Vector database cost calculator: Pinecone vs Qdrant vs Weaviate vs Milvus vs pgvector

How it is calculated

s_full = d × b
s_mem  = d × q
g_hnsw = 2 × M × 4 × 1.05      (IVF and DiskANN use dataset values)

RAM  = N × (s_mem + g) × f_engine ÷ 2^30      (+ N × meta ÷ 2^30 if metadata is kept in RAM)
Disk = N × (s_full + s_mem [if quantized] + meta + g) × f_disk ÷ 2^30
Totals = per replica × replicas

pick the smallest tier with RAM and disk at or above the need;
if none fits, use several nodes of the largest tier
cost = tier price × nodes × replicas

storage = N × (s_full + meta) in GB × storage rate
reads   = queries per month × read units per query × rate
writes  = upserts per month × write units per record × rate
cost    = max(monthly minimum, storage + reads + writes)

RAM ≥ 1.2 × N × (s_mem + g) ÷ 2^30
cost = (instance hourly rate × 730 + storage) × replicas

one-off = N × chunk tokens ÷ 10^6 × embedding price
monthly = one-off × upserts % ÷ 100

Assumptions

• N is the number of vectors, d the dimensions, b the bytes per element of the stored type (4 or 2) and q the bytes per dimension of the in-memory copy (1, 0.25 or 0.125 when quantized; otherwise b). 1 GiB = 2^30 bytes.
• f_engine (1.3) covers engine overhead and f_disk (1.5) covers snapshots and write-ahead log. Both come from the dataset.
• IVF-PQ keeps only a product-quantized copy in memory. DiskANN keeps its graph on disk. IVF and DiskANN overheads are planning assumptions.
• Usage-priced engines bill reads by namespace size; the read units per query shown are an assumption you can override. top-k is recorded but does not change any price here.
• Qdrant is modelled as self-hosted on AWS instances, not a Qdrant Cloud quote (its price list is not published in readable form). Pinecone, Weaviate and Zilliz rates are the lowest listed 'from' or range rates, and pgvector is priced on Supabase compute add-ons, so other regions, plans and hosts can cost more. Check the notes on each engine.
• A result that uses a modelled or unchecked engine price, or an IVF or DiskANN index (whose per-vector overhead is a planning assumption), carries a DATASET_UNVERIFIED notice naming what to treat as an estimate.
• Totals shown for RAM and disk are the per-replica figure as displayed times replicas. Network, backups, support and taxes are not included.

The data behind it

Three ways to reduce this number

1. Use scalar or binary quantization with rescoring: about 4 and 32 times smaller vectors in memory.
2. Truncate dimensions if your embedding model supports it.
3. Keep payloads on disk and index only the fields you filter on.

Use it from code

The same engine is available as a REST endpoint and as an MCP tool named calc_rag_vector_db_sizing_comparator. Your scenario as an API call:

curl -X POST 'https://api.x12i.com/v1/calc/rag-vector-db-sizing-comparator' \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
  "input": {
    "total_chunks": 5000000,
    "vector_dimensions": 1536,
    "vector_dtype": "float32",
    "quantization": "none",
    "metadata_bytes_per_vector": 500,
    "index_type": "hnsw",
    "hnsw_m": 16,
    "replicas": 2,
    "queries_per_second": 20,
    "top_k": 10,
    "monthly_upserts_pct": 10,
    "payload_in_ram": false,
    "engines": [
      "pinecone",
      "qdrant",
      "weaviate",
      "milvus",
      "pgvector"
    ],
    "avg_chunk_tokens": 400,
    "embedding_usd_per_mtok": 0.1,
    "preferred_vector_db": "none"
  },
  "options": {
    "explain": false
  }
}'

Copy as API call Schema: GET /v1/calc/rag-vector-db-sizing-comparator/schema

Frequently asked questions

Related calculators