tsvector vs pgvector: The Postgres Naming Collision That Confuses Newcomers

Postgres ships with two features that share the word vector but are completely unrelated. Mixing them up is one of the most common confusions newcomers hit when reading about modern Postgres patterns. `tsvector` is short for text search vector and has been in Postgres since 2007. It contains zero machine learning. It performs linguistic preprocessing: tokenize input text, strip stop words like 'the' and 'a', apply Snowball stemmer rules so 'running' becomes 'run' and 'dogs' becomes 'dog', and track word positions. Calling `SELECT to_tsvector('english', 'The running dogs')` returns something like `'dog':3 'run':2`. It is used with `tsquery` and the `@@` match operator for full-text search, and a GIN index scales it to hundreds of millions of documents. This is the feature that lets Postgres replace Elasticsearch for keyword search. `vector`, from the pgvector extension released in 2021, is a different beast entirely: a typed column holding a fixed-length array of floats, typically 384, 768, or 1536 dimensions. The values come from an AI embedding model such as OpenAI text-embedding-3-small, sentence-transformers, Cohere, or Voyage. Queries use cosine distance (`<=>`) or inner product (`<#>`) operators with HNSW or IVF indexes. This is what powers Semantic Search and replaces dedicated Vector Databases like Pinecone. The quick disambiguation rule: if you see `tsvector`, `to_tsvector()`, `tsquery`, `@@`, or `ts_rank`, that is native text search with no AI involved. If you see `vector(1536)`, `<=>`, `<#>`, HNSW, or the pgvector extension, that is semantic search and it does require running an AI model to generate the embeddings. The collision is historical. Postgres named `tsvector` back when 'vector' meant 'collection of values arranged in order' in the generic mathematical sense. By the time pgvector arrived 14 years later, 'vector' had acquired its modern ML-era connotation of 'point in high-dimensional embedding space.' Both names are now locked in. A common modern pattern, often called hybrid search, runs both in the same query: filter and rank by `tsvector` keyword match, then re-rank by `pgvector` semantic distance. Doing this in one Postgres transaction is one of the strongest arguments for consolidating onto Postgres rather than splitting search across two systems.