_API_KEY
wrangler secret put OPENROUTER_API_KEY
3

Test connection

Generate a PRF from any episode to verify AI providers are working.

​
AI generation endpoints

​
PRF generation

Endpoint: POST /api/ai/prf Input: 
{
  "transcript": "Episode transcript text...",
  "model": "gemini-2.0-flash-exp",
  "systemPrompt": "Generate a PRF document..."
}
Output: 
{
  "content": "PRF HTML content..."
}
Usage: 
import { generatePRF } from '@/services/ai'

const prf = await generatePRF(transcript)

​
Viral hooks generation

Endpoint: POST /api/ai/hooks Requires: Transcript for fact verification Input: 
{
  "prf": "PRF content with transcript...",
  "model": "gemini-2.0-flash-exp",
  "systemPrompt": "Generate viral hooks..."
}
Output: 
{
  "content": "Hooks HTML content..."
}
Usage: 
import { generateHooks } from '@/services/ai'

const hooks = await generateHooks(
  transcript,
  prf,
  episodeNumber,
  guestName
)

​
Episode metadata (agentic)

Endpoint: POST /api/agent/metadata Uses: Server-sent events (SSE) for progress updates Input: 
{
  "episodeId": "episode-123",
  "episodeNumber": "385",
  "guestName": "Chris Pacifico",
  "transcript": "Full transcript...",
  "prf": "Optional PRF content..."
}
Output (streamed): 
// Status event
{
  "type": "status",
  "step": "Analyzing transcript",
  "detail": "Extracting key themes...",
  "progress": 30
}

// Result event
{
  "type": "result",
  "content": {
    "title": "Episode 385: Chris Pacifico on Tech Debt Management",
    "shortDescription": "Chris shares...",
    "longDescription": "On this episode...",
    "keyTakeaways": ["Takeaway 1", "Takeaway 2", "Takeaway 3"],
    "showNotes": [
      { "timestamp": "00:00", "description": "Introduction" },
      { "timestamp": "05:30", "description": "Tech debt discussion" }
    ]
  }
}
Usage: 
import { generateEpisodeMetadata } from '@/services/ai'

const metadata = await generateEpisodeMetadata(
  episodeId,
  episodeNumber,
  guestName,
  transcript,
  prf,
  (step, detail, progress) => {
    console.log(`${progress}%: ${step} - ${detail}`)
  }
)

​
Visual suggestions (parallel generation)

Endpoint: POST /api/agent/suggestions Uses: Server-sent events (SSE) for progress updates Generates: 4 Data Viz + 4 Cinematic + 2 Quote Cards = 10 total Input: 
{
  "episodeId": "episode-123",
  "episodeNumber": "385",
  "guestName": "Chris Pacifico",
  "transcript": "Full transcript...",
  "prf": "PRF content...",
  "hooks": "Viral hooks...",
  "history": ["previous-style-1", "previous-style-2"],
  "datavizPrompt": "Custom prompt for data viz...",
  "cinematicPrompt": "Custom prompt for cinematic...",
  "quoteCardsPrompt": "Custom prompt for quote cards..."
}
Output: 
{
  "suggestions": [
    {
      "visualType": "dataviz",
      "type": "infographic",
      "sourceText": "Selected text from PRF...",
      "sourceSection": "PRF Section 2",
      "spec": {
        "layout": "Card Grid",
        "template": "Strategic Framework",
        "title": "Tech Debt Management Framework",
        "colorSystem": "Corporate Professional",
        "iconStyle": "isometric",
        "aspectRatio": "16:9",
        "prompt": "Full Kie.ai prompt..."
      }
    }
  ],
  "counts": {
    "dataviz": 4,
    "cinematic": 4,
    "quoteCard": 2
  },
  "errors": {
    "dataviz": null,
    "cinematic": null,
    "quoteCard": null
  }
}

​
Fact-checking

Endpoint: POST /api/agent/fact-check Purpose: Verify generated content against transcript to prevent hallucinations Input: 
{
  "transcript": "Full transcript...",
  "prf": "PRF content...",
  "guestName": "Chris Pacifico",
  "coHostName": "Doug",
  "items": [
    {
      "type": "quote",
      "content": "Tech debt is like credit card debt",
      "source": "Chris Pacifico"
    },
    {
      "type": "statistic",
      "content": "70% of IT budgets go to maintenance",
      "source": "Chris Pacifico"
    }
  ]
}
Output: 
{
  "overallScore": 85,
  "passedValidation": true,
  "results": [
    {
      "item": { "type": "quote", "content": "..." },
      "status": "verified",
      "confidence": 95,
      "transcriptEvidence": "Exact quote from transcript..."
    }
  ],
  "summary": "2 items verified, 0 issues found",
  "criticalIssues": []
}

