Models

Kizuki uses two models: an answer model that proposes concepts and links and chooses questions, and a meaning-search model that turns passages into numbers for search. By default both run in Ollama on your computer, so your material never leaves it. For bigger models without a fast computer, Kizuki can use Vercel AI Gateway instead, after asking you.

Settings

DEFAULT_SETTINGS points both at Ollama, http://localhost:11434/v1, with qwen3.5:2b for answers and nomic-embed-text for meaning search. Settings live in settings.json in the data folder (settingsSchema):

• provider for each model: openai-compatible for any server that speaks the OpenAI format (Ollama, MLX), or gateway for Vercel AI Gateway. Settings saved before this field existed read as openai-compatible.
• baseURL and model for each model.
• apiKeyEnv: the name of an environment variable that holds an API key, ending in _API_KEY. Kizuki refuses anything that is not such a name, so a key is never saved. chatModel reads the key from the environment at the moment it calls.
• reasoning: none turns thinking off, which is much faster on small models. default leaves it to the model.
• replyShape: server lets the server enforce the reply's JSON shape (Ollama, AI Gateway). prompt makes Kizuki describe the shape in the instructions and check the reply itself (MLX).
• sendOutAllowed: true once you agreed to send your material to a model off your computer. See below.

readSettings treats a broken settings.json as an error that names the file, never as a silent reset to the defaults.

Changing models

/model shows both models and their addresses. PRESETS holds three ready-made choices, each switching both models at once:

/model answer <name> and /model search <name> change one model's name and keep its address. With AI Gateway, any model name the gateway lists works, such as anthropic/claude-sonnet-5.5.

Meaning-search numbers from one model can't be compared with another's. When the meaning-search model changes, Kizuki says the search file is out of date; /rebuild rebuilds it from the logs (rebuildIndex).

Checking the models

checkModels asks each server for its model list and reports a problem with a fix: "Start Ollama with: ollama serve", "Install it with: ollama pull qwen3.5:2b", or the name of an API key variable that is not set. For AI Gateway it asks the gateway's own model list. Kizuki checks at start and after every /model change, prints any problems, and keeps going, so you can still review and browse without a working model.

Talking to the models

Kizuki calls the models through the AI SDK:

• For openai-compatible servers it uses the OpenAI-compatible provider. makeAsk sends POST {baseURL}/chat/completions with temperature 0.2 and the reply shape. With reasoning none it also sends Qwen's switch to turn thinking off (requestBodyFor, providerOptionsFor). makeEmbedder sends POST {baseURL}/embeddings.
• For AI Gateway it uses the AI SDK's own gateway provider (createGateway), which the AI SDK docs recommend over pointing an OpenAI-format provider at the gateway. Kizuki sends no temperature there, because hosted reasoning models reject one.
• For nomic models each text starts with search_document: or search_query: (embedPrefix).

Requests carry your material's sentences, your corrections and readings, and, in a session, your explanation and answers. API keys go in the Authorization header only when the settings name a key variable.

Before material leaves your computer

isLocalUrl counts an address as local only when its host is localhost, 127.0.0.1, or ::1. sendsMaterialOut is true when either model's address is not local.

When /model would switch to such settings and you have not agreed before, Kizuki asks: "These models run off your computer. Kizuki will send them your material, your explanations, and your answers. Send them?" No changes nothing. Yes saves the settings with sendOutAllowed set. writeSettings refuses settings that send material out without it, so the rule holds even if something other than /model writes the settings.