RAG that never goes stale: continuous vector sync with Drasi

This article addresses the common problem of stale embeddings in RAG pipelines, where nightly reindexing wastes resources. It introduces Drasi's SyncVectorStore reaction, which uses continuous queries and change data capture to update embeddings only for changed rows. The author provides a working sample using Azure Database for PostgreSQL, Azure OpenAI embeddings, and an in-memory vector store, including YAML configuration and practical tips like keyField for idempotency and Handlebars templates for document shaping. The key benefit is that embedding costs scale with data churn, not corpus size, and deletions are handled automatically. The article also notes a bug limiting vector store backend to InMemory and a secret reference issue with connectionString.