An embedding represents an unstructured data item as a large, ordered vector of floating-point numbers. Its coordinates locate the item in a model-defined conceptual space, making the data suitable for exploratory analysis and similarity operations. section 6.1.4
Use embeddings to:
- Generate vectors for all or a reasonable subset of photos, text, or other unstructured records with a foundational model. section 6.1.4
- Project high-dimensional vectors with t-SNE or PCA to inspect the distribution of data items, while treating information loss as a risk. section 6.1.4
- Index vectors with FAISS or Annoy and use nearest-neighbor queries to extract structure from a corpus. section 6.1.4
Exploring Shakespeare
Encode Shakespeare’s plays with all-MiniLM-L12-v2, then inspect nearest-neighbor connections or significant links between plays. section 6.1.4
array([ 8.19933936e-02, 9.97491628e-02, 6.05560839e-02, 6.95289299e-02,
4.54569124e-02, -5.78593016e-02, 2.58211885e-02, -5.37960902e-02,
1.54499486e-02, 1.98997390e-02, 3.31314690e-02, 4.48754244e-02,
-1.08558014e-02, 2.36593257e-03, -1.36038102e-03, 5.81134520e-02,
4.76119894e-04,....
= Indeed, there is Fortune too hard for Nature, when
Fortune makes Nature’s natural the cutter-off of
Nature’s wit.
Inspect a sample embedding alongside the text it represents; the raw coordinates are not directly useful without indirect analysis. section 6.1.4
