Direct answer: vector search cost has four layers: embedding, storage, read/write units and rerank. At official prices, embedding 1 million chunks (500 tokens each) costs about USD 10–80 (OpenAI small 10, large 65, Gemini 75, Pinecone managed 40–80); 1 million 1536-dimension vectors take about 6.1 GB, costing roughly USD 2/month to store; initial writes about 4–4.5; read units for 300,000 queries about 5. The line that dominates is rerank: at USD 2 per 1,000 requests, 300,000 queries cost USD 600 — dozens of times the other three combined.
This guide is for technical and procurement teams building RAG, semantic search or recommendations. Data cut-off: 1 November 2026; prices come from official Pinecone, OpenAI and Google Cloud pages, in USD, excluding tax, with no FX conversion. The examples are editorial models, not quotes; dimensions, chunk counts and query volume must be replaced with your own data.
Embedding and storage are cheap; read/write units depend on query volume; rerank is billed per request and is the easiest line to lose control of. Confirm how many read units a query consumes and whether every query really needs reranking.
How is vector search billed?
Embedding per token, storage per GB-month, reads and writes per unit (per million), rerank per request (per thousand) and egress per GB. Five units, so calculate each and add them up.
| Item | Official unit and price (Pinecone Standard) | Note |
|---|---|---|
| Storage | USD 0.33/GB-month | Index data; metadata overhead follows the docs |
| Writes | USD 4–4.50 per million write units | Varies by cloud and region |
| Reads | USD 16–18 per million read units | Varies by cloud and region |
| Rerank | USD 2 per 1,000 requests | Per request, not per document |
| Egress | USD 0.10/GB, 100 GB/month included | Cross-region restore priced separately |
| Managed embedding | llama-text-embed-v2 USD 0.16/M tokens; multilingual-e5-large 0.08 | Or bring your own embeddings |
Source: Pinecone pricing (1 November 2026). How many units a query or vector consumes follows the official docs and invoice; this article does not infer the conversion ratio.
What does embedding 1 million chunks cost?
At 500 tokens per chunk (500 million tokens): OpenAI small about USD 10, large about 65, Gemini Embedding about 75, Pinecone managed 40 (e5) to 80 (llama).
| Option | Unit price (per 1M tokens) | 500M tokens | Note |
|---|---|---|---|
| OpenAI text-embedding-3-small | USD 0.02 | 10 | 1536 dimensions, value tier |
| OpenAI text-embedding-3-large | USD 0.13 | 65 | 3072 dimensions, double storage |
| Google Gemini Embedding | USD 0.15 (0.00015/1,000) | 75 | Online requests |
| Pinecone multilingual-e5-large | USD 0.08 | 40 | Managed inference |
| Pinecone llama-text-embed-v2 | USD 0.16 | 80 | Managed inference |
Editorial model: 1,000,000 chunks × 500 tokens = 500 million tokens; unit price × 500. Excludes retries, failures, chunk overlap and metadata; changing models means re-embedding and paying again.
What do storage and read/write units cost?
1536 dims × 4 bytes × 1M ≈ 6.1 GB, about USD 2/month; 3072 dims is about 12.3 GB and USD 4/month. Initial writes cost about 4–4.5, and read units for 300,000 queries about 5 (assuming 1 read unit per query).
| Item | Assumption | Volume | Amount |
|---|---|---|---|
| Storage (1536 dims) | 4 bytes/dim, 1M vectors | About 6.1 GB | About USD 2.0/month |
| Storage (3072 dims) | Same | About 12.3 GB | About USD 4.1/month |
| Initial writes | 1 vector ≈ 1 write unit (confirm in docs) | 1M | USD 4–4.5 |
| Monthly updates | 10% of chunks rewritten | 100k | USD 0.4–0.45 |
| Query reads | 300k queries, 1 read unit each | 300k | USD 4.8–5.4 |
Editorial model: capacity = vectors × dimensions × 4 bytes; storage at USD 0.33/GB-month. Read/write unit conversion is an assumption to confirm with Pinecone's docs and invoice; metadata, replicas and index overhead can raise real storage.
Why is rerank the most expensive line?
Because rerank is billed per request: USD 2 per 1,000, so 300,000 requests cost USD 600, while embedding, storage and read/write together are usually tens of dollars. Whether every query reranks decides the order of magnitude.
| Monthly queries | Rerank cost (USD 2/1,000) | Embedding + storage + read/write |
|---|---|---|
| 10,000 | 20 | About USD 12–92 (one-off/month) |
| 100,000 | 200 | About USD 17–97 |
| 300,000 | 600 | About USD 12–92 |
| 1,000,000 | 2,000 | About USD 60–140 |
Editorial model: rerank at USD 2/1,000 requests; the comparison column roughly adds first-year embedding (10–80) + storage 2 + writes 4.5 + read units 5.4, which grows with queries but stays far below rerank. Real rerank may process multiple documents per request; unit definitions follow the vendor.
When should you avoid a dedicated vector database?
For small corpora, existing Postgres, or keyword-only retrieval, pgvector, full-text or in-memory search is cheaper; a dedicated vector database earns its keep on scale, filtering, hybrid search and managed operations.
| Scenario | Start with | Why | Do not ignore |
|---|---|---|---|
| Under 100k chunks | pgvector / in-memory index | Saves managed fees and sync work | Backup, permissions, scaling |
| Existing Postgres team | pgvector | Reuse operations, transactional consistency | Index parameters, recall |
| Keyword/ID lookup | Full-text index | No embeddings needed | Synonyms, typos |
| Millions + metadata filters | Managed vector DB | Filtering, hybrid search, SLA | Read units and egress |
| High-concurrency semantic search | Managed + rerank budget | Quality and latency | Rerank volume control |
Editorial guidance, not a vendor ranking; compare recall, latency and cost per query on the same query set before launch.
