Free Vector Databases for RAG in 2026

A RAG app stores one vector per chunk of your documents and, for every question, asks the database for the nearest few. Seven providers give you somewhere to keep those vectors for free: five managed vector databases (Qdrant, Zilliz, Weaviate, Upstash Vector, Cloudflare Vectorize) and two general databases with vector search built in (Supabase Postgres with pgvector, Turso). Every number below was checked on the vendor's own pricing or docs page on 2026-10-21. You still need vectors to put in it: the free embeddings API guide covers that step.

Compare the free vector database tiers

The "vectors that fit" column for Upstash, Vectorize, Supabase and Turso is our arithmetic: a 768-dimension float32 vector is 3,072 bytes, 200M / 768 is about 260,000, and 5M / 768 is about 6,500. Qdrant and Zilliz publish their own estimate; Weaviate caps objects directly. Smaller embeddings (384 dimensions) fit twice as many.

Which one to pick

• Most vectors for free: Qdrant (1 GB RAM cluster) or Zilliz (5 GB), both about 1M vectors at 768 dimensions. Qdrant deletes a free cluster after 4 weeks of inactivity, so keep it in use or snapshot it.
• Serverless app, no server to keep awake: Upstash Vector. 10,000 queries and updates a day over a plain REST API, so it works from serverless functions.
• Already on Cloudflare Workers: Vectorize is in the Workers Free plan, but 5M stored dimensions is only about 6,500 vectors at 768 dimensions; it suits a small docs bot.
• Want SQL and your app data in one place: Supabase (Postgres with pgvector) or Turso (vector search built into libSQL). Filtering by user, date or tag is a plain WHERE clause.
• Built-in embeddings: Weaviate Cloud can embed your text for you (2,000 embedding requests a day), at the cost of a single collection and 100,000 objects.

Steps

Qdrant Cloud

1. Sign up at cloud.qdrant.io (no card) and create a free cluster.
2. Create an API key for the cluster and copy the cluster URL.
3. Create a collection sized to your embedding model: curl -X PUT "$QDRANT_URL/collections/docs" -H "api-key: $QDRANT_API_KEY" -H "Content-Type: application/json" -d '{"vectors": {"size": 768, "distance": "Cosine"}}'
4. Upsert points, then search with POST /collections/docs/points/query.

Upstash Vector

1. Sign in at console.upstash.com and create a Vector index; choose the dimension and similarity metric of your embedding model.
2. Copy the REST URL and token from the index page.
3. Query: curl "$UPSTASH_VECTOR_REST_URL/query" -H "Authorization: Bearer $UPSTASH_VECTOR_REST_TOKEN" -d '{"vector": [0.1, 0.2, ...], "topK": 5, "includeMetadata": true}'

Supabase (pgvector)

1. Create a free project at supabase.com.
2. In the SQL editor run create extension vector with schema extensions;
3. Add a column such as embedding extensions.vector(768), add an HNSW index, and order by distance (<=> for cosine) to fetch the nearest rows.

Turso

1. Sign up at turso.tech (no card) and create a database.
2. Declare the column as F32_BLOB(768) and index it with CREATE INDEX docs_idx ON docs (libsql_vector_idx(embedding));
3. Query with vector_top_k('docs_idx', vector32('[...]'), 5).

Zilliz, Weaviate and Vectorize follow the same shape: create the free cluster or index in the console, set the dimension to your embedding model's, copy the endpoint and key, and use the vendor's client.

When you outgrow the free tier

Three startup programs in the catalog cover a vector store for a funded company: Pinecone for Startups (fewer than 100 employees, Series A or earlier; free Standard tier, Pro Support and usage credits of an unpublished amount), MongoDB for Startups (Atlas credits, less than 7 years old and Series A or earlier; 12 months to activate, 12 months to use) and the Neon Startup Program (Postgres with pgvector; up to $1,000 self-funded or up to $200K in Neon and Databricks credits, valid 12 months). Or self-host: Qdrant (Apache-2.0), Milvus (Apache-2.0) and pgvector (PostgreSQL License) run on your own server with no vector cap.

Limits and the catch

• Inactivity kills free clusters. Qdrant suspends after 1 week and deletes after 4 weeks; Supabase pauses a project after 1 week. A demo nobody opens for a month may be gone.
• Pick the embedding size first. Upstash and Vectorize stop at 1,536 dimensions, and changing the model means re-embedding every document.
• Backups are thin or manual. Qdrant's free cluster has manual snapshots through the API only, and Turso keeps 1 day of point-in-time restore. Keep your source documents so you can rebuild the index.
• One node, no high availability. Qdrant's free tier is a single node without high availability, and Qdrant upgrades it with downtime; fine for a prototype, not for a product with users.
• Free tiers change. Offers are re-checked monthly; the verified date on each offer page is the day we last confirmed it. The RAG stack builder pairs a vector store with an embedding model and an LLM from public listings.