Start a project
Techniques Reviewed 24 Sep 2026 · 16 sources

vector database

/ˈvɛk.tɚ ˈdeɪ.təˌbeɪs/ noun Plural vector databases
Definition

Software that stores embeddings, the lists of numbers a model uses to represent content, and quickly finds the ones most similar to a query.

Examples
  • Pinecone, a commercial service launched in 2021
  • Milvus, an open-source system first developed by Zilliz
  • Postgres with the pgvector extension installed
  • The store of help-article passages behind a support chatbot
Document cards each turn into a strip of numbers and a dot on a map. A question lands on the map as a new dot, and a circle around it picks out the three nearest dots and the cards they came from.
Inkling · Stored items found by closeness to a questionFig. V1
Current meaning
2023
A store that chatbots search for passages to add to a prompt.
Also called
Vector store
The name Jimmy Lin and colleagues use in their 2023 critique.
§01

In plain words

In AI, a vector database is software that stores embeddings and finds those most similar to a query, often to supply a language model with documents.

An ordinary database answers exact requests: this customer number, that date, every row containing a given word. Systems built for such lookups are poorly suited to comparing long lists of numbers.1 A vector database answers a different kind of request: find the stored items most like this one.

It holds embeddings, lists of numbers that a model produces so that content with similar meaning gets similar numbers. Each embedding is stored with a link to the text, image or product it stands for. Given a query, the database returns the items whose embeddings sit closest. A question about sending back a jacket can then find a page on returns that shares none of its words.

The same step drives recommendations and image search.2 Since 2023 it is best known as the lookup inside retrieval-augmented generation, which fetches passages and places them in a large language model’s prompt. Investment followed. Vespa, Weaviate, Pinecone and Chroma raised a combined $200 million that year, and established databases such as MongoDB, Elastic, Redis and Oracle added vector search.3

That prompted a debate over whether a separate product is needed. MongoDB’s chief executive has called many dedicated vector products features rather than products.4 Microsoft’s documentation separates a pure vector database, kept apart from the source data, from vector search built into an existing database.5

A vector database is often confused with an embedding. The embedding is the list of numbers for one item. The vector database stores many such lists and searches them.

§02

How it works

A vector database works in two stages. When data is loaded, an embedding model turns each item, such as a passage of text, into a list of numbers. The database stores that list with an ID and details such as the source and date.

When a query arrives, it goes through the same model. The database compares the query’s numbers with the stored ones, by distance or a similar score, and returns the closest items.

Checking every stored item gives exact results but is slow once a collection reaches millions or billions of items. So the database builds an index for approximate nearest neighbor search, which gives up a little accuracy for a large gain in speed. A common index, HNSW, links items into layers of graphs. A search enters at a sparse top layer and moves down toward the closest matches. Other methods compress each list into a short code, as Facebook’s Faiss library can.

A library such as Faiss provides an index and the search. A vector database wraps an index in database features, such as updating records while searches continue and spreading data across machines.2 It also has to handle hybrid queries, which combine similarity with ordinary conditions such as a date range.6

Fig. V2 · Ranked by closenessOut of 1 · Illustrative, not measured

Can I send back a jacket that does not fit?

  1. Returns and exchanges are accepted within 30 days 0.89
  2. Our jackets run small, so check the size chart 0.64
  3. Store hours are 9 to 5 on weekdays 0.08
Similarity between a question’s embedding and three stored passages, on a scale from 0 to 1. The top match shares no words with the question. The passages and scores are invented.
§03

Origin and first known use

The name joins vector, from the Latin for carrier, to database. Mathematicians have used vector since 1846 for a quantity with size and direction; machine learning uses it for any ordered list of numbers, including an embedding. The idea came before the name. Facebook’s Faiss library built a nearest neighbor graph of a billion vectors in 2017 without calling itself a database. The name spread with products: TechCrunch described Pinecone as a vector database in January 2021, and LF AI & Data called Milvus one that June. After ChatGPT, it became best known as the store that supplies documents to a language model.

  1. 1846 Vector, from the Latin for carrier, gains its mathematical sense: a quantity with size and direction. Merriam-Webster dates this first known use to 1846.
    Merriam-Webster, “vector” merriam-webster.com →: source for 1846
  2. 2016 Malkov and Yashunin describe HNSW, which stacks graphs of stored items in layers. A search starts in the top layer and works down, scaling logarithmically.
    Malkov and Yashunin, “Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs” arXiv 1603.09320 →: source for 2016
  3. 2017 Facebook AI researchers release Faiss, a library that can search compressed vectors. On 4 GPUs it links 1 billion vectors into a nearest neighbor graph in under 12 hours.
    Johnson, Douze and Jégou, “Billion-scale similarity search with GPUs” arXiv 1702.08734 →: source for 2017
  4. 2021 TechCrunch reports that Pinecone, founded by Edo Liberty, has built a vector database that indexes and searches high-dimensional vectors for machine learning applications.
    Ron Miller, “Pinecone lands $10M seed for purpose-built machine learning database”, TechCrunch, 27 January 2021 techcrunch.com →: source for 2021
  5. 2021 Pgvector, an extension that adds vector search to the Postgres database, has its first release. Version 0.5.0 adds an HNSW index in 2023.
    pgvector changelog, versions 0.1.0 (20 April 2021) and 0.5.0 (28 August 2023) github.com →: source for 2021
  6. 2023 OpenAI’s retrieval plugin lets ChatGPT search a vector database that the developer picks from six: Milvus, Pinecone, Qdrant, Redis, Weaviate or Zilliz.
    OpenAI, “ChatGPT plugins”, 23 March 2023 openai.com →: source for 2023
  7. 2023 Lin and colleagues run vector search with OpenAI embeddings on the HNSW indexes of Lucene, an existing search library, and see no compelling case for a dedicated vector store.
    Lin, Pradeep, Teofili and Xian, “Vector Search with OpenAI Embeddings: Lucene Is All You Need” arXiv 2308.14963 →: source for 2023
  8. 2023 A survey by Pan, Wang and Li counts over 20 commercial vector database management systems, all produced within the past five years.
    Pan, Wang and Li, “Survey of Vector Database Management Systems” arXiv 2310.14021 →: source for 2023
§04

