ellmer routes ollama via the OpenAI-compatible /v1/ endpoint, which ignores think=FALSE. This helper calls /api/chat directly where think is honoured.
Usage
.llm_ollama_native(
text,
system_prompt,
model = NULL,
think = FALSE,
options = list(),
base_url = rep_if(Sys.getenv("OLLAMA_BASE_URL"), "http://localhost:11434", replace =
""),
timeout = llm_timeout()
)Arguments
- text
The text to send to the LLM (vector of strings, or data frame with the text in a column)
- system_prompt
A system prompt to set the behavior of the assistant
- model
the ollama model
- think
whether to use thinking mode (very slow)
- options
further options to pass to to the model
- base_url
the local URL
- timeout
request timeout in seconds; defaults to
llm_timeout(). A stuck/overloaded local Ollama server otherwise blocks here with no log output for as long as curl is willing to wait – this turns that into a visible, retriable per-call failure instead.