OpenAI ChatGPT API integration with per-guild model selection and modular prompt system.

Model Selection

API endpoint: https://api.openai.com/v1/models

Model filtering:
Bot filters available models to chat-capable models only.

Exclusion patterns:

• Embedding models: /(embedding|text-embedding)/i
• Audio models: /(whisper|tts|audio|transcribe)/i
• Image models: /(dall|image|vision|clip)/i
• GPT-3 variants: /^gpt-3/i
• Turbo variants: /turbo/i
• Preview variants: /preview/i
• Fine-tuned models: /ft:/i
• Omni/sprite models: /(omni|sprites)/i

Inclusion pattern:
Models must match: /^(gpt|o[34]|gpt-5(?:.d+)?)/i

Selection interface:

• Administrator runs /model command
• Bot fetches and filters models
• Dropdown shows up to 25 chat-capable models
• Selection saved per Discord guild

Storage:
data/model_settings.json – Guild ID to model ID mapping

Model operations:

• getModelForGuild(guildId) – Returns configured model ID for guild
• setModelForGuild(guildId, modelId) – Saves model selection

Validation:
No validation beyond filtering. Assumes filtered models support chat completions API.

Modular System Prompt

Prompt assembly:
System prompt built from multiple .txt files in data/ directory.

Loading process:

1. Scan data/ directory for .txt files
2. Read each file’s content
3. Sort files alphanumerically
4. Concatenate with double newline separator
5. Store in memory

File ordering:
Use numeric prefixes for consistent ordering:

• 00_base_instructions.txt
• 10_style_parameters.txt
• 20_technical_details.txt
• 30_constraints.txt

Dynamic reloading:

• Loads on bot startup
• Reloads after /editprompt saves changes
• No bot restart required

Fallback behavior:
If no .txt files exist or readable:

• Falls back to MJPIN_OPENAI_SYSTEM_PROMPT environment variable
• If environment variable missing, uses default: "You are a helpful AI that generates Midjourney prompts."

Section editing:
Via /editprompt command:

• Lists all .txt files as selectable sections
• Opens modal editor for selected file
• Saves changes directly to file
• Triggers automatic reload

Chat Completions

API endpoint: https://api.openai.com/v1/chat/completions

Request format:

{
  "model": "guild_specific_model_id",
  "messages": [
    {
      "role": "system",
      "content": "assembled_system_prompt"
    },
    {
      "role": "user",
      "content": "user_input_from_prompt_command"
    }
  ]
}

Authentication:
Bearer token in Authorization header from MJPIN_OPENAI_API_KEY environment variable.

Message structure:
Two-message conversation:

1. System message with assembled prompt
2. User message with input from /prompt command

Response handling:

• Extracts content from response.data.choices[0].message.content
• Truncates to 2000 characters if longer (Discord limit)
• Returns truncated text with ellipsis

Error handling:

• Extracts error from response.data.error.message
• Falls back to generic error.message
• Prefixes with "OpenAI API error: "
• Returns sanitized message to user

Configuration

Required environment variables:

• MJPIN_OPENAI_API_KEY – OpenAI API bearer token

Optional environment variables:

• MJPIN_OPENAI_SYSTEM_PROMPT – Fallback system prompt if no .txt files

Prerequisites:

• Guild administrator must configure model via /model command before first /prompt use
• System prompt files should exist in data/ directory (user-created)

Initialization:

• System prompt loads automatically on bot startup
• Model selection required per guild before prompt generation
• No global model default

Prompt Generation Workflow

User perspective:

1. User runs /prompt input:"describe desired image"
2. Bot defers reply (operation may exceed 3 seconds)
3. Bot retrieves guild’s configured model
4. Bot retrieves assembled system prompt
5. Bot sends request to OpenAI API
6. Bot returns generated prompt to user

Error scenarios:

• No model configured: "OpenAI model is not configured for this server. Ask an admin to run /model to set it."
• No API key: "OpenAI API key not set in environment."
• API failure: "OpenAI API error: [sanitized_error_message]"

Response formatting:

• Full response if under 2000 characters
• Truncated with "…" if over 2000 characters
• No markdown formatting applied by bot