What are best practices for a production pgvector deployment?

-- Production-ready setup summary:

-- 1. Enable extension
CREATE EXTENSION IF NOT EXISTS vector;

-- 2. Create schema with appropriate type
CREATE TABLE documents (
    id         BIGSERIAL   PRIMARY KEY,
    content    TEXT        NOT NULL,
    category   TEXT,
    created_at TIMESTAMPTZ DEFAULT NOW(),
    embedding  HALFVEC(1536)   -- halfvec for 2x storage savings
);

-- 3. Add metadata index for filtering
CREATE INDEX ON documents (category);

-- 4. Bulk load data (skip vector index during this phase)
-- COPY documents FROM '/data/docs.csv';

-- 5. Update statistics after load
ANALYZE documents;

-- 6. Build vector index with tuned memory/parallelism
SET maintenance_work_mem = '4GB';
SET max_parallel_maintenance_workers = 7;
CREATE INDEX ON documents
    USING hnsw (embedding halfvec_cosine_ops)
    WITH (m = 16, ef_construction = 64);
RESET maintenance_work_mem;
RESET max_parallel_maintenance_workers;

-- 7. Verify index is used:
EXPLAIN SELECT id FROM documents
ORDER BY embedding <=> '[...]'::halfvec LIMIT 5;