Vector Database Sizing Calculator

About the Vector Database Sizing Calculator

Plan the memory and storage of a retrieval-augmented generation (RAG) index before you buy the servers. Enter how many chunks you will embed, the embedding dimensions and precision, the index type and its settings — HNSW with M, or IVF with its lists and product-quantization code size — the metadata stored with each vector, the number of copies and how fast the data grows.

The calculator shows the raw vectors, what the index adds, the memory an in-memory index needs, the disk it takes and a year of growth, each line with its formula and your numbers. You can also add the cost of creating the embeddings from the tokens per chunk and a price you enter, and read the trade-offs of quantization and shorter embeddings.

Common uses

• Choosing a server or a managed plan with enough memory for a RAG index.
• Comparing HNSW with IVF-PQ, or float32 with float16, int8 and binary vectors, before you build.
• Planning how much memory and disk a year of new documents needs.
• Estimating the one-off and monthly cost of embedding your documents.

The formulas

• Vectors: chunks × dimensions × bytes per dimension (float32 4, float16 2, int8 1, binary ⅛). pgvector’s README gives the same sizes for its types: 4 × d + 8 bytes for vector, 2 × d + 8 for halfvec and d ÷ 8 + 8 for bit.
• HNSW: “(d * 4 + M * 2 * 4) bytes per vector” (Faiss guidelines): the vector plus M × 2 links of 4 bytes on the bottom layer. hnswlib puts the graph at roughly M × 8–10 bytes per element, the upper layers adding the rest.
• IVF: each vector also stores an 8-byte id (Faiss indexes: IVF-Flat 4 × d + 8, IVF-PQ code size + 8), and the index keeps one float32 centroid per list.
• PQ: with 8-bit codes the code size in bytes equals the number of sub-vectors, and each sub-vector has 256 centroids: 256 × d floats of codebooks in all.
• Memory adds up the parts an in-memory index needs; disk adds the metadata and, for IVF-PQ, the full vectors kept for re-ranking. Both are multiplied by the number of copies.

Choosing an index

Flat compares the query with every vector: exact, and fine for tens of thousands of vectors. HNSW is a graph that finds neighbours fast at the cost of M × 8 bytes per vector; M 12–48 suits most data, and ef_construction and ef_search trade build and query time for recall without changing memory (hnswlib). IVF-Flat splits the vectors into lists and searches only a few of them. Faiss’s guidelines suggest 4√N to 16√N lists below a million vectors, then 65,536 lists up to 10 million, 262,144 up to 100 million and 1,048,576 up to a billion, trained on 30 to 256 vectors per list; pgvector suggests rows ÷ 1000 up to a million rows and √rows above. IVF-PQ also compresses each vector into a short code, which saves the most memory and needs re-ranking on the full vectors for the best recall.

Smaller vectors: precision and dimensions

Lower precision shrinks memory in proportion: float16 halves it, int8 quarters it and binary codes are a thirty-second of float32. How much recall you lose depends on the model and the data, so test it on your own queries. Some embedding models are trained so that their vectors can be shortened: OpenAI’s embeddings guide reports that a text-embedding-3-large embedding shortened to 256 dimensions still beats a full 1,536-dimension text-embedding-ada-002 one on the MTEB benchmark. pgvector indexes up to 2,000 dimensions for vector, 4,000 for halfvec and 64,000 for bit columns.

Limitations

• An estimate from the published per-vector formulas: every database adds its own overhead (segment files, write-ahead logs, deleted vectors waiting for compaction, page alignment), so leave headroom.
• Building an index needs extra memory for a while (pgvector, for example, builds HNSW fastest when the graph fits in maintenance_work_mem); that peak is not included.
• HNSW is counted with its bottom layer of links (the Faiss formula); hnswlib’s M × 8–10 bytes includes the upper layers, a few per cent more.
• Recall, latency and the prices of managed vector databases are not estimated.

Privacy

Everything happens in your browser. What you enter or open here is not uploaded or stored by MySmartCoPilot.