Semantic search over a data catalog, with Postgres and pgvector holding the embeddings and Ollama generating them on CPU. There is no vector database in this design, and adding one would have been the easier decision to defend in a review.

The constraint driving all three notes is that a search feature should not buy an operational dependency. Postgres was already deployed and already backed up, and whoever gets paged already knew how it behaves under load. A dedicated vector store would have been a second system to run for one feature.

Three parts: why a data MCP server needs vector search at all, the embedding pipeline that keeps the index current as the catalog changes, and running Ollama and pgvector on Kubernetes. The work comes out of txn2/mcp-data-platform, also available hosted as Plexara.