How to use this calculator
- Enter or choose number of vectors.
- Enter or choose dimensions.
- Enter or choose precision.
- Enter or choose metadata per vector.
- Enter or choose index overhead.
- The monthly cost updates as you type. Use Copy result or Share to keep it.
Method
Raw bytes = Vectors × (Dimensions × Bytes per dimension + Metadata); Stored = Raw × (1 + Overhead%) × Copies; Cost = Stored GB × Price + Queries × Query price + Writes × Write price
- Bytes per dimension
- 4 (float32), 2 (float16), 1 (int8), 0.125 (binary)
- Copies
- primary plus replicas
Worked example
- Per vector: 768 × 4 + 200 = 3,272 bytes; × 10 million = 32.72 GB.
- With 50% index overhead: 49.08 GB × 0.25 = 12.27.
- Queries 5 × 2.00 = 10; writes 1 × 1.00 = 1; total 23.27.
Ten million 768-dimension float32 vectors need about 49 GB with index overhead, about 23.27 a month at these example prices.
What the result means
Dimensions and precision drive size. Quantising float32 to int8 cuts raw vector storage by 75%, usually at a small cost in recall; check the trade-off on your own queries.
Assumptions and limits
- All vectors have the same dimensions and average metadata size.
- Index overhead applies to the whole stored record.
- Many managed services price by pods, nodes or read and write units instead of GB; convert your plan to these terms or enter an equivalent rate.
- Memory-resident indexes may need RAM sized to the index, which is priced differently from disk.
Questions
How big is one embedding?
Dimensions × bytes per value. A 768-dimension float32 vector is 3,072 bytes; a 1,536-dimension one is 6,144 bytes.
Does embedding generation cost go here?
Not directly. Creating embeddings is a one-off token cost per document; use the LLM Token Cost Calculator with your embedding model’s price.