The goal of rollama is to wrap the Ollama API, which allows you to run
different LLMs locally and create an experience similar to
ChatGPT/OpenAI’s API. Ollama is very easy to deploy and handles a huge
number of models. Checkout the project here:
https://github.com/ollama/ollama.
While there are several R packages for working with LLMs, rollama
takes an opinionated approach centred on local, open-weight models:
prioritising privacy, reproducibility, and ease of use for research
tasks. The package and its learning materials are particularly focused
on annotating text and images — making it a natural fit for (social)
scientists who want to use LLMs without relying on proprietary APIs or
sending sensitive data to third-party servers. It also offers deep
integration with the Ollama ecosystem beyond just chat, including model
management features like creating, copying, and pushing custom models.
See the Similar packages section for a comparison
with related packages.
You can install this package from CRAN:
install.packages("rollama")Or you can install the development version of rollama from
GitHub. This version is updated
more frequently and may contain bug fixes (or new bugs):
# install.packages("remotes")
remotes::install_github("JBGruber/rollama")However, rollama is just the client package. The models are run in
Ollama, which you need to install on your system, on a remote system
or through Docker. The easiest way
is to simply download and install the Ollama application from their
website. Once Ollama is running, you can see if
you can access it with:
rollama::ping_ollama()
#> ▶ Ollama (v0.17.4) is running at <http://localhost:11434>!For beginners we recommend to download Ollama from their website. However, if you are familiar with Docker, you can also run Ollama through Docker. The advantage of running things through Docker is that the application is isolated from the rest of your system, behaves the same on different systems, and is easy to download and update. You can also get a nice web interface. After making sure Docker is installed, you can simply use the Docker Compose file from this gist.
If you don’t know how to use Docker Compose, you can follow this video to use the compose file and start Ollama and Open WebUI.
The first thing you should do after installation is to pull one of the
models from https://ollama.com/library. By calling pull_model()
without arguments, you are pulling the (current) default model —
“llama3.1 8b”:
library(rollama)pull_model()There are two ways to communicate with the Ollama API. You can make single requests, which does not store any history and treats each query as the beginning of a new chat:
# ask a single question
query("Why is the sky blue? Answer with one sentence.")
#>
#> ── Answer from llama3.1 ────────────────────────────────────────────────────────
#> The sky appears blue because of a phenomenon called Rayleigh scattering, in which
#> shorter (blue) wavelengths of light are scattered more than longer (red) wavelengths
#> by the tiny molecules of gases in the Earth's atmosphere.With the output argument, we can specify the format of the response. Available options include “text”, “list”, “data.frame”, “response”, “httr2_response”, and “httr2_request”:
# ask a single question and specify the output format
query("Why is the sky blue? Answer with one sentence." , output = "text")
#>
#> ── Answer from llama3.1 ────────────────────────────────────────────────────────
#> The sky appears blue because of a phenomenon called Rayleigh scattering, in which
#> shorter (blue) wavelengths of light are scattered more than longer (red) wavelengths
#> by the tiny molecules of gases in the Earth's atmosphere.Or you can use the chat function, treats all messages sent during an R
session as part of the same conversation:
# hold a conversation
chat("Why is the sky blue? Give a short answer.")
#>
#> ── Answer from llama3.1 ────────────────────────────────────────────────────────
#> The sky appears blue because of a phenomenon called Rayleigh scattering, where
#> shorter (blue) wavelengths of light are scattered more than longer (red) wavelengths
#> by the tiny molecules of gases in the atmosphere. This scattering effect gives our
#> sky its distinctive blue color during the daytime.
chat("And how do you know that? Give a short answer.")
#>
#> ── Answer from llama3.1 ────────────────────────────────────────────────────────
#> I was trained on vast amounts of text data, including scientific information and
#> explanations from experts in various fields, such as physics and atmospheric science.
