Your corpus.
Its memory footprint.
Model the space behind your embeddings. See how vector dimensions, storage precision and index overhead add up before planning infrastructure.
Five inputs.
One sizing scenario.
Estimated RAM envelope
10,000,000 vectors · 1536 dimensions · 4 bytes per element · HNSW
- Raw vector storage
- 57.22 GiB
- Index overhead
- 1.19 GiB
- Planning buffer / 30%
- 17.52 GiB
Raw vectors plus modeled index overhead, multiplied by 1.5. This does not include a measured database deployment.
Arithmetic over stated assumptions, not a benchmark. GiB uses 1,073,741,824 bytes. Actual database memory, disk allocation, metadata and workload requirements can differ.
Same corpus.
Different index models.
Capacity arithmetic only.
This table does not compare search quality or latency.
| Index model | Raw vectors | Index overhead | RAM + 30% | Disk × 1.5 | Open index scenario |
|---|---|---|---|---|---|
| Flat | 57.22 | 0.00 | 74.39 | 85.83 | Use index ↗ |
| HNSW | 57.22 | 1.19 | 75.94 | 87.62 | Use index ↗ |
| IVFFlat | 57.22 | 0.02 | 74.41 | 85.86 | Use index ↗ |
Flat has no separate index-overhead term in this simplified model. HNSW and IVFFlat overhead use different heuristics; this is not a recommendation to choose the smallest estimate.
The limits belong
beside the result.
Storage arithmetic
Raw bytes = vector count × dimensions × bytes per element. Divide by 1024³ for GiB. The calculation counts vector elements, not documents, source content or their metadata.
Index overhead heuristics
HNSW overhead = vector count × M × 8 bytes. IVFFlat overhead = square root of vector count × dimensions × bytes per element. Flat adds no separate term. These simplified terms do not reproduce a particular database implementation.
RAM buffer and disk multiplier
RAM = (raw vectors + index overhead) × 1.3. Disk = (raw vectors + index overhead) × 1.5. The 30% RAM buffer and 1.5 disk multiplier are planning assumptions, not measured limits or service guarantees.
What this model does not establish
Sharding, recall, query throughput, index build memory, replicas, backups, metadata, concurrency and provider costs are outside the calculation. Validate your actual database and workload before purchasing or provisioning infrastructure.