Upstash Vector

Upstash Vector is a serverless vector database from Upstash, the company known for serverless Redis and Kafka offerings. It follows the same pay-per-request pricing model that made Upstash Redis popular: you pay only for the operations you perform, with no minimum fees or idle costs. This makes it particularly attractive for applications with variable or unpredictable query patterns, such as AI agents that may go hours without queries then burst during active conversations.

The API is REST-based and designed for simplicity. You create an index specifying the embedding dimension and distance metric, then use upsert, query, and fetch operations. Each vector can carry a metadata payload for filtered search. Upstash provides SDKs for Python, TypeScript, and Go, plus a REST API that works from any HTTP client — including edge runtimes like Cloudflare Workers, Vercel Edge Functions, and Deno Deploy where traditional database drivers can't run.

This edge-runtime compatibility is Upstash Vector's key differentiator. Modern AI agents often run on serverless and edge platforms where maintaining persistent database connections is impossible. Upstash Vector's stateless HTTP API eliminates connection management entirely. Combined with Upstash Redis for caching and session state, teams can build fully serverless agent architectures with zero always-on infrastructure.

Upstash Vector includes a built-in embedding generation feature: you can send raw text instead of pre-computed vectors, and Upstash handles the embedding using models like BGE or multilingual E5. This reduces integration complexity for simple RAG applications. The index supports namespace-based isolation for multi-tenant scenarios and metadata filtering with operators for equality, range, and set membership.

For agent frameworks, Upstash provides integrations with LangChain and LlamaIndex. The @upstash/rag-chat package bundles vector search, LLM calls, and conversation history into a single high-level API — useful for quickly building conversational agents with persistent memory.

Performance and scale are constrained compared to dedicated vector databases: query latencies are typically 10-50ms (higher than in-memory solutions like Pinecone or Qdrant), and the maximum index size is limited compared to distributed solutions like Milvus. The trade-off is zero operational overhead and true per-request pricing. Upstash Vector is best suited for serverless deployments, edge-first architectures, and applications where cost predictability matters more than absolute query performance.