Vector stores

All stores implement the same protocol: add(ids, vectors, texts, metadatas), search(vector, k), and len(). Scores are always larger-is-closer. Pass a store to build_retriever(..., store=...) or HybridRetriever(embedder, store).

Store	Extra	Persistence	Delete
InMemoryVectorStore	none	process memory	yes
FaissVectorStore	faiss	process memory	no
QdrantVectorStore	qdrant	server, cloud, local folder, or memory	yes
PgVectorStore	pgvector	PostgreSQL	yes
ElasticsearchStore	none	Elasticsearch 8	yes
OpenSearchStore	none	OpenSearch 2	yes

The Elasticsearch and OpenSearch stores also serve BM25 for hybrid search; see Elasticsearch and OpenSearch. Stores with delete support work with incremental indexing.

Qdrant

from persian_rag.vectorstores import QdrantVectorStore

server = QdrantVectorStore("persian_docs", dimension=384, url="http://localhost:6333")
cloud = QdrantVectorStore("persian_docs", dimension=384, url="https://...", api_key="...")
on_disk = QdrantVectorStore("persian_docs", dimension=384, path="./qdrant-data")
in_memory = QdrantVectorStore("persian_docs", dimension=384)

String ids are mapped to deterministic UUIDs; the original id, text, and metadata live in the payload. The collection is created on first use (distance="cosine", "dot", or "euclid"). When reopening an existing collection the dimension is read from it. Pass client= to reuse a configured QdrantClient.

PostgreSQL with pgvector

from persian_rag.vectorstores import PgVectorStore

store = PgVectorStore(
    "postgresql://user:pass@localhost:5432/db",
    dimension=384,
    table="persian_chunks",
    distance="cosine",  # or "l2", "inner_product"
    create_index=True,  # HNSW index
)

The store creates the vector extension (when the role may), a table with id, embedding, text, and metadata JSONB, and optionally an HNSW index. Pass connection= to reuse an autocommit psycopg 3 connection. Table names are validated, and all values are sent as query parameters.

To run a local database for development:

docker run -e POSTGRES_PASSWORD=postgres -p 5432:5432 pgvector/pgvector:pg16

FAISS

from persian_rag import FaissVectorStore

store = FaissVectorStore(dimension=384, metric="cosine")