Skip to content

Settings: Modules

RPG Roleplay uses several internal AI sub-modules, each responsible for a distinct task — primary narration, intent parsing, vector embedding, image generation, and more. Assign a specific model to any module, or set your personal default GM model. Modules without an explicit override inherit from the primary GM.

Navigation: top bar SettingsModule Models (left sidebar).


Key Concepts

Primary GM Default Model

The first row in the list is the Primary GM Default Model. It is saved to your personal preferences and applies to all new sessions and any saved game that has not had its model individually changed. All other modules inherit from this model unless overridden.

Module Override

Each module row has its own model selector. Selecting Follow Primary GM (or leaving it blank) clears the override and restores inheritance. The vector embedding and image generation modules are exceptions — they require an explicit selection and do not support inheritance.

Model List Source

The selector only lists models from providers you have configured in API Keys that have been successfully synced. If a model is missing, go to API Keys and confirm that the corresponding provider key has been added correctly.


Module Reference

ModulePurposeRecommended tier
Primary GM DefaultCore narrative model for player conversationFlagship
Context sub-agentStructures player intent and retrieval plans from ambiguous inputMid-range
Command parsing agentTranslates /set natural language commands into structured operationsMid-range or small
Console assistantDedicated model for the admin sidebar consoleSame tier as primary GM
Narrative extractorSecondary parse of GM output to extract state operations (step 2 of two-step GM mode)Cheap/small
Character card generatorCreative tool in the sidebar for generating or refining character cardsMid-range to flagship
AI field organizerUsed when importing Tavern cards with Organize fields with AI checked — converts free text into structured fieldsMid-range or small
Consistency scorerScores character card generation output for consistency (0–1, threshold 0.6)Cheap/small
Acceptance validatorChecks whether GM output satisfies curator acceptance conditionsCheap/small
Compact summarizerCondenses long session history into phase summaries for GM long-term recallMid-range or small
Black swan event agentSub-agent that proactively triggers world eventsMid-range
Generic sub-agent fallbackUsed by any sub-agent that does not have a dedicated model assignedMid-range
Vector embedding (RAG)Embedding model for memory retrieval and semantic search after book parsingDedicated embedding model
Image generationAll AI image generation — chat images, character avatars, script covers, character sheetsDedicated image model

Common Tasks

Change the Primary GM Default Model

  1. Go to Settings → Module Models.
  2. Find the Primary GM Default Model row and click the selector on the right.
  3. Choose a model from the list. The change saves immediately and takes effect the next time a new session is created or an existing session without a model override is loaded.

Assign a Model to a Specific Module

  1. Find the module row and click its selector.
  2. Choose from the list of synced models or type a model ID directly.
  3. The change saves immediately — no confirmation button is needed.

Clear a Module Override (Restore Inheritance)

  1. Find the module’s selector and choose Follow Primary GM (usually the first item or the blank option).
  2. The change takes effect immediately. The module will track future changes to the primary GM model automatically.

Configure the Vector Embedding Model

The vector embedding module requires a model that supports the embedding API — standard chat models are not compatible.

  1. In API Keys, add a key for a provider that offers embedding (e.g. Vertex AI, OpenAI, Cohere).
  2. In the Vector Embedding (RAG) row, select the appropriate embedding model.
  3. Note: switching the embedding model requires re-embedding any scripts that were previously processed. Old vectors are not compatible with the new model.

Users without an embedding key will see RAG retrieval fall back to a degraded mode. Admin and VIP users have a platform fallback (Gemini free tier); the configuration page shows the current fallback status below that row.

Configure the Image Generation Model

  1. In API Keys, add a key for a provider that supports image generation (e.g. Vertex AI, Doubao, Tongyi).
  2. In the Image Generation row, select the target model. This selection is synchronized with the model picker in the image generation dialog.
  3. If no image generation key is configured, RPG Roleplay prompts you to add one when you attempt to generate an image — it does not fail silently.

FAQ

Q: The model I want isn’t in the list. A: The list only includes models from providers with a successfully synced key. Go to API Keys, add the key for that provider, and click Fetch Models to sync.

Q: What is the narrative extractor and do I need it? A: The narrative extractor is the second step of the two-step GM mode. The primary GM generates narrative text; the extractor then parses that text into structured state operations. Enabling it improves state consistency but increases API cost by roughly 20% per turn. The default single-step GM is sufficient for most users — consider enabling two-step mode only if you encounter persistent state inconsistencies.

Q: When do module model changes take effect? A: Changes are saved immediately. Each backend module reads its configuration on the next invocation. Active conversations are not interrupted — the new model applies starting from the next message.

Q: Can Anthropic or DeepSeek models be used for vector embedding? A: No. Neither provider offers an embedding API, so neither appears in the vector embedding model selector.

Q: What is the generic sub-agent fallback for? A: When a sub-agent has no dedicated model assigned, the platform first checks the primary GM setting. The generic sub-agent fallback provides an additional middle layer for unified management of sub-agents without individual configurations. In most cases, leaving it to follow the primary GM is sufficient.