เลือกตัวชี้วัดระยะห่างใน pgvector

Similarity search ranks vectors by distance. The metric you choose must match how your embedding model was trained, or results will be subtly wrong.

pgvector provides three distance operators:

• <-> Euclidean (L2) distance
• <=> Cosine distance
• <#> Negative inner product

Cosine measures angle, ignoring magnitude. It is the default for most text embedding models like OpenAI and Sentence Transformers.

L2 measures straight-line distance and is sensitive to magnitude. Use it when vector length carries meaning, such as raw feature vectors.

The inner product operator <#> returns the negative dot product (so smaller is more similar). It is fastest and ideal for normalized vectors.

Always check your embedding model's documentation:

• OpenAI text-embedding-3: cosine
• Normalized vectors: inner product (equivalent to cosine, faster)
• Geometric/spatial features: L2

If your vectors are unit-normalized, cosine and inner product give the same ranking. Inner product is cheaper, so normalize once at insert time.

Indexes are metric-specific. You must declare the operator class so the index matches your query operator.

If your index uses vector_l2_ops but you query with <=>, Postgres ignores the index and does a slow sequential scan. Keep them aligned.

Use EXPLAIN to confirm the planner uses your vector index for the chosen metric.

Rule of thumb:

• Text/semantic search: cosine
• Pre-normalized vectors: inner product
• Raw numeric features: L2

You learned pgvector's three distance operators, when to use cosine, L2, and inner product, and why your index operator class must match your query operator for fast searches.