Providers and models¶
eYssen AI talks to language-model providers over their web APIs. This page explains what to configure once, so that every other AI feature (chat, writing assistant, knowledge sources, voice, images) has a model to work with.
Providers¶
Open and create one provider per account you use. The supported types are OpenRouter, OpenAI, Anthropic, xAI and Google Gemini; any OpenAI-compatible gateway works through the Base Url field. The API key is write-only: once saved it is never shown again.
A hosted provider needs its API key. If it has none, the request fails at once
with the message “The AI provider has no API key. Set it under AI >
Configuration > Providers.” instead of being sent and rejected by the provider.
A self-hosted endpoint such as Ollama, LM Studio or vLLM needs no key. That is
the case when the Base Url points to the local machine
(localhost), to a private network address, to a name ending in .local,
.localhost, .internal or .lan, or to a plain host name without dots,
such as http://ollama:11434/v1.
Click Refresh Models on the provider to import its model catalogue (and prices where the provider publishes them). Refreshing never overwrites a price you entered by hand.
Kinds of models¶
Every model has a Kind. The kind decides where the model can be chosen:
Chat models answer conversations and drive the assistant, the writing assistant and all the one-click features. Choose one under in AI Model. The assistant calls tools, so pick a model that supports tool calling.
Image generation models create pictures for the AI panel, the editor and the product and Knowledge cover features. Set the model by hand to this kind, then pick it in Image Model under . Anthropic offers no image models; use another provider for images.
Embedding models turn text into vectors for knowledge sources and similar tasks (see Knowledge sources). OpenAI, Google Gemini and OpenRouter provide them; xAI and Anthropic do not.
Speech to text models transcribe audio for dictation and call transcription (see Voice: dictation and call transcription). Any provider with an OpenAI-compatible audio transcription route works, including a custom base URL such as Groq or a self-hosted Whisper server. Anthropic and OpenRouter are refused.
What Refresh Models does for each kind¶
Models whose id contains
embedare filed as Embedding automatically and never offered to the chat loop. OpenRouter lists embedding models on a separate endpoint, so a refresh of an OpenRouter provider also fetches that list and adds those models; if that second request fails, the normal refresh still succeeds.Models whose id contains
whisperortranscribeare filed as Speech to text.Image models are not detected: open the model and set its Kind to Image generation yourself. A refresh never changes a kind you set by hand.
Where each model is chosen¶
AI Model and Image Model: , block Model.
Embedding model: same page, block Knowledge sources.
Speech-to-text model: same page, block Voice.
Reasoning and cost¶
Only models flagged as reasoning-capable react to Think longer in the AI panel (see The AI panel). Flag a model only if its API documents a reasoning effort or a thinking budget: other models reject the extra parameter.
Each model has unit prices (per million tokens, per image or per audio minute). They feed the cost cap in the settings and the log entries. A model without prices is not blocked, but the cost cap cannot count it.
Important
Everything a model is asked is sent to the provider you chose. Pick providers whose data terms fit your obligations, and keep the log level on Ai Log Level metadata only unless you need payloads.