Vector Stores in Spring AI

The VectorStore Interface

Two methods cover 95% of RAG use cases:

• add(List<Document> documents) — embeds each document and stores it with its metadata
• similaritySearch(SearchRequest request) — embeds the query and returns the top-K most similar documents

Qdrant: The Power RAG Choice

Qdrant is an open-source, Rust-based vector database with a gRPC API. Spring AI communicates with it via gRPC (port 6334) for low latency. Qdrant stores vectors in collections — Power RAG uses a single collection named power_rag_docs.

VectorStoreConfig

@Bean
public QdrantVectorStore vectorStore(QdrantClient qdrantClient,
                                     EmbeddingModel embeddingModel) {
    return QdrantVectorStore.builder(qdrantClient, embeddingModel)
        .collectionName(collectionName)
        .initializeSchema(true)
        .build();
}

initializeSchema(true) tells Spring AI to create the Qdrant collection on startup if it does not already exist. This makes deployment zero-config — just start the service and the collection is ready.

Adding Documents

List<Document> docs = chunks.stream()
    .map(c -> new Document(c.getText(), c.getMetadata()))
    .toList();
vectorStore.add(docs);

When vectorStore.add(docs) is called, Spring AI automatically:

1. Calls the EmbeddingModel to embed each document's text
2. Stores the vector + text + metadata as a point in the Qdrant collection

Docker Compose Setup

qdrant:
  image: qdrant/qdrant:latest
  ports:
    - "6333:6333"  # HTTP REST (for Qdrant Web UI / manual queries)
    - "6334:6334"  # gRPC (used by Spring AI)
  volumes:
    - qdrant_data:/qdrant/storage

The named volume qdrant_data persists vectors across container restarts. Without it, all indexed documents would be lost every time the container is recreated.

Metadata in Spring AI Documents

The second argument to new Document(text, metadata) is a Map<String, Object>. Metadata is stored alongside the vector in Qdrant as a payload. Common metadata fields in Power RAG:

• file_name — original filename
• document_id — UUID linking back to the PostgreSQL documents table
• page_number — for PDFs
• row_number — for Excel
• section_title — heading text when available
• chunk_index — position within the document