Rhobots

BERTopic-Style Topic Modelling in Pure R

Author

Janpieter van der Pol

🤖 Rhobots

The full BERTopic pipeline — transformer embeddings, UMAP dimensionality reduction, HDBSCAN density clustering, and class-based TF-IDF — running entirely in R. No Python. No conda. No reticulate.

📦 CRAN v0.1.6 🔬 MIT License 🖥️ R ≥ 4.1 🪟 Windows · 🍎 macOS · 🐧 Linux

The Pipeline

Rhobots implements the four-stage BERTopic pipeline as a set of modular, pluggable R objects. Each stage can be swapped out independently — use PCA instead of UMAP, k-means instead of HDBSCAN, or bring your own representation model.

<span class="icon">📝</span>
<div class="label">Embed</div>
<div class="pkg">torch · safetensors</div>

→

<span class="icon">📉</span>
<div class="label">Reduce</div>
<div class="pkg">uwot · UMAP</div>

→

<span class="icon">🔵</span>
<div class="label">Cluster</div>
<div class="pkg">Rcpp · HDBSCAN</div>

→

<span class="icon">🏷️</span>
<div class="label">Represent</div>
<div class="pkg">c-TF-IDF · POS</div>

Resources

Hands-on

🔬

Full Tutorial

A complete walkthrough on 3,328 African research abstracts — from raw text to labelled topics, quality sweep, interactive visualisations, and LLM-generated topic names.

How it works

📉

How UMAP Works

Step-by-step technical explainer with illustrations: the manifold hypothesis, fuzzy membership weights, SGD optimisation, and the effect of every parameter.

How it works

🔵

How HDBSCAN Works

Core distances, mutual reachability distance, Borůvka MST, the condensed cluster tree, EOM extraction, and what every parameter actually controls.

Why Rhobots?

<span class="fi">🐍</span>
<p><strong>No Python dependency.</strong> The entire pipeline — embeddings, UMAP, HDBSCAN,
c-TF-IDF — runs in R through <code>torch</code>, <code>uwot</code>, and <code>dbscan</code>.</p>

<span class="fi">🤗</span>
<p><strong>Any Hugging Face model.</strong> Load BERT, SciBERT, BGE, E5, SPECTER2, or any
BERT-architecture model directly from the Hub with <code>load_hf_bert()</code>.</p>

<span class="fi">🔢</span>
<p><strong>Automatic topic count.</strong> HDBSCAN finds the right number of topics for your
data — no need to pre-specify <em>k</em>.</p>

<span class="fi">🔍</span>
<p><strong>Hyperparameter sweep.</strong> <code>sweep_topics()</code> evaluates many
parameter combinations in one call, reusing cached embeddings.</p>

<span class="fi">🌿</span>
<p><strong>Noise-aware.</strong> Documents that don't fit any topic get label <em>−1</em>
rather than being forced into a cluster.</p>

<span class="fi">🔌</span>
<p><strong>Pluggable architecture.</strong> Swap any stage — use PCA instead of UMAP,
k-means instead of HDBSCAN, or a custom representation model.</p>

How It Compares

	LDA (`topicmodels`)	STM (`stm`)	Rhobots
Needs Python	No	No	No
Pre-trained embeddings	No	No	Yes
Topic count specified	Yes (fixed k)	Yes (fixed k)	Automatic
Handles noise	No	No	Yes (label −1)
Non-linear structure	No	No	Yes (UMAP + HDBSCAN)
Domain-specific encoders	No	No	Yes (any HF model)
Hyperparameter sweep	No	No	Yes

Installation

Step 1 — Install from CRAN:

install.packages("Rhobots")

Step 2 — Install the torch backend (once per machine, ~560 MB download):

library(Rhobots)
rhobots_install()
# Restart your R session when prompted

Note

Windows users must first install the Microsoft Visual C++ Redistributable 2022 before running rhobots_install().

Step 3 — Try the built-in demo:

rhobots_demo()

The demo downloads four Gutenberg books, embeds them, sweeps parameters, fits a topic model, and produces an interactive visualisation — no data files needed.

Quick Start

library(Rhobots)

# 1. Load a sentence encoder
enc <- load_hf_bert("sentence-transformers/all-MiniLM-L6-v2")

# 2. Embed your documents (result cached to disk)
emb <- embed_texts_cached(enc, docs, cache_file = "emb.rds", normalize = TRUE)

# 3. Sweep parameters to find the best combination
sw  <- sweep_topics(docs, emb,
                    n_neighbors  = c(10, 15, 20),
                    n_components = c(5, 10),
                    min_pts      = c(10, 20, 30))
print(sw)

# 4. Fit with the best parameters
fit <- fit_bertopic(docs, emb,
                    umap_n_neighbors  = sw$best$n_neighbors,
                    umap_n_components = sw$best$n_components,
                    hdbscan_min_pts   = sw$best$min_pts)

# 5. Inspect topics
print(fit)
visualize_barchart(fit)
visualize_topics(fit)

Citation

If you use Rhobots in research, please cite:

van der Pol, J.P.G. (2026). Rhobots: BERTopic-Style Topic Modeling Without Python.
R package version 0.1.6. https://github.com/JPvdP/Rhobots

The algorithm is based entirely on the Python BERTopic package by Maarten Grootendorst.