AI configuration

The assistant is configured in the ai: section of the config file. This page lists every key, the default model of each provider, where API keys come from and how to set up the embedder for similarity search.

Quick start

ai:
  provider: googleai          # googleai, vertexai, anthropic, openai or ollama
  model: gemini-3.8-flash     # optional: the provider's default is used when empty
  api_key: env:GEMINI_API_KEY # a secret reference, never the key itself
  send_rows: ask              # ask, always or never
  embedder: googleai/text-embedding-005  # optional: text similarity search

Keys

Key	Default	Meaning
`provider`	`googleai`	One of `googleai`, `vertexai`, `anthropic`, `openai`, `ollama`.
`model`	the provider’s default (below)	The provider’s model name.
`api_key`		A secret reference (`env:`, `file://`, dotenv, `keyring:`). A literal key, or a value that is not a reference, is a load error.
`base_url`		An `http://` or `https://` URL, without credentials: an OpenAI-compatible server for `openai`, the server address for `ollama`.
`project`		The Google Cloud project, for `vertexai`.
`location`	`us-central1`	The Google Cloud region, for `vertexai`.
`send_rows`	`ask`	Whether row data may be sent to the model: `ask`, `always` or `never`. See send_rows. `send_sample_rows: true\|false` is an alias for `always\|never`; set one or the other, not both.
`max_rows`	`50`	Rows in one tool result, 1 to 200.
`max_tool_calls`	`8`	Tool calls in one answer, 1 to 64. When the cap is reached the model is told to answer with what it has.
`context_tokens`	`8000`	Approximate size of the schema context sent with each question, 1000 to 1000000.
`embedder`		The embedding model for text similarity search: see Embedder.

An unknown key under ai: is a load error that suggests the closest key. No message ever quotes the value of a key.

Default models

Provider	Default model
`googleai`	`gemini-3.8-flash`
`vertexai`	`gemini-3.8-flash`
`anthropic`	`claude-sonnet-5-5`
`openai`	`gpt-5`
`ollama`	`llama3.3`

Set model to use another one. The panel header and the tobira ask footer show the model in use.

Keys and credentials

googleai, anthropic, openai. The key comes from api_key when it is set. Otherwise tobira reads the provider’s environment variable: GEMINI_API_KEY and then GOOGLE_API_KEY for googleai, ANTHROPIC_API_KEY for anthropic, OPENAI_API_KEY for openai. With neither, asking fails with an authentication error that names the variable.
openai with base_url. For an OpenAI-compatible server (a local gateway, for example) a key is optional. With none configured, tobira sends a placeholder key, which such servers usually ignore.
vertexai. Uses Application Default Credentials (gcloud auth application-default login, or GOOGLE_APPLICATION_CREDENTIALS) with project and location. It takes no API key; an api_key is ignored. Credentials are looked up on first use, not at startup.
ollama. Talks to http://localhost:11434, or to base_url. It takes no key; an api_key is ignored. Use a model that supports tool calling, because the answer mode needs it.

Provider error messages are redacted: they never contain the key, a password or a secret reference.

Embedder

embedder names the embedding model that turns text into vectors, for similarity search by text on vector databases: :similar <text> in the terminal UI, and the similar tool of the assistant and the MCP server.

Write it as provider/model, for example googleai/text-embedding-005, openai/text-embedding-3-small or ollama/nomic-embed-text, or as a model name of provider. anthropic has no embedding models.
api_key and base_url apply only when the embedder uses provider. Another provider reads its key from its environment variable.
The key is read on the first text search, not at start.
Without an embedder, searching by text says “text similarity needs ai.embedder”. Searching with a vector still works.