PostgreSQL with pgvector is the pragmatic choice for teams already using PostgreSQL. Full-stack architecture: Supabase (PostgreSQL + pgvector + Auth + Realtime) + Next.js + OpenAI embeddings. Setup: CREATE EXTENSION vector; CREATE TABLE documents (id bigserial primary key, content text, embedding vector(1536), metadata jsonb, created_at timestamp default now()); CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops) WITH (m = 16, ef_construction = 64); The HNSW index enables sub-millisecond approximate nearest neighbor search. Ingestion pipeline: chunk document -> generate embeddings via OpenAI API -> upsert to PostgreSQL with metadata. Query pipeline: embed user query -> SELECT content, 1-(embedding <=> $1) as similarity FROM documents ORDER BY embedding <=> $1 LIMIT 5 -> pass retrieved chunks + query to LLM. Row Level Security (Supabase): implement per-user document isolation with RLS policies - each user only searches their own documents. Performance: with HNSW index, pgvector handles 1M vectors at <50ms p99 latency on standard Supabase plan. Scaling: vertical scaling to Supabase Pro for larger datasets, or horizontal partitioning by user_id. When to choose pgvector: existing PostgreSQL infrastructure, <5M vectors, need ACID transactions with vector operations, cost-sensitive.