AI providers
These parts of Gryffi use a language model:
- AI guides, in journeys and on their own
- the journey wizard, which drafts a journey from a document
- Copy & Translate
- descriptions of images, so guides can talk about them (optional)
- reading steps aloud with AI voices (optional; without it, the browser reads aloud)
- the help assistant for makers
Everything else works without AI. You can leave AI off in the setup wizard and add it later on the Installation page.
Any OpenAI-compatible API
Gryffi talks to the OpenAI API format, which most providers and local model servers support. You need the API address, a key, a chat model and an embedding model.
| Provider | API address | Notes |
|---|---|---|
| Mistral | https://api.mistral.ai/v1 | Processes data in the EU. Embedding model mistral-embed |
| OpenAI | https://api.openai.com/v1 | Embedding model text-embedding-3-small |
| OpenRouter | https://openrouter.ai/api/v1 | Has no embeddings: set a separate embedding provider |
| Ollama | http://host.docker.internal:11434/v1 | On the same server. Any API key works. Embedding model for example nomic-embed-text |
| vLLM, LM Studio and others | The address of their OpenAI-compatible endpoint |
Press Test connection: Gryffi lists the models of your provider, so you can pick the chat model from that list.
Gryffi runs in a container. localhost there is the container itself, not your server. For a model server on the same machine use host.docker.internal, and make sure that server listens on an address the container can reach (for Ollama: OLLAMA_HOST=0.0.0.0).
Embeddings
AI guides search your documents with embeddings. Gryffi stores vectors of up to 1536 numbers; models with shorter vectors work too. Models with longer vectors, such as text-embedding-3-large, do not.
When you change the embedding model or provider, Gryffi rebuilds the index of all knowledge sources. The Installation page shows the progress. Until it is done, guides find fewer answers.
Costs and limits
You pay your provider for what Gryffi uses. To keep costs in hand, set Maximum AI questions per month on the Installation page. Admins get an email at 80% and at 100%.
What goes to the provider
Questions of participants, the matching parts of your documents, and the text of journeys you translate or generate go to the provider you chose. Choose a provider whose terms fit your data, or run models yourself.