Chat With (Local) LLM — chat • certellm

These functions initialise and interact with the active LLM, backed by ellmer::chat(). The connection is configured via named presets defined in the secrets YAML under llm.presets; individual provider, model, and url arguments override the preset when supplied.

initiate_llm(
  preset = read_secret("llm.default.preset"),
  provider = NULL,
  model = NULL,
  url = NULL,
  system_prompt = read_secret("llm.system_prompt"),
  ...
)

list_presets()

new_chat(input, ...)

chat(input, new = FALSE, ...)

chat_in_browser(new = FALSE, ...)

chat_in_console(new = FALSE, ...)

get_chat_object()

get_provider()

get_system_prompt()

get_tools()

get_tokens()

Arguments

preset: Name of a preset defined under llm.presets in the secrets YAML. Defaults to llm.default.preset. Each preset must contain a provider (e.g. "ollama", "anthropic") and a model field; an optional url sets a non-default base URL (e.g. an internal Ollama server). Individual provider, model, and url arguments override the preset when supplied.
provider: Override the provider from the preset.
model: Override the model from the preset.
url: Override the base URL from the preset. Passed as base_url to the provider function; omit for providers with fixed endpoints (e.g. Anthropic).
system_prompt: System prompt text. Defaults to llm.system_prompt from the secrets file.
...: Arguments passed on to ellmer::chat().
input: Text for input.
new: Initiate new LLM instance.

Details

Functions:

initiate_llm() creates a new ellmer Chat object, appends user context to the system prompt when the user's name and job title are found in the secrets file, registers the session tools (get_df_summary, list_objects, get_colnames), and probes the model to verify actual tool support. Models that do not support tools are silently recreated without them, so the function always leaves a working instance. The active object is stored in pkg_env$chat_object.
list_presets() prints all configured presets with their provider, model, and URL, marking the default.
chat() sends a single message to the active LLM, initiating a new session first if none exists or if new = TRUE.
new_chat() is a shorthand that always initiates a fresh session before sending the message.
chat_in_browser() and chat_in_console() open an interactive session via ellmer::live_browser() and ellmer::live_console() respectively.
get_chat_object(), get_provider(), get_system_prompt(), get_tools(), and get_tokens() are thin accessors that expose the underlying ellmer Chat object's state for inspection.