Agent Constructors

Convenience constructors for Agent objects based on different backends, such as Ollama or OpenAI. Users should call one of these functions to gain access to an LLM. For the local backends, every attempt is made to install the model and even the backend in some cases, so that users can call these functions in a script and expect the script to run anywhere, more or less reproducibly.

ollama_agent(name, pull = NA, server = ollama_server(), ...)
llama_cpp_agent(path, mode = c("chat", "embedding"), alias = NULL,
                server_path = NULL, port = 0L, ...)
openai_agent(name = "gpt-4o-mini", ...)
azure_openai_agent(name = "gpt-4o", url = getOption("wizrd_azure_openai_url"),
                   ...)

Arguments

name: Name of the model as identified by the backend
pull: Whether to pull the model from Ollama's repository. If NA (the default), prompt the user, except when the session is non-interactive, in which case TRUE is assumed.
server: OllamaServer object. By default, looks for a server running on localhost on the default port. If no server is found, attempts to start the server. Note that if this ends up starting the server, the server process will be killed when the returned object goes out of scope. Thus, it is safest to start the server explicitly using ollama_server and keep a reference around (or just run ollama serve in a shell).
path: A string pointing to a llamafile or a file containing weights in a format compatible with llama.cpp. This can be a simple path on the file system, a URL or the name of an Ollama model. If it resembles a URL, the file is downloaded and cached in a standard place (see R_user_dir). If it resembles an Ollama identifier, the path to the model is automatically resolved, pulling the model if necessary.
mode: Whether the model is to be used for "chat" or "embedding". Embedding models are accessed via the llamafile "v2" web server, while chat models are run through the default llama.cpp web server.
alias: An alternative, more human-friendly name for the model, instead of using the path.
server_path: The path to the llama.cpp server binary. Ignored if path is a (self-contained) llamafile. Otherwise, defaults to the automatic, user-local installation of the llamafile utility, downloading and installing it if necessary and (when interactive) the user approves.
port: The listening port for the llama.cpp HTTP server. By default, automatically finds an open port.
url: The URL of the Azure endpoint, typically ending with the hostname.
...: Model parameters, see LanguageModelParams.

Commercial Providers

Using a commercial provider, like via openai_agent or azure_openai_agent, requires some configuration, depending on the provider:

openai_agent: OPENAI_API_KEY holds the API key
azure_openai_agent: AZURE_OPENAI_API_KEY holds the API key, and option(wizrd_azure_openai_url=) sets the account-specific URL (up to the hostname).

Issues with llamafile

While a llamafile is supposed to work seamlessly across platforms, OS-specific issues may arise. Examples include:

On Linux, issues with binary format registration may result in a wine popup appearing when attempting to run a llamafile.
On Windows, llamafiles over 4GB in size are not supported.

See the llamafile README for workarounds.

Value

An Agent object

Author

Michael Lawrence

Note

These functions rely on an unexported lower level API that will eventually be exported to enable extension to additional backends.

Examples

if (FALSE) { # \dontrun{
    model <- ollama_agent("llama3.2:3b-instruct-q4_K_M") # or llama()
    model |>
        instruct("Answer questions about this dataset:", mtcars) |>
        predict("What is the relationship between horsepower and mpg?")
} # }