What is halfvec and when should you use it to reduce storage costs?

-- Standard vector: 4 bytes per dimension
-- halfvec:           2 bytes per dimension (50% storage reduction)

-- Create table with halfvec column:
CREATE TABLE documents_compact (
    id        BIGSERIAL   PRIMARY KEY,
    content   TEXT,
    embedding HALFVEC(1536)
);

-- For 1M documents at 1536 dimensions:
-- vector(1536):   1,000,000 * 1536 * 4 bytes = ~6 GB
-- halfvec(1536):  1,000,000 * 1536 * 2 bytes = ~3 GB
-- Saving: ~3 GB per million documents

-- Insert (same syntax as vector):
INSERT INTO documents_compact (content, embedding)
VALUES ('Example', '[0.1, 0.2, 0.3, ...]');
-- Values are automatically rounded to 16-bit precision

-- HNSW index for halfvec:
CREATE INDEX ON documents_compact
    USING hnsw (embedding halfvec_cosine_ops);  -- note: halfvec_cosine_ops

-- Convert from vector to halfvec:
SELECT embedding::halfvec(1536) FROM documents WHERE id = 1;

-- When to use halfvec:
-- Storage cost is significant (millions of documents, high dimensions)
-- Slight precision loss is acceptable (typically negligible for text)
-- When to stick with vector:
-- Highest accuracy required (scientific data, financial embeddings)
-- Precision is critical