vector-db-sizer 0.1.0

Analytical vector database storage estimator

vector-db-sizer

What it is

vector-db-sizer is an analytical CLI estimator for vector database disk and RAM sizing.

When to use it

Use it for fast pre-implementation sizing work, such as:

• early architecture decisions;
• comparing vector dimensions;
• comparing engines;
• comparing index types;
• estimating metadata/payload impact;
• generating Markdown/CSV/JSON artifacts for architecture discussions.

What it does not do

• No live database connections.
• No ingestion or load execution.
• No latency/recall benchmarking.
• No pricing calculations.
• No production guarantee.

Quick start

uv sync
uv run vector-db-sizer validate examples/qdrant_text_hnsw.yaml
uv run vector-db-sizer estimate examples/qdrant_text_hnsw.yaml --format markdown --out report.md
uv run vector-db-sizer estimate examples/multi_scenario.yaml --format csv --out comparison.csv

Input YAML

name: qdrant_text_hnsw

dataset:
  source_type: text
  total_tokens: 50000000
  chunk_tokens: 512
  chunk_overlap: 64

embedding:
  kind: dense
  dimensions: 1536
  dtype: float32

database:
  engine: qdrant
  index_type: hnsw

Single-scenario example

uv run vector-db-sizer estimate examples/qdrant_text_hnsw.yaml --format markdown

Multi-scenario example

uv run vector-db-sizer estimate examples/multi_scenario.yaml --format csv
uv run vector-db-sizer estimate examples/multi_scenario.yaml --format json

Output formats

• json (machine-readable)
• markdown (human report)
• csv (comparison table)

Supported engines

• generic
• pgvector
• qdrant
• milvus
• elasticsearch
• opensearch
• weaviate
• pinecone

How to interpret the report

• Raw vectors: uncompressed/base vector bytes.
• Quantized vectors: additional quantized representation when modeled.
• Record payload: IDs + metadata/text/provenance payload bytes.
• Index disk: index structure bytes on disk.
• Engine overhead: engine/profile-level overhead approximation.
• Final disk estimate: replicated storage plus WAL/snapshot/safety factors.
• Final RAM estimate: vectors + payload + index + overhead RAM approximation.
• Warnings: profile caveats and scenario assumptions to review.
• Confidence: per-component confidence levels for planning.

Confidence levels

• high: formulaic or type-level estimate.
• medium: useful engineering approximation.
• low: heuristic and engine-dependent; validate with pilot load.

Production sizing warning

The estimates are analytical and should be calibrated with a representative pilot load before production capacity planning.

Development

uv sync
uv run pytest
uv run ruff check .

Current limitations

• Engine profiles are approximate.
• No vendor pricing model.
• No actual DB measurements from live systems.
• No latency/recall estimation.
• No automatic database selection.