Struct EmbeddingSpaceId 

pub struct EmbeddingSpaceId {
    pub identity_hash: String,
    pub model_artifact: String,
    pub tokenizer_artifact: String,
    pub preprocessing: Value,
    pub pooling: Value,
    pub normalization: Value,
    pub dimension: u32,
}

Full embedding-space identity. Two providers with the same dimension and different identities produce incomparable vectors, so the identity is more than a width.

The fields below are exactly the ones the embedding_space table records (DESIGN § Table embedding_space), so a provider’s declaration and the durable row cannot describe different things.

Fields

identity_hash: String

Canonical hash over the artifact/preprocessing identity below. Derived by EmbeddingSpaceId::new — never assembled by hand, or two providers describing one space would disagree about its name.

model_artifact: String

Exact model artifact: name plus version or content hash.

tokenizer_artifact: String

Exact tokenizer artifact, on the same terms.

preprocessing: Value

Declared preprocessing, pooling and normalization configuration. A different pooling rule over identical weights still yields vectors that must not be compared, so these are part of the identity rather than documentation of it.

pooling: Value

normalization: Value

dimension: u32

Implementations

impl EmbeddingSpaceId

pub fn new( model_artifact: impl Into<String>, tokenizer_artifact: impl Into<String>, preprocessing: Value, pooling: Value, normalization: Value, dimension: u32, ) -> EmbeddingSpaceId

Build an identity and derive its canonical hash.

The hash lives here rather than in each provider because it is the name readiness compares against: the ONNX plugin, a remote plugin and the deterministic fake must all arrive at the same string for the same space, and at different strings for different ones.

A provider’s preprocessing, pooling and normalization blobs are frozen once it is released. Their shape is written in the provider’s code, and the hash is over the blob as built (after key order and integral numbers are normalized), not over what it means: adding a key, renaming one or changing how a value is spelled gives every deployment of the new version a different identity, and the gear blocks vector search over what the old version stored until it is re-embedded. Values taken from configuration belong in a blob, since a different setting there is a different space; a change of shape needs the same deliberation as a change of model.

Trait Implementations

impl Clone for EmbeddingSpaceId

fn clone(&self) -> EmbeddingSpaceId

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for EmbeddingSpaceId

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Eq for EmbeddingSpaceId

impl PartialEq for EmbeddingSpaceId

fn eq(&self, other: &EmbeddingSpaceId) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for EmbeddingSpaceId