Vector Databases: Pinecone vs. Weaviate vs. pgvector

Vector Databases: Pinecone vs. Weaviate vs. pgvector Every RAG tutorial written in 2023 opened the same way: pip install pinecone client. Two years later, that reflex is the most expensive line of code in your side project. Pinecone still works fine — it just costs $50/month minimum before you've served a single user, and half the workloads people spin it up for would run happily on a $25 Supabase instance they already have. This post is the version of the comparison I wish someone had handed me when I was evaluating vector stores for AIQORA's semantic search. I've since talked to enough solo builders picking the wrong one — usually Pinecone, occasionally Weaviate for reasons they can't articulate — to think the market advice is stuck in an older era. The right answer in 2026 depends on exactly three things: how many vectors, how much money, and whether you need hybrid search. Everything else is noise. Below: what each of the big three actually costs, where each one breaks, and which two newer options (Turbopuffer, LanceDB) are quietly stealing the mid market from all three. The 2026 landscape shifted, and the old advice is stale The 2023 pitch for a dedicated vector database was that Postgres couldn't hack it above a few hundred thousand embeddings. That was true — HNSW support in pgvector was slow and memory hungry. It stopped being true around pgvector 0.7, and by 0.8 (shipped late 2024) the index build times dropped by roughly 30x for typical workloads. The second shift: embedding dimensions came down. OpenAI's text embedding 3 small at 512 dims via Matryoshka truncation costs one third the storage of the old 1536 dim ada 002 defaults, and quality holds. Cohere's v4 and the Voyage 3 lineup are similar. So the "you'll blow past…