If you're building anything that involves AI "remembering" or searching through information — a chatbot that answers questions about your docs, a recommendation engine, a semantic search bar — you'll eventually run into the term vector database. It sounds intimidating, but the idea is simple: instead of storing text and looking it up by exact keyword match, a vector database stores a mathematical fingerprint of the meaning of your data (an "embedding") and lets you search by similarity. Ask it for things "like this," and it finds them, even if the wording is completely different.
There are now dozens of options, from fully-managed clouds to open-source engines you run yourself. This guide compares 7 real, currently-active vector databases so you can pick the one that fits your project instead of the one with the loudest marketing.
The short version
If you want the least setup and don't mind a managed service, Pinecone remains the safest default. If you care about open-source and self-hosting, Qdrant and Weaviate are the two most mature choices. If you're already deep in the AI/ML Python stack and want something dead simple to embed in a prototype, Chroma is the fastest to get running. If you're operating at real scale and cost is the bottleneck, turbopuffer's object-storage architecture is worth a serious look — it's what powers search in Cursor, Notion and Linear.
Pinecone
Pinecone is the tool most people mean when they first hear "vector database" — a fully managed cloud service, no servers to run, no cluster to tune.
For who: developers who want to ship a RAG feature or semantic search without managing infrastructure.
Price: Freemium — you can create a free index to start, then pay per usage as you scale.
Forces: free tier to get started, genuinely fast search, and scales up without you having to think about sharding.
Limites: costs can climb quickly once you're past hobby scale, and the pricing model isn't always obvious to newcomers.
Verdict: the safe, no-surprises choice if you'd rather pay for convenience than manage your own database.
Qdrant
Qdrant is an open-source vector database written in Rust, offered both as self-hosted software and as a managed cloud.
For who: teams who want open-source flexibility (self-host now, migrate to cloud later) with strong raw performance.
Price: Freemium — free to self-host, or start free on Qdrant Cloud.
Forces: fully open-source, no vendor lock-in, and benchmark-fast thanks to its Rust core.
Limites: there's a learning curve if you're not used to running your own infrastructure, and it's less of a household name than Pinecone.
Verdict: a strong pick if you want Pinecone-level performance without being tied to one vendor's cloud.
Weaviate
Weaviate is another open-source vector database, built specifically with AI-native workflows in mind (built-in support for hybrid search and model integrations).
For who: teams building AI-first products who want an open-source engine with no cloud lock-in from day one.
Price: free to self-host.
Forces: open source, designed around AI use cases specifically, and no vendor lock-in.
Limites: documentation can be thinner than the bigger commercial players, and it's still building broader name recognition.
Verdict: a solid open-source alternative to Qdrant — worth comparing both before you commit to self-hosting.
Milvus
Milvus is an open-source vector database built for large-scale AI applications, with a managed cloud version (Zilliz Cloud) for teams that don't want to self-host.
For who: teams that expect to scale to very large datasets and want an engine proven at that scale.
Price: free to self-host, with a paid cloud version available.
Forces: open-source, built specifically to handle large AI application workloads, and the cloud version is fast to spin up.
Limites: pricing on the managed side isn't fully transparent up front, and it's relatively less well-known outside ML-infrastructure circles.
Verdict: worth evaluating if you expect serious scale and want an option that's been battle-tested at that size.
Chroma
Chroma is an open-source database built for AI retrieval — vector, full-text, regex and metadata search — that you can self-host or run on Chroma Cloud.
For who: developers prototyping a RAG app who want something they can `pip install` and be querying in minutes.
Price: paid, starting around $5/GB/month for hosted storage (the open-source core is free to self-host).
Forces: open-source, flexible between self-hosting and cloud, and purpose-built for AI application patterns.
Limites: technical documentation has rough edges, and hosted costs can add up if you're storing a lot of data.
Verdict: probably the fastest path from zero to a working prototype — a lot of RAG tutorials use it for exactly that reason.
LanceDB
LanceDB is an open-source database built to store and search the large piles of images, text and embeddings that AI systems train on, with production users like Netflix, Uber, Runway and Databricks.
For who: AI/ML engineering teams building RAG or large-scale multimodal search and training pipelines, not just a simple text-search feature.
Price: the open-source core is free (Apache 2.0); LanceDB Cloud/Enterprise pricing isn't public — you'll need to contact sales.
Forces: 11,000+ GitHub stars with real production adoption, $30M Series A funding, free and self-hostable core, and native support for vector, full-text and hybrid search plus SQL filters — plus SOC 2, GDPR and HIPAA certification.
Limites: enterprise/cloud pricing isn't published, it can be overkill if you just need simple vector search rather than a full data lakehouse, and its managed cloud track record is shorter than Pinecone's.
Verdict: the strongest option if your "vector search" is really part of a bigger multimodal data pipeline, not a standalone feature.
turbopuffer
turbopuffer is a vector + full-text search database built directly on object storage (like S3), designed to be cheap at scale with sub-10ms latency. It's what runs search inside Cursor, Notion and Linear.
For who: AI product teams running search or RAG at real production scale, where per-GB storage cost matters.
Price: paid, usage-based, no free tier — Launch $16/month, Scale $256/month, Enterprise $4,096/month (includes a 99.95% SLA).
Forces: object-storage architecture makes it much cheaper at scale than dedicated-server rivals, true hybrid vector + BM25 search under 10ms, and production references most developers already trust (Cursor, Notion, Linear, Anthropic).
Limites: no free tier to try it casually, and it's closed-source with no self-hosting option.
Verdict: if your product has already outgrown a hobby-scale vector database and cost-per-GB is starting to hurt, this is the one to benchmark against your current bill.
Side-by-side
| Tool | Model | Starting price | Best for |
|---|---|---|---|
| Pinecone | Managed cloud | Free tier, then usage-based | Zero-infrastructure default |
| Qdrant | Open-source + cloud | Free self-hosted / free cloud tier | Open-source performance |
| Weaviate | Open-source | Free to self-host | AI-native open source |
| Milvus | Open-source + cloud | Free to self-host | Large-scale workloads |
| Chroma | Open-source + cloud | ~$5/GB/mo hosted | Fast prototyping |
| LanceDB | Open-source + enterprise | Free core, custom enterprise | Multimodal data pipelines |
| turbopuffer | Managed, closed-source | $16/mo, no free tier | Cost-efficient scale |
Which one should you actually pick?
There's no single "best" here — the right choice depends on what stage you're at. Prototyping something this weekend? Chroma or Pinecone's free tier will get you there fastest. Building something you plan to self-host long-term? Qdrant or Weaviate are the mature open-source options. Already running a multimodal AI pipeline at a real company? LanceDB or Milvus are built for that scale. And if your current vector database bill is becoming a real line item, turbopuffer's architecture is specifically designed to bring that number down.