Vector Search
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Vector search (also called semantic search) is a retrieval technique that finds relevant documents or code by comparing dense vector embeddings in high-dimensional space, rather than matching exact tokens or keywords. It is widely considered the archetypal component of RAG pipelines, though practitioners argue this conflation understates the full scope of modern retrieval.
Common Misconception: Vector Search as Synonym for RAG
Kuba Rogut of Turbopuffer identifies a pervasive misunderstanding in the field: "what a lot of people think rag is is just simple vector search." 0:46 He argues this framing is too narrow — retrieval in practice encompasses vector search alongside full text search, grepping, globbing, regex, and filters, all used iteratively by an agent. 3:20
Demonstrated Performance Value
Despite being only one component of a broader retrieval toolkit, vector search has shown measurable impact in production systems. Rogut cites Cursor as a concrete case: adding semantic search to Cursor's composer model produced a 24% increase in answer accuracy and a 2.6% gain in code retention in large codebases. 8:36 Cursor implements this by embedding the user's entire codebase and making it available for semantic search queries at inference time. 5:42
Role in Agent Pipelines
The framing from Rogut positions vector search not as a standalone solution but as one retrieval primitive among many that an agent may invoke depending on query type and context. This reflects a broader shift from static, single-step RAG toward iterative, multi-modal retrieval architectures where vector search handles semantic similarity tasks while complementary techniques (keyword search, regex, structured filters) handle cases where embedding-based retrieval falls short.