Skip to content

configuration.md

Latest commit

 

History

History
42 lines (33 loc) · 1.79 KB

configuration.md

File metadata and controls

42 lines (33 loc) · 1.79 KB
Raw

Configuration

Precedence: CLI flags, then environment variables, then the config file.

Config file location: %APPDATA%\thoth\config.toml on Windows, ~/.config/thoth/config.toml on Linux/macOS.

base_url = "http://localhost:11434/v1"
model = "qwen3:8b"
# api_key = "..."       # if the server requires one
# temperature = 0.7
# max_turns = 40        # agent steps per request before pausing
# num_ctx = 32768       # context window (only applies to Ollama)
# think = false         # disable thinking (only applies to Ollama);
                        # saves a lot of tokens on small windows

# Use Google for web_search instead of DuckDuckGo. Google has no free
# scraping, so this needs a Programmable Search Engine:
#   1. create an engine at https://programmablesearchengine.google.com (the "cx" id)
#   2. get an API key at https://developers.google.com/custom-search/v1/introduction
# Free tier is 100 queries/day. If Google fails, thoth falls back to DuckDuckGo.
# google_api_key = "..."
# google_cx = "..."

Environment variables: THOTH_BASE_URL, THOTH_MODEL, THOTH_API_KEY, THOTH_NUM_CTX, THOTH_THINK, THOTH_GOOGLE_API_KEY, THOTH_GOOGLE_CX.

CLI flags: --base-url, -m/--model, --api-key, --temperature, -p/--prompt for one-shot mode. See thoth --help.

Context window notes

num_ctx and think only work with Ollama, because thoth talks to Ollama through its native API where those are per-request options. Other servers use the standard OpenAI API, which has no such fields; the window is whatever the server was started with (llama-server -c 16384).

If the model stalls, rambles, or forgets files it just read, the window is almost always too small. Raise num_ctx (costs RAM/VRAM; ollama ps shows the active size), or set think = false for reasoning models.