Embeddings
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Embeddings are vector representations of text (or other data) used to enable semantic search and RAG (Retrieval-Augmented Generation) pipelines, allowing systems to retrieve contextually relevant information based on meaning rather than exact keyword matching.
Embeddings as Cached Compute
Kuba Rogut of Turbopuffer characterizes embeddings and semantic search as a form of cached compute: the idea being that embedding a document pre-computes a representation that can be reused across many queries, rather than re-running full inference each time. In his framing, this caching property is what makes embeddings economically attractive — their value scales with query volume, since the upfront cost of computing and storing embeddings is amortized over repeated retrievals. ↗
This framing has implications for how embeddings compete with or complement approaches like long-context models: as the cost of running large context windows at query time decreases, the relative advantage of pre-computed embeddings may shift depending on the volume and nature of queries a system handles.