pgvector: 'function min(vector) does not exist'

ERROR: function min(vector) does not exist means you wrote min or max over a pgvector column. The vector type has no ordering, so those aggregates are not defined for it. Aggregate the thing you actually wanted, usually a dimension count or a distance, which are plain numbers.

If the goal was "how wide are the embeddings", use max(vector_dims(embedding)): vector_dims returns an integer per row and max over integers is fine. If the goal was "the nearest chunk", do not aggregate at all: order by the distance operator and take the first row.

The error

select d.slug, count(*), vector_dims(min(c.embedding))
from knowledge_chunk c join knowledge_document d on d.id = c.document_id
group by 1;

ERROR:  function min(vector) does not exist
LINE 1: ...slug, count(*) as chunks, vector_dims(min(c.embedding)) ...
HINT:  No function matches the given name and argument types. You might need to add explicit type casts.

Why it happens

PostgreSQL's min and max are defined per type, and they need a type that can be ordered. Numbers, text and dates have an order; a 768-dimensional vector does not, so pgvector defines no min(vector). It does define avg and sum, because adding vectors and dividing by a count means something. The hint about explicit type casts is the generic message PostgreSQL prints when no overload matches, and it is misleading here: no cast will make a vector orderable.

The fix

-- the dimension of the stored vectors, per document
select d.slug, count(*) as chunks, max(vector_dims(c.embedding)) as dims
from knowledge_chunk c join knowledge_document d on d.id = c.document_id
group by 1;

-- the nearest chunk to a query vector: no aggregate, just an ordered limit
select c.text, c.embedding <=> $1 as distance
from knowledge_chunk c
order by c.embedding <=> $1
limit 6;

Apply the scalar function to the column first, then aggregate the scalar. For nearest neighbours, the pattern is order-by-distance with a limit, which is also what lets the HNSW index do the work; an aggregate would force a full scan even if it existed. If you really need one representative vector per group, avg(embedding) is defined and gives the centroid, which is occasionally what a "min" was reaching for.

Where it bit us

Season four, Part 6, while proving the knowledge table's contents for the post. Nothing in the application was affected; the EF Core query uses CosineDistance and an ordered Take, which compiles to the second form above.

Frequently asked

Which aggregates does pgvector support on a vector column?

avg and sum, which produce a vector. min and max are not defined because vectors have no ordering. For anything else, apply a scalar function such as vector_dims or a distance operator first and aggregate the number.

How do I check the dimension of embeddings stored in pgvector?

Use vector_dims(embedding) per row and aggregate that, for example max(vector_dims(embedding)) grouped by document. The column's declared type, vector(768), also fixes the width for every row.

How do I find the nearest vectors in pgvector without an aggregate?

Order by the distance operator and limit: order by embedding <=> $1 limit 6 for cosine distance. That form uses the HNSW or IVFFlat index; aggregates would not.

More decoded errors in the Fixes category, or start the season that produced this one at Part 1.