Generation Pipeline

The custom course generation runs through three n8n workflows on worker-1, each handling a distinct stage of the pipeline.

Purpose

Transform user inputs (issue, goal, feel, why) into a personalized hypnotherapy audio session through AI-powered script generation with safety guardrails.

Architecture

All workflows run on n8n worker-1 (Digital Ocean) and are triggered via webhook by the submit-custom-course edge function.

User Input → Health Check → Script Generation → Safety Check → Audio Generation

Stage 1: Health Check

Webhook: /webhook/api/v1/custom-course-health-check Model: mistral-medium Purpose: Validate that user inputs are safe and appropriate for hypnotherapy before generating scripts.

What It Checks

• Safety — Flags self-harm, medical conditions, dangerous content
• Clarity — Are the inputs specific enough to generate a good script?
• Rewrite suggestions — If inputs are vague or problematic, suggests improvements

Response Format

{
  "success": true,
  "blocked": false,
  "data": {
    "is_safe": true,
    "is_clear": true,
    "needs_rewrite": false,
    "safety_category": "NONE",
    "reasoning": "..."
  }
}

Fail-Safe

If the health check workflow is unreachable or returns an error, the system defaults to allowing the request through (doesn’t block the user due to workflow issues).

Cost

~$0.0007 per check (~768 input tokens, ~219 output tokens)

Stage 2: Script Generation

Webhook: /webhook/api/v1/custom-script Model: ft:mistral-medium (fine-tuned on HypnoELP course data) Purpose: Generate three personalized hypnotherapy scripts using different techniques.

Input

{
  "issue": "...",
  "goal": "...",
  "feel": "...",
  "why": "...",
  "title": "...",
  "scripts": [
    {"number": 1, "technique": "REHEARSAL"},
    {"number": 2, "technique": "FUTURE_VISUALIZATION"},
    {"number": 3, "technique": "EMOTIONAL_LEVERAGE"}
  ],
  "user_id": "..."
}

Output

Three complete hypnotherapy scripts (script1, script2, script3) plus a safety object from the post-generation safety check.

Safety Check (Post-Generation)

After generating scripts, a safety check runs using mistral-small to verify the generated content is appropriate. If any script fails, the entire request is rejected.

Cost

~$0.0026 per generation (~450 input tokens, ~1,200 output tokens)

Stage 3: Audio Generation

Webhook: /webhook/api/v1/custom-audio Model: ElevenLabs TTS Purpose: Convert the selected script into a professional hypnotherapy audio file.

Input

{
  "text": "full script text",
  "voice": "male" | "female",
  "title": "session title",
  "client_id": "user_id"
}

Output

Audio file URL hosted on Cloudflare R2 (hypnoelp-custom-files bucket).

Voices

• Male and Female options via ElevenLabs TTS

Available Techniques

Total Cost Per Custom Course

Audio generation cost (ElevenLabs) is separate and varies by script length.

Costs are logged to the custom_course_costs table on the HypnoELP-AI Supabase project (yxudnnytcjuafeclsofg).

Handshakes

Dependencies

• n8n (workflow orchestration)
• Mistral API (LLM — small, medium, fine-tuned)
• ElevenLabs API (TTS)
• Cloudflare R2 (audio file storage)
• HypnoELP-AI Supabase (cost logging)

Status

Production. All three stages operational. Fine-tuned Mistral V2 model deployed. V3 retraining pending (4 courses missing from training data).