In the wild

6 sightings · 2021–2026 · oldest first
  1. Blog23 Jun 2021

    “Milvus is an open-source vector database built to manage embedding vectors generated by machine learning models and neural networks.”

    Andrew Bringaze, LF AI & Data Foundation blog lfaidata.foundation →: Andrew Bringaze, LF AI & Data Foundation blog, 23 Jun 2021Archived →: copy of Andrew Bringaze, LF AI & Data Foundation blog, 23 Jun 2021
  2. Blog23 Mar 2023

    “The retrieval plugin allows ChatGPT to search a vector database of content, and add the best results into the ChatGPT session.”

    OpenAI, ChatGPT plugins announcement openai.com →: OpenAI, ChatGPT plugins announcement, 23 Mar 2023Archived →: copy of OpenAI, ChatGPT plugins announcement, 23 Mar 2023
  3. Headline27 Apr 2023

    “Pinecone drops $100M investment on $750M valuation, as vector database demand grows”

    TechCrunch techcrunch.com →: TechCrunch, 27 Apr 2023Archived →: copy of TechCrunch, 27 Apr 2023
  4. Paper29 Aug 2023

    “The main goal of our work is to challenge the prevailing narrative that a dedicated vector store is necessary…”

    Lin et al., arXiv preprint arXiv 2308.14963 →: Lin et al., arXiv preprint, 29 Aug 2023Archived →: copy of Lin et al., arXiv preprint, 29 Aug 2023
  5. Headline20 Apr 2024

    “Why vector databases are having a moment as the AI hype cycle peaks”

    TechCrunch techcrunch.com →: TechCrunch, 20 Apr 2024Archived →: copy of TechCrunch, 20 Apr 2024
  6. Docs22 May 2026

    “A vector database is a database designed to store and manage vector embeddings, which are mathematical representations of data in a high-dimensional space.”

    Microsoft, Azure Cosmos DB documentation learn.microsoft.com →: Microsoft, Azure Cosmos DB documentation, 22 May 2026Archived →: copy of Microsoft, Azure Cosmos DB documentation, 22 May 2026
§05

Not to be confused with

  • Embedding

    An embedding is the list of numbers that stands for one item. A vector database stores many embeddings, each linked to the item it came from, and searches them for the ones nearest to a query.

    See embedding →

  • Relational database

    A relational database stores rows and columns and returns the records that match a query exactly. A vector database ranks stored items by how close their embeddings sit to the query’s. Extensions such as pgvector give Postgres both abilities.

  • Vector search library

    A library such as Faiss supplies the index and search step inside one program. A vector database wraps such an index in storage, updates and scaling across machines, as a service that applications query.

  • Knowledge graph

    A knowledge graph records named things and the stated links between them, such as a company and its founder. A vector database records no links; it finds items whose embeddings sit close together.

§06

Questions

  • What is a vector database used for?

    It finds stored items that are close in meaning to a query, which drives recommendations and image search. Since 2023 its best-known use is retrieval-augmented generation, where it supplies passages for a language model’s prompt. OpenAI’s ChatGPT retrieval plugin, announced in March 2023, searched a vector database chosen by the developer and added the best results to the conversation.

  • How is a vector database different from a regular database?

    A relational database stores rows and columns and returns records that match a query exactly, such as a customer number. Facebook AI researchers called such SQL systems impractical for similarity search, because they are built for hash lookups and one-dimensional ranges. A vector database stores embeddings and returns the items closest to the query’s embedding. Established databases, including Postgres through the pgvector extension and MongoDB, now offer vector search as well.

  • Is a separate vector database needed for RAG?

    Not always. In 2023 Jimmy Lin and colleagues ran vector search with OpenAI embeddings on Lucene, an existing search library, and saw no compelling reason to add a dedicated vector store. MongoDB’s chief executive has called many dedicated vector products features rather than products. Microsoft’s documentation distinguishes a pure vector database, kept apart from the source data, from vector search integrated into an existing database.

  • What is the difference between a vector database and an embedding?

    The embedding is the data, and the vector database is where it is kept and searched. An embedding model turns one item, such as a paragraph or a photo, into a list of numbers. A vector database holds many of those lists, each linked to its original item, and returns the ones nearest to a query’s list. Retrieval-augmented generation uses both, in that order.

  • How does a vector database search millions of items quickly?

    It avoids comparing the query with every stored item. It builds an index for approximate nearest neighbor search, which gives up a little accuracy for speed. HNSW, described by Malkov and Yashunin in 2016, arranges items in layered graphs, and a search starts at the top layer and works down. Facebook’s Faiss library can compress vectors, and on four GPUs it has linked a billion of them into a nearest neighbor graph.

§07

Watch

§09

Cite this entry

“Vector database.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/dictionary/vector-database.

First published 24 Sep 2026 · Last reviewed 24 Sep 2026 · Report an error →