Vector Database Configuration for NVIDIA RAG Blueprint

Why it matters

The vector store is a deployment choice behind a stable retrieval interface, and switching it has a cost. This guide shows both: the settings that select a backend and the data that does not move with them.

What to practise

Point the RAG and ingestor servers at a backend with APP_VECTORSTORE_NAME and APP_VECTORSTORE_URL, first for the default Elasticsearch, then for the milvus Compose profile or the Helm values eck-elasticsearch.enabled: false and nv-ingest.milvusDeployed: true. Note that documents must be uploaded again after a switch. Then read the VDBRag contract for a custom backend, passed in through vdb_op: in the retrieval-only variant, retrieval_langchain() must return documents carrying source, content_metadata and collection_name.

How it fits

The backend choice comes before 6.2.1’s index selection. The re-upload after a switch is a reindexing cost; compare it with 6.2.3’s replacement of a collection behind an alias, which keeps the serving name stable while the new collection is built and indexed.