The course

The trade-off

What quantization is, and how we measure whether it worked.

• 0.1 Fewer bits, on purpose
• 0.2 Memory vs accuracy
• 0.3 The ruler: recall and compression

Scalar quantization

The simplest compressor: a range of floats onto 256 integer levels.

• 1.1 A range into 256 levels
• 1.2 Calibrating the range

k-means

Grouping points into clusters and learning a codebook.

• 2.1 Grouping points
• 2.2 Codebooks: a vocabulary of vectors

Product quantization

Split a vector, replace each chunk with a codebook index.

• 3.1 Splitting a vector into chunks
• 3.2 Encoding to bytes
• 3.3 Reconstruction error

Searching squashed data

Estimate closeness from compressed vectors, and draw the curve.

• 4.1 Asymmetric distance
• 4.2 Recall vs compression: the curve
• 4.3 Re-ranking the candidates

The frontier

The 2026 methods, and the debate over what is genuinely new.

• 5.1 Rotate, then round (TurboQuant)
• 5.2 RaBitQ and the error bound
• 5.3 The novelty debate