Embedding Visualization
Why Visualize?
Understand clustering patterns
Identify outliers
Debug retrieval issues
Communicate insightsDimensionality Reduction
High dimensions (1536) → 2D/3D for visualization
t-SNE:
Preserves local structure
Good for clusters
Non-deterministicUMAP:
Faster than t-SNE
Preserves global structure better
Good default choicePCA:
Fast, deterministic
Linear reduction
May lose nuanceImplementation
from sklearn.manifold import TSNE
from umap import UMAP
# t-SNE
tsne = TSNE(n_components=2, perplexity=30)
coords = tsne.fit_transform(embeddings)
# UMAP
umap = UMAP(n_components=2, n_neighbors=15)
coords = umap.fit_transform(embeddings)
Visualization Tools
Matplotlib/Seaborn: Basic plotting
Plotly: Interactive 3D
TensorBoard Embedding Projector
Atlas/Nomic: Specialized tools