Home › Glossary

What is a vector database (in plain English)?

Plain-English definition · Updated 2026-11-16. Sources are the vendors' own pages and documentation: Pinecone's pricing page and "Understand Pinecone cost" docs (read/write-unit formulas, index-size formula), Weaviate's pricing page and pricing FAQ, and Zilliz Cloud's list-price guide. Rates were re-checked against the official pages on 2026-11-16; Pinecone's numbers were additionally cross-checked against two independent pricing trackers.

AITop is an independent guide. Prices and features change fast — always check the vendor's page.

The 30-second answer

A vector database is a database built to store embeddings and answer nearest-neighbour queries over them — the storage layer under RAG, semantic search and recommendations. The 2026 buyer's trap is that the four mainstream managed options bill on four incompatible models: Pinecone bills storage plus per-query read units ($0.33/GB-month; $16–18 per million read units on Standard, where one read unit is billed per GB of namespace scanned); Weaviate Cloud bills total vector dimensions plus GiB stored ($45/month Flex minimum); Zilliz Cloud bills per-CU-hour compute; open-source engines bill whatever nodes you run. Work Pinecone's own numbers and the reversal appears: one million 1536-dim vectors with metadata form a 7.15 GB index costing about $2.36/month to store — but 50,000 queries a day against it meter roughly 10.7 million read units, about $171–193/month. The query bill scales with your corpus, not your traffic.

What does a vector database actually mean?

A vector database does three jobs. It stores records that carry an embedding — the list of numbers an embedding model produces, typically 768 to 3,072 dimensions — alongside each record's ID and metadata. It builds an approximate-nearest-neighbour (ANN) index over those vectors, most commonly an HNSW graph, so that "which records are most similar to this query vector" is answered in milliseconds without comparing against every row. And it filters those results by metadata, so a search can be constrained to one tenant, one folder or one date range. That is the whole machine; everything else in the marketing is configuration.

What makes it a database decision rather than a library decision is the operational wrapper: replication, multi-tenancy, backups, and a bill you can forecast. Postgres with the pgvector extension, or a raw ANN library inside your application, covers the first two jobs free. What you are buying from a managed vector database is the third — plus the opinionated metering that comes with it. That metering is where the actual purchase decision lives, because the 2026 market has not converged on one pricing axis; it has fragmented across four, and they rank workloads differently.

The four models, from the vendors' own pricing pages (November 2026): Pinecone meters four things on serverless — storage, read units, write units, and egress — at $0.33 per GB-month of index size, $16–18 per million read units and $4–4.50 per million write units on the Standard plan (rates vary by cloud and region; Enterprise runs $24–27 and $6–6.75 with a $500/month minimum and a 99.95% SLA). Weaviate Cloud bills the total number of vector dimensions you store, plus storage in GiB and backups — its FAQ explains the choice as "a simple number you can calculate, rather than abstract Read/Write units or CPU cycles" — with a $45/month Flex minimum and a $400/month Premium minimum, and discounted dimension rates for quantized indexes. Zilliz Cloud (the managed Milvus) bills compute in Capacity Units (CU) per hour plus per-GB storage. And the open-source route — self-managed Milvus, Qdrant, Weaviate, or pgvector — bills whatever nodes you run it on, per node rather than per query.

Why does it matter when you're choosing a tool?

Because on the per-query model, the price of one query depends on how big your corpus is, not on how hard the query is. Pinecone's cost documentation is explicit about the meter: "A query uses 1 RU for every 1 GB of namespace size, with a minimum of 0.25 RUs per query." The same page adds the counterintuitive part — parameters that shape the response, like top_k, include_metadata and include_values, do not affect query cost. A query that returns 5 results and one that returns 100 from the same namespace meter identically. You are paying for the scan, not the answer.

Run the arithmetic end to end (ours, from Pinecone's published formula and rates). Pinecone's own size formula counts ID + metadata + dimensions × 4 bytes per record, and its documentation works the same example we will: 1,000,000 records, 1536 dimensions, 1,000 bytes of metadata, 8-byte IDs = a 7.15 GB index. Storage at $0.33/GB-month costs about $2.36/month. Now query it: a 7.15 GB namespace meters ~7.15 RUs per query. At 50,000 queries per day — 1.5 million queries a month — that is roughly 10.7 million RUs, or $171–193/month at Standard rates. Reads out-bill storage by about 75×.

