RAG Vector DB Cost 2027 Calculator
Calculate vector database cost for RAG systems — document count, chunks per document, embedding dimensions, monthly queries, and provider pricing for Pinecone, Weaviate, Qdrant Cloud, and self-hosted pgvector. Plus embedding generation cost from your model.
What Drives Vector DB Cost
Three cost drivers: storage (total vector count times dimensions times bytes per dimension, typically 4 bytes for FP32 or 1 byte for binary quantized), query volume (queries per second peak times time-of-day pattern), and embedding generation (one-time at ingest, recurring on updates). Storage cost dominates for small-query systems; query cost dominates for high-traffic chatbots. Embedding cost is a one-time line that returns when you re-embed at model upgrades.
Total Monthly Cost
Storage Cost = Vectors × Dimensions × Bytes × Provider Storage $/GB
Query Cost = Monthly Queries × Provider Per-Query $
Embedding Cost = Tokens Embedded × Embedding Model Price (one-time)
Provider Pricing as of 2026
Pinecone Serverless: storage USD 0.33/GB/month, reads USD 8.25 per million query units. Pinecone Pods: USD 70-700/month per pod. Weaviate Cloud: starts USD 25/month, scales by data size. Qdrant Cloud: USD 25/month starter, USD 0.014/hour per node larger. pgvector self-hosted: free software, USD 50-500/month managed Postgres infrastructure depending on size. Turbopuffer: USD 3.33/GB/month with included query allowance, popular for budget-constrained RAG.
Embedding Model Pricing
OpenAI text-embedding-3-small: USD 0.02 per million tokens (1536 dimensions). OpenAI text-embedding-3-large: USD 0.13 per million tokens (3072 dimensions). Cohere embed-english-v3.0: USD 0.10 per million tokens (1024 dimensions). Voyage AI voyage-3: USD 0.06 per million tokens (1024 dimensions). Self-hosted BGE-large: free per-query plus GPU hosting. Most B2B RAG systems use text-embedding-3-small as the default — cheap, good enough for English.
Common Cost Mistakes
Three errors that blow up vector DB budgets: (1) re-embedding on every nightly refresh instead of diffing for changes — multiplies embedding cost 30-100x, (2) over-chunking documents (5 chunks per page when 2 would do) — 2.5x storage cost for no quality gain, (3) running production on serverless when you have steady QPS — pods or self-hosted is 50-80 percent cheaper at sustained 10+ QPS. Always run a 30-day pilot before committing to a tier.
Sources: Pinecone Pricing 2026, Weaviate Cloud Pricing 2026, Qdrant Cloud Pricing 2026, OpenAI Embeddings Pricing 2026, Voyage AI Pricing 2026. Last updated: April 2026.
Frequently Asked Questions
Pinecone Serverless or Pods?
Serverless wins below 10 sustained queries-per-second and bursty workloads. Pods win above 50 QPS or when you need predictable monthly cost. Run 30 days on serverless first, then migrate to pods if your QPS pattern justifies fixed pricing.
How much does pgvector save versus managed Pinecone?
50-80 percent at scale. A 100M-vector Postgres + pgvector deployment on a 16-vCPU managed Postgres instance costs USD 800-1500/month total. Equivalent Pinecone Pod is USD 2500-5000. Trade-off: you operate the database and tune the IVFFlat or HNSW indexes yourself.
Should I use binary quantization to save storage?
Yes if storage cost dominates your bill. Binary quantization (Qdrant, Weaviate, Turbopuffer support) shrinks vectors 32x with minor recall loss (typically 2-5 percent). Production-friendly for most B2B RAG use cases.
What dimension count should I pick?
1024 dimensions (Cohere, Voyage) is the sweet spot for most B2B use cases — strong quality, half the storage cost of 3072 (OpenAI 3-large). Use 1536 (OpenAI 3-small) for general English, 3072 (3-large) only if you need maximum recall on diverse content.
How often do I need to re-embed?
Re-embed only when (a) you change embedding models, (b) source documents change, or (c) you change chunking strategy. Avoid blanket nightly re-embeds — use a diff-based ingest pipeline that only embeds changed chunks. Most teams overspend 5-10x by re-embedding too eagerly.
Is this tool private?
Yes. All calculations happen in your browser. Document counts, query volumes, and provider choices are never sent, stored, or shared.