2 · Embedding Spaces & Chunking

2. Embedding Geometries & Chunking Strategies

Dense vectors, cosine distance, semantic chunking, and sliding window overlaps.

9 min read · 2 MCQs

Vector Space Geometry

Embedding models map unstructured text into continuous high-dimensional vector spaces (e.g. 1536 dimensions). Cosine similarity measures the angle between vectors rather than Euclidean magnitude, evaluating semantic alignment independent of text length.

Chunking Trade-Offs

Small chunks (100–250 tokens) provide high vector precision but lack surrounding context. Large chunks (1000+ tokens) capture broad context but dilute semantic signal. Recursive character chunking with 15–20% overlap balances boundaries around natural paragraphs.

Chapter quiz

2 questions · pass mark 75%

1. Why is cosine similarity preferred over Euclidean distance for embedding comparisons?
2. What problem does chunk overlap mitigate in RAG ingestion pipelines?