Now hold traffic constant and change corpus size, because that is the trap. The same 1.5 million queries a day-month against a 1 GB namespace meter 1.5M RUs — $24–27/month. Identical traffic, one-seventh the corpus, one-seventh the read bill. This is the opposite of how most SaaS bills behave, and it inverts the usual optimisation instincts: shrinking top_k saves nothing, while splitting one mega-index into per-customer namespaces, or cutting metadata payload (which counts toward the billed GB), moves the number linearly. It also explains why Weaviate's dimension-based model and Zilliz's per-CU model exist: they price the corpus and the hardware, letting query volume ride free up to capacity — which is exactly the wrong deal for spiky, low-QPS workloads and the right one for hot, read-heavy ones. There is no universally cheaper option; there is only the axis your workload is heaviest on.

What is true in 2026 specifically?

Is Pinecone still the default, and what does it really cost?

Pinecone's serverless platform is the per-query model at its purest, and its pricing page keeps the entry price low enough to be misleading without the fine print. The free Starter tier includes 2 GB of storage, 2M write units and 1M read units a month — enough for a prototype. Standard requires no published minimum but starts metering at the rates above, plus $0.10/GB of egress with 100 GB/month included; Enterprise carries a $500/month minimum, 99.95% uptime SLA, and BYOC. Two cost details hide in the documentation rather than the pricing page: every fetch request costs 1 RU per 10 records (minimum 1), and every upsert request bills 1 WU per 1 KB of request with a 5 WU minimum — so 1,000 upserts of 1536-dim records with 1 KB metadata each meter about 7,140 WUs per the vendor's own table. Index writes are not the bill you should watch; queries are.

What do the alternatives actually price?

Weaviate Cloud's FAQ justifies its dimension meter directly: total vector dimensions are "a good indicator of the resources required — a simple number you can calculate." The trade is legibility versus load-blindness: your Flex bill starts at $45/month whether you run 10 queries or 10 million, until hardware capacity forces you up the tiers (Premium from $400/month with a 99.9% SLA, dedicated deployments above that). Quantization discounts the dimension rate itself — compressed indexes "yield greater efficiency, reflected in discounted list prices." Zilliz Cloud splits compute (per CU-hour, with separate on-demand query and indexing CUs) from storage (per GB-month), which is the model closest to renting hardware; it suits steady production traffic and punishes idle clusters less than a fixed node does. And the open-source route — self-managed Milvus, Qdrant or Weaviate, or pgvector inside the Postgres you already run — has zero license cost and per-node economics that win decisively at high query rates, at the price of owning the 3 a.m. pager yourself.

When do you not need one at all?

The honest 2026 answer: more often than the landing pages admit. Under roughly 10 million vectors at moderate query rates, pgvector on a Postgres instance you already pay for usually costs less in total than any managed vector bill, with recall that is competitive for most RAG workloads. Dedicated vector databases earn their keep on scale, multi-tenancy, latency SLAs and hybrid search features. A useful tells-based rule (ours): if your embedding bill is bigger than your query bill and your corpus fits on one machine, you have a Postgres problem, not a vector database problem. If read units are growing faster than your storage and your namespaces approach tens of GB, the per-query model has found your workload, and the model choice becomes a real decision. And whichever storage you pick, the vectors themselves are disposable — re-embedding after a model swap is the real switching cost, which is why the embedding choice deserves more deliberation than the database choice.

Quick checklist

Where you'll hit it

The vector database is the line item under everything retrieval-shaped on this site. In embedding it is where the "6 KB per vector" storage warning lands as an actual invoice; in reranking it is the stage whose per-query meter sits on top of the vector store's per-query meter — stack both and the retrieval bill can outgrow the model bill. In fine-tuning vs RAG it is the infrastructure the RAG half quietly assumes. In the best AI customer-service tools ranking it is what "searches your knowledge base" actually runs on, and the reason per-resolution pricing looks cheap until you check what the vendor pays for retrieval. Related terms: embedding is the content it stores, RAG is the workload that makes it load-bearing, and inference cost is the other half of the RAG bill. Full list on the AI tools glossary.

The bottom line

A vector database stores embeddings and answers nearest-neighbour queries over them — and in 2026 the four mainstream managed options bill on four different meters: Pinecone on storage plus read units where one query costs 1 RU per GB of namespace scanned ($0.33/GB-month storage, $16–18 per million RUs on Standard), Weaviate on total vector dimensions ($45/month Flex minimum), Zilliz on per-CU-hour compute, and open-source engines on the nodes you run. Pinecone's own formulas produce the number to remember: a million 1536-dim vectors cost about $2.36/month to store and about $171–193/month to query 50,000 times a day. On per-query billing, the bill tracks your corpus, not your traffic — so pick the meter that matches what your workload actually does at scale, and do not buy a managed cluster for a corpus that fits on one Postgres box.