What should procurement and acceptance check?
Unit definitions, storage basis, read/write prices, rerank request definition, egress and backups, and the cost of re-embedding when models change — not just the monthly plan or embedding headline.
| Acceptance item | Suggested definition | Evidence to keep |
|---|---|---|
| Unit definition | Read/write units per query and per vector | Official docs, sample invoice |
| Storage basis | Whether metadata, replicas and index overhead are included | Console usage, invoice |
| Rerank | Called per query? How a request is counted | Call logs, invoice |
| Re-embedding | Frequency and cost when models or chunking change | Change log |
| Egress and backup | Cross-region, backup, restore prices | Price list, drill |
| Recall and latency | recall@k and P95 on one query set | Evaluation report |
| Permissions | Namespaces, RBAC, audit | Config, audit logs |
| Exit | Export format and migration drill | Export sample, contract |
Editorial guidance, not legal advice; assess personal data and cross-border transfer separately.
Who should read this — and who should not?
For teams building RAG, semantic search or recommendations. Not for a few hundred documents, or projects without a query volume and recall target.
Read if: you are in engineering, data or procurement; you are comparing Pinecone, pgvector and self-hosting; you need a budget for 1M chunks and 300k queries. Skip if: your corpus is tiny; you have no evaluation set; you want this article to pick a vendor. Supports: itemised embedding/storage/read-write/rerank modelling, hybrid retrieval choices and acceptance criteria.
Lingyi does not rank vector databases or promise recall. To connect retrieval, permissions and model routing to an enterprise knowledge base, see LingHarness; to fold retrieval cost into project budgets, talk to us.
Related reading: RAG knowledge-base total cost, knowledge-base content governance, and data collection cost; this article only covers the retrieval layer (embedding/storage/read-write/rerank).
FAQ
How is vector search billed?
Embedding per token, storage per GB-month, reads and writes per unit (per million), rerank per request (per thousand) and egress per GB. Pinecone list prices: storage USD 0.33/GB-month, writes 4–4.5 per million, reads 16–18 per million, rerank USD 2 per 1,000 requests.
What does embedding 1M chunks cost?
At 500 tokens per chunk (500M tokens): OpenAI small about USD 10, large about 65, Gemini Embedding about 75, Pinecone e5 about 40, llama about 80. Changing models or chunking means re-embedding and paying again.
Is storage expensive?
No. 1536 dims × 4 bytes × 1M ≈ 6.1 GB, about USD 2/month at 0.33/GB-month; 3072 dims ≈ 12.3 GB, about USD 4/month. Metadata, replicas and index overhead can raise actual usage.
Why does rerank exceed everything else?
Because it is billed per request: USD 2 per 1,000, so 300,000 queries cost USD 600, while embedding, storage and read/write together are usually tens of dollars. Controlling "rerank every query" is the key to the bill.
How many read/write units does a query use?
The vendor bills in units, but how many units a query or a vector consumes must be confirmed in the docs and invoice; this article's "1 query = 1 read unit, 1 vector = 1 write unit" is an example assumption, not an official definition. Run a real query batch and reconcile.
Do small corpora need a vector database?
Not necessarily. Under 100k chunks, with existing Postgres, or for keyword-only search, pgvector, full-text or in-memory search is cheaper; a dedicated vector DB pays off on scale, metadata filtering, hybrid search and managed operations.
How should we estimate our own monthly bill?
Four lines: one-off embedding (tokens × price), storage (GB × 0.33), reads/writes (queries/updates × unit price), rerank (requests × 2/1,000). Run 1,000 real queries, record unit consumption and recall, then extrapolate.
Method and limitations
This article uses only official pages accessible on 1 November 2026: Pinecone pricing, OpenAI API pricing and Google Cloud Vertex AI pricing. All amounts are converted from public list prices and example assumptions; no enterprise agreements or reseller quotes are used. USD is not converted to RMB and no vendor ranking is implied. Pinecone's read/write unit conversion, storage basis, regional prices and rerank request definition follow its docs and invoice; the "1 query = 1 read unit, 1 vector = 1 write unit" figures are example assumptions. Chunk counts, tokens, dimensions, query volume and update ratios are examples, not benchmarks. The article does not assess recall, latency or stability and is not legal advice.
Sources and basis
Data cut-off: 1 November 2026. The following are official pages; prices, unit definitions and terms follow the original text and invoice.
- Pinecone, Pricing: pinecone.io/pricing. Storage USD 0.33/GB-month, writes 4–4.50 per million, reads 16–18 per million, rerank USD 2/1,000, egress 0.10/GB; managed embedding 0.08–0.16 per million tokens.
- OpenAI, API pricing: platform.openai.com/docs/pricing. text-embedding-3-small USD 0.02 and text-embedding-3-large USD 0.13 per million tokens.
- Google Cloud, Vertex AI generative AI pricing: cloud.google.com/vertex-ai/generative-ai/pricing. Gemini Embedding input USD 0.00015 per 1,000 (online), about USD 0.15 per million tokens.