​
Model selection

The system automatically selects the best available model: 
function getModel() {
  if (env.GOOGLE_GENERATIVE_AI_API_KEY) {
    return 'gemini-2.0-flash-exp'
  }
  if (env.ANTHROPIC_API_KEY) {
    return 'claude-3-5-sonnet-20241022'
  }
  throw new Error('No AI provider configured')
}

​
Cost optimization

​
Token usage

Google Gemini pricing (as of 2024):
  • Input: Low cost per 1M tokens
  • Output: Low cost per 1M tokens
  • Flash models optimized for speed and cost
Anthropic Claude pricing:
  • Input: Higher cost per 1M tokens
  • Output: Higher cost per 1M tokens
  • Higher quality for complex reasoning

​
Reduce costs

1

Truncate long transcripts

For visual suggestions, only send first 3000 characters of transcript:
const truncated = transcript.slice(0, 3000)
const context = truncated.length < transcript.length
  ? `${truncated}\n[...truncated for context window...]`
  : truncated
2

Cache RAG results

Store Pinecone query results to avoid repeated embedding calls:
const cacheKey = `rag:${query}`
const cached = await cache.get(cacheKey)
if (cached) return cached

const results = await queryKnowledgeBase(query)
await cache.set(cacheKey, results, 3600) // 1 hour
return results
3

Batch generations

Generate multiple visual suggestions in parallel to amortize prompt costs:
// Generates 10 suggestions in 3 parallel calls
await generateVisualSuggestionsV2(...)
4

Use Google Gemini as primary

Google Gemini Flash models are significantly cheaper than Claude for equivalent quality on most tasks.

​
Prompt management

Prompts are stored in the Settings page and can be customized per agent:

​
Agent types

  • PRF - Podcast Repurposing Framework generation
  • Hooks - Viral hooks extraction
  • LinkedIn - LinkedIn post creation
  • Instagram - Instagram post creation
  • Infographic - Infographic spec generation
  • Thumbnail - YouTube thumbnail spec generation
  • Timeline - Career timeline spec generation
  • Suggestions - Visual suggestions orchestration

​
Preset fields

Prompts support variable interpolation:
  • {{episodeNumber}} - Episode number
  • {{guestName}} - Guest name
  • {{recentStyles}} - Recent visual styles for variety
  • {{brandVoice}} - YBH brand voice guidelines
  • {{targetAudience}} - IT leaders (CIOs, CTOs, IT Directors)

​
Edit prompts

  1. Navigate to Settings > AI Configuration
  2. Select agent type
  3. Edit system prompt
  4. Save changes
  5. Test with new episode
Changes to prompts affect all future generations. Test thoroughly before deploying to production.

​
Error handling

​
Retry logic

The system automatically retries failed generations: 
try {
  const prf = await generatePRF(transcript)
} catch (error) {
  if (error.message.includes('rate limit')) {
    // Wait and retry
    await sleep(5000)
    return generatePRF(transcript)
  }
  throw error
}

​
Fallback behavior

  1. Try Google Gemini (if key available)
  2. Fall back to Anthropic Claude (if key available)
  3. Return error if both fail

​
Common errors

Rate limit exceeded: 
Error: Rate limit exceeded. Please try again in 60 seconds.
Solution: Wait and retry, or upgrade API plan. Invalid API key: 
Error: API key is invalid or expired.
Solution: Regenerate API key from provider dashboard. Context too long: 
Error: Input exceeds maximum context length.
Solution: Truncate transcript or split into chunks.

​
Monitoring

Track AI usage in application logs: 
import { logger } from '@/utils/logger'

logger.debug('[AI] Generating PRF', { episodeId, model })
const prf = await generatePRF(transcript)
logger.info('[AI] PRF generated', { episodeId, length: prf.length })

​
Best practices

  • Always verify facts: Use fact-checking endpoint before approving content
  • Include transcript: Pass transcript to all generation endpoints for accuracy
  • Monitor token usage: Track costs across Google, Anthropic, and OpenRouter
  • Test prompts: Validate prompt changes on multiple episodes before deployment
  • Cache aggressively: Store expensive generation results in Sanity
  • Use streaming: Implement SSE for long-running operations to show progress

Build docs developers (and LLMs) love

Get started for freeTalk to us