"Which vector database should I use?" is the single most-asked question in every RAG project I've consulted on. The marketing pages all look the same. The benchmarks online are mostly synthetic.
So I ran the same 1M-document real estate RAG workload across FAISS, Pinecone, Weaviate, and ChromaDB. Same embedding model (text-embedding-3-large, 3072 dimensions), same queries, same hardware. Here's what I found.
The Setup
- Dataset: 1,000,000 real estate listings, avg 800 tokens per doc
- Embeddings: OpenAI text-embedding-3-large (3072 dim)
- Index type: HNSW where available, IVF-PQ for FAISS
- Query set: 10,000 real production queries
- Hardware: AWS c6i.4xlarge (self-hosted) or equivalent managed tier
FAISS, The Speed King
Latency (p95): 8ms
Recall@10: 0.94
Cost: Server costs only (~$250/mo)
Operability: ⭐⭐
FAISS is the fastest thing I tested, by a wide margin. When you run it in-process with your application, you can't beat it on latency. The tradeoff is that FAISS is a library, not a database. You get no multi-tenancy, no filtering language, no metadata queries, no replication. You build all of that yourself.
Use when: you have a single-process workload, embeddings are relatively static, and you care about absolute speed more than features.
Pinecone, The Managed Option
Latency (p95): 45ms
Recall@10: 0.96
Cost: $2,100/mo at this scale (s1.x1)
Operability: ⭐⭐⭐⭐⭐
Pinecone is what you pick when you don't want to think about infrastructure. Sign up, get an API key, upsert vectors, query. It scales horizontally without you lifting a finger, has solid metadata filtering, and the developer experience is polished.
The cost, however, is real. At 1M vectors with 3072 dimensions, I was paying more per month than the entire compute budget for the rest of my stack. For smaller projects it's a steal. At scale, the math gets uncomfortable.
Use when: you're a small team, you don't want to run infrastructure, and your dataset is under 10M vectors.
Weaviate, The Feature-Rich Middle Ground
Latency (p95): 22ms
Recall@10: 0.95
Cost: $600/mo self-hosted, $1,400/mo Weaviate Cloud
Operability: ⭐⭐⭐⭐
Weaviate is my current favorite. It has built-in hybrid search (BM25 + dense), a proper GraphQL/REST API, multi-tenancy, and first-class support for multi-modal embeddings. It runs in Kubernetes like any other service.
The thing that pushed me to Weaviate was the modules system, built-in rerankers, cross-encoders, generative modules. I stopped gluing tools together and let Weaviate handle the pipeline.
Use when: you need hybrid search, you're comfortable running infrastructure, and you want features that Pinecone doesn't have.
ChromaDB, The Prototyper's Best Friend
Latency (p95): 55ms
Recall@10: 0.93
Cost: Free / server costs (~$150/mo)
Operability: ⭐⭐⭐
ChromaDB wins on developer experience for small projects. It runs embedded, serverless, or in Docker. The Python API is the cleanest of the four. I use it for every prototype and internal tool.
That said, it's not yet a match for Weaviate or Pinecone at 1M+ scale. Query latencies climbed noticeably past 500k vectors. I wouldn't put it in front of production traffic over a certain threshold, but for anything under that, it's glorious.
Use when: prototyping, internal tools, datasets under 500k vectors.
The Summary Table
p95 Recall $/mo Ops
FAISS 8ms 0.94 $250 ⭐⭐
Pinecone 45ms 0.96 $2,100 ⭐⭐⭐⭐⭐
Weaviate 22ms 0.95 $600 ⭐⭐⭐⭐
ChromaDB 55ms 0.93 $150 ⭐⭐⭐My Production Choice
For Offplan's 300k+ page RAG system, I run Weaviate self-hosted on Kubernetes. Hybrid search + rerankers + operational simplicity beats everything else at that scale.
For Cortivex's knowledge graph, I use ChromaDB because it's embedded and the dataset is small.
For a new client who just wants a working RAG system in a week, I recommend Pinecone every time. The time savings justify the cost until they don't.
There's no "best" vector database. There's only the one that fits your latency budget, your scale, your operational capacity, and your wallet. Pick based on your constraints, not hype.
What I'm Watching in 2026
- LanceDB, disk-based, multi-modal, impressive benchmarks
- Qdrant, Rust performance, payload filtering is best-in-class
- pgvector + PostgreSQL, if you already run Postgres, this might be all you need
- Turbopuffer, object-storage-backed, interesting cost model
The space is still moving. Bookmark this post and check back in six months, the numbers will have shifted.