Upstash Vector
Serverless vector database by Upstash with per-request pricing. Stores and queries high-dimensional vector embeddings for similarity search, RAG (retrieval-augmented generation), and semantic search use cases. Designed for edge and serverless deployments with HTTP API.
Score Breakdown
⚙ Agent Friendliness
🔒 Security
HTTPS enforced. Read-only and read-write tokens enable least-privilege access. SOC 2 Type II. Token isolation per index limits blast radius of key compromise.
⚡ Reliability
Best When
You need serverless, edge-compatible vector search with per-request pricing and no infrastructure management, for RAG or semantic search in agent workflows.
Avoid When
You need high-throughput (>500 QPS) vector search or have large collections requiring complex indexing strategies.
Use Cases
- • Store document embeddings for RAG pipelines where agents retrieve relevant context before generating responses
- • Semantic similarity search for agent memory systems — find the most relevant past interactions or knowledge chunks
- • Hybrid search (vector + metadata filters) for agent knowledge bases with structured and unstructured data
- • Edge-compatible vector search from Cloudflare Workers, Vercel Edge Functions, and Deno Deploy without connection pooling
- • Zero-infrastructure vector search for agent prototypes — no cluster to manage, pay-per-request pricing
Not For
- • Very high throughput vector search (>1000 QPS) — managed Pinecone, Qdrant, or Weaviate clusters scale better
- • On-premises deployments — Upstash Vector is cloud-only
- • Collections larger than 1B vectors — check Upstash limits for your plan
Interface
Authentication
API key passed as Authorization Bearer header or UPSTASH_VECTOR_TOKEN env variable. Read-only and read-write tokens available. Token is per-index — each vector index has its own token.
Pricing
True pay-per-request pricing. No idle costs. Competitive with Pinecone for low-to-medium workloads. Free tier suitable for prototyping.
Agent Metadata
Known Gotchas
- ⚠ Vector dimensions must match the index configuration — inserting a 1536-dim vector into a 768-dim index will error; agents must create separate indexes per embedding model
- ⚠ Upstash Vector does not embed text — agents must generate embeddings before inserting; Upstash has a separate Embeddings API for this but it's a separate call
- ⚠ Free tier has hard query limits per day — agents hitting the free tier limit will receive 429 errors rather than degraded service
- ⚠ Namespaces are supported but not the default — agents querying without specifying a namespace search the default namespace only
- ⚠ Metadata values must be strings, numbers, or booleans — complex nested JSON metadata must be serialized to strings before storage and deserialized after retrieval
Alternatives
Full Evaluation Report
Detailed scoring breakdown, competitive positioning, security analysis, and improvement recommendations for Upstash Vector.
Scores are editorial opinions as of 2026-03-06.