Learning path · Embeddings & Representation · 35
Embeddings
Dense vector representations of text (or other modalities) where semantic similarity approximates geometric closeness.
Why it matters
- Foundation of semantic search, RAG retrieval, and clustering.
- Choice of embedding model affects recall on domain jargon.
- Separate from generative LLM—you often use both.
Key ideas
- Vector representations
- Similarity search
- Domain adaptation
Embeddings map sentences or documents into fixed-length vectors optimized so related meanings sit nearby. They power "find passages like this question" without keyword overlap. Pick models trained on text similar to your corpus; legal, medical, and code domains often need specialized embedders. Re-embed when you change models—vectors are not portable across unrelated embedding spaces. Version embedding models in index metadata and block silent cross-version queries that return nonsense similarity scores. Ship only after eval gates pass on representative production failures.
Updated 2026-08-09 · Full learning path