Literature Map

The literature map is an interactive visualization that shows how the papers in your knowledge base relate to each other based on their content. Papers that cover similar topics appear close together; papers on different topics are farther apart.

How it works

The map is built from the abstract embeddings of your papers — the same vector representations used for search and retrieval.

1. Dimensionality reduction (PCA) — each paper's 1024-dimensional abstract embedding is projected down to two dimensions using principal component analysis. The first principal component (PC1) becomes the x-axis and the second (PC2) becomes the y-axis. Together they capture the two directions of greatest variation in your corpus.

2. Clustering (k-means) — papers are grouped into clusters using k-means on their 2D positions. The number of clusters is chosen automatically using the formula k = round(sqrt(n / 2)), clamped to a minimum of 2 and a maximum of 8. For example, a KB with 50 papers produces 5 clusters, while one with 10 papers produces 2.

3. Cluster labeling — each cluster gets an LLM-generated label describing its theme, produced during the background computation job. If no LLM is configured, labels fall back to the most frequent significant words from the cluster's paper titles.

The map is pre-computed and cached. It updates automatically when papers are added or removed and embeddings are recalculated.

Reading the map

• Dots represent individual papers, colored by cluster.
• Cluster labels appear near the center of each cluster.
• Proximity indicates similarity — papers close together share more conceptual overlap in their abstracts.
• The scree chart below the map shows how much variance each principal component explains. The first two PCs (highlighted) are used for the plot axes. If they explain a large share of the variance, the 2D layout is a faithful representation; if not, there is more structure in higher dimensions that the flat map cannot show.
• The cluster legend lists all clusters with their colors and labels.

Navigating the map

• Hover over a dot to see the paper's title, first author, year, and journal.
• Click a dot to go to that paper's detail page.
• Scroll to zoom in and out.
• Drag to pan around.
• Zoom buttons (+/−) in the top-right corner step through zoom levels.
• Double-click to reset the view.
• Click a cluster in the legend to highlight just that cluster's papers (click again to show all).

Requirements

The literature map requires at least two papers with computed abstract embeddings. If embeddings are not yet ready, the page will show a message indicating that the map will be generated once processing is complete.

An LLM connection is optional — without one, cluster labels fall back to the word-frequency method, which still produces useful (if less polished) labels.