Embedding Storage Calculator

An embedding dimension is the size of the vector that represents a piece of text. text-embedding-3-small uses 1536 dimensions by default (can be reduced to 512). Each dimension is a float32 (4 bytes). Higher dimensions capture more semantic nuance but cost more storage.

Storage = number of vectors × dimension × 4 bytes (float32). For 1M vectors at 1536 dims, that's 1M × 1536 × 4 = ~6 GB of raw vector data. Add ~20% overhead for index structures (HNSW, IVF). Quantized (int8) storage halves the size.

float32 gives full precision. int8 quantization (product quantization) reduces storage by 4× with a small accuracy drop (~1-3% on recall@10). For production RAG at scale, int8 is usually worth the trade-off. OpenAI's text-embedding-3 supports native dimension reduction as a better alternative.

For managed cloud: Pinecone (simplest), Weaviate Cloud, Qdrant Cloud. For self-hosted: pgvector (if you're already on Postgres), Chroma (local dev), Qdrant, Milvus. For serverless scale: Turbopuffer, Zilliz. Choose based on scale, existing infra, and query pattern.