Update:Our Privacy Policy has changed

Serverless vector database for AI and search

Upstash Vector stores embeddings and finds the closest matches with low latency. Upsert raw text and let a built-in model embed it, combine semantic and keyword search, filter on metadata, and pay per request, so an idle index costs only its storage.

Why Upstash Vector

A vector database that scales from a prototype to production without a cluster to size.

Built-in embeddings

Pick an embedding model when you create the index, then upsert and query plain text. No separate embedding API to call.

Hybrid search

Combine dense semantic vectors with sparse keyword vectors in one index for results that match both meaning and exact terms.

Metadata filtering

Attach metadata to each vector and filter queries with a familiar SQL-like syntax.

Namespaces

Partition one index into isolated namespaces, one per tenant, user, or dataset, and query each on its own.

Scales on disk

A disk-based index keeps large datasets fast and affordable, from thousands to millions of vectors.

REST API

Connectionless HTTP access from serverless and edge runtimes, with SDKs for TypeScript, Python, and Go.

AI framework ready

First-class integrations with LangChain, LlamaIndex, the Vercel AI SDK, Flowise, and Langflow.

Pay per request

Scale to zero on pay as you go: $0.40 per 100K requests and $0.25 per GB stored, with a free tier to start.

Add vector search to your app

Upsert and query with the @upstash/vector and upstash-vector SDKs, or call the REST API from anywhere.

import { Index } from "@upstash/vector";

const index = new Index()

const question = "What is Quantum Mechanics?"

// Find related documents
const context = await index.query({
  data: question,
  topK: 5,
  includeMetadata: true,
});

const prompt = `Question: ${question} - Context: ${JSON.stringify(context)}`;

const response = await openai.chat.completions.create({
  model: "gpt-5.4",
  messages: [
    { 
      role: "system", 
      content: "You are a helpful assistant, answer the question using the context." 
    },
    { 
      role: "user", 
      content: prompt 
    },
  ],
});

What is a vector database?

A database that finds things by meaning, not by matching text.

A vector database stores embeddings: lists of numbers that an AI model produces to represent the meaning of text, images, or other data. Items with similar meaning end up close together, so the database can answer “what is most like this?” by finding the nearest vectors to a query.

That is the retrieval step behind RAG, semantic search, recommendations, and agent memory. Doing it well at scale takes an approximate nearest neighbor index, metadata filters, and a way to keep embeddings in sync with the source data, which is a lot to run yourself.

Upstash Vector is that database as a serverless service. Create an index, choose an embedding model or bring your own vectors, and query over HTTP from any runtime. Storage scales on disk, and on pay as you go you are billed per request, so an index that is not being queried costs only its storage.

Frequently asked questions

Upstash Vector is a serverless vector database for storing embeddings and running similarity search. You create an index in the console and upsert and query over a REST API or the TypeScript, Python, and Go SDKs. It is used for RAG, semantic search, recommendations, and memory for AI agents.

A vector database stores embeddings, the numeric representations of meaning that AI models produce for text, images, and other data, and finds the vectors closest to a query. Because similar items have nearby vectors, it can search by meaning instead of exact keywords, which is the retrieval step behind RAG and semantic search.

No. When you create an index you can choose a built-in embedding model, then upsert and query with plain text and Upstash Vector embeds it for you. You can also bring vectors from any model, such as OpenAI or Hugging Face, and upsert them directly.

Yes. A hybrid index stores a dense vector for semantic meaning and a sparse vector for keyword relevance for each item, and combines both scores at query time. This helps when exact terms like product names, codes, or identifiers must match.

Yes. Each vector can carry metadata, and queries accept a filter written in a SQL-like syntax, so you can restrict results by fields such as category, date, or user while still ranking by similarity.

Use namespaces. A namespace is an isolated partition inside an index, and queries only see the namespace they target. This is a common way to build multi-tenant AI apps on a single index.

Yes. Upstash Vector has integrations for LangChain, LlamaIndex, the Vercel AI SDK, Flowise, and Langflow, so it can be dropped in as the vector store for existing RAG pipelines.

The free plan includes up to 10,000 queries and updates a day and 1 GB of data. Pay as you go is $0.40 per 100K requests and $0.25 per GB of storage, with 200 GB of bandwidth free each month. A query or fetch counts as one request however many vectors it returns, while an upsert or delete counts one request per vector. A Fixed plan is $60 a month for up to 1M queries and updates a day, and Pro plans are available for high-volume workloads. See the pricing page for details.

Indexes can be created in N. Virginia and Ireland on AWS, and in Iowa on GCP. Fixed and Pro plans are available in the AWS regions.

Create your first index in seconds

Start on the free tier with 10,000 queries and updates a day and 1 GB of data. No credit card required.