#> This training enables me to recall and summarize established knowledge on topics
#> like the color of the sky.If you are done with a conversation and want to start a new one, you can do that like so:
new_chat()You can set a number of model parameters, either by creating a new model, with a modelfile, or by including the parameters in the prompt:
query("Why is the sky blue? Answer with one sentence.", output = "text",
model_params = list(
seed = 42,
num_gpu = 0)
)
#>
#> ── Answer from llama3.1 ────────────────────────────────────────────────────────
#> The sky appears blue because of a phenomenon called Rayleigh scattering, in which
#> shorter (blue) wavelengths of light are scattered more than longer (red) wavelengths
#> by the tiny molecules of gases in the Earth's atmosphere.You can configure the server address, the system prompt and the model
used for a query or chat. If not configured otherwise, rollama assumes
you are using the default port (11434) of a local instance
(“localhost”). Let’s make this explicit by setting the option:
options(rollama_server = "http://localhost:11434")You can change how a model answers by setting a configuration or system message in plain English (or another language supported by the model):
options(rollama_config = "You make short answers understandable to a 5 year old")
query("Why is the sky blue?")
#>
#> ── Answer from llama3.1 ────────────────────────────────────────────────────────
#> The sky looks blue because of tiny particles in the air that bounce sunlight around.
#> It's like when you shine a flashlight through a prism and it makes colors appear -
#> it's kind of like that! The blue ones just happen to be the strongest, so we see them most.By default, the package uses the “llama3.1 8B” model. Supported models
can be found at https://ollama.com/library. To download a specific
model make use of the additional information available in “Tags”
https://ollama.com/library/llama3.2/tags. Change this via
rollama_model:
options(rollama_model = "llama3.2:3b-instruct-q4_1")
# if you don't have the model yet: pull_model("llama3.2:3b-instruct-q4_1")
query("Why is the sky blue? Answer with one sentence.")
#>
#> ── Answer from llama3.2:3b-instruct-q4_1 ───────────────────────────────────────
#> The sky looks blue because of something called light, and when it comes from the
#> sun, it scatters and goes all around up in the air, making it look blue!The make_query function simplifies the creation of structured queries,
which can, for example, be used in annotation
tasks.
Main components (check the documentation for more options):
text: The text(s) to classify.prompt: Could be a (classification) questionsystem: Optional system prompt providing context or instructions for the task.examples: Optional prior examples for one-shot or few-shot learning (user messages and assistant responses).
Zero-shot Example
In this example, the function is used without examples:
# Create a query using make_query
q_zs <- make_query(
text = "the pizza tastes terrible",
prompt = "Is this text: 'positive', 'neutral', or 'negative'?",
system = "You assign texts into categories. Answer with just the correct category."
)
# Print the query
print(q_zs)
#> [[1]]
#> # A tibble: 2 × 2
#> role content
#> <chr> <glue>
#> 1 system You assign texts into categories. Answer with just the correct categor…
#> 2 user the pizza tastes terrible
#> Is this text: 'positive', 'neutral', or 'neg…
# Run the query
query(q_zs, output = "text")
#>
#> ── Answer from llama3.2:3b-instruct-q4_1 ───────────────────────────────────────
#> Negative- Use rollama for annotation tasks
- Annotate images
- Get text embedding
- Use more models (GGUF format) from Hugging Face
There are some similar R packages for working with LLMs:
- ellmer is part of the Posit/tidyverse ecosystem and supports a broad range of providers (OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock, Ollama, and many more). It is especially good for interactive console chats, chatbots in Shiny, advanced tool-calling, and structured data extraction via schema functions.
- tidyllm supports multiple providers (Anthropic Claude, OpenAI, Google Gemini, Groq, Mistral, Perplexity, and Ollama) with a pipeline-oriented, side-effect-free interface. Highlights include media handling (PDFs, images, video/audio for Gemini), batch processing for cost savings, and a tidy workflow designed for integration into data analysis pipelines.
- ollamar has the closest overlap
with
rollamaas it wraps the Ollama API in its entiority. The design philosphy however, is quite different, as the package closely mirrors the official Ollama Python/JavaScript libraries. It therefore feels more familiar for users of these languages, but potentially less familiar for many R users.
Please cite the package using the pre print DOI: https://doi.org/10.48550/arXiv.2404.07654