Constructor EmbeddingEndpoint

Namespace

MailFathom.AI.Embeddings

Assembly

MailFathom.AI.dll

EmbeddingEndpoint(string, EmbeddingProfileIdentity, Uri?, string, bool)

One route to a vector space: where to send a request, what to route it to, and how to authenticate.

public EmbeddingEndpoint(string Alias, EmbeddingProfileIdentity Identity, Uri? Address, string RoutedModelName, bool SupportsRequestedDimension)

Parameters

Alias string

The deployment's own name for this endpoint, which is what a log, a metric, and a failure message call it.

Identity EmbeddingProfileIdentity

The geometry this endpoint produces vectors in. Every endpoint of one chain declares the same one, and startup refuses a chain where they differ.

Address Uri

The base address requests are sent to, or null for the provider's own default.

RoutedModelName string

What is sent as the model of a request, which for a cloud deployment is the deployment's name rather than the vendor's model identifier.

SupportsRequestedDimension bool

Whether the endpoint honours a requested vector width, so a narrower space is asked for rather than cut out of a wider answer.

Remarks

The alias exists because everything else here is either an address or a credential, and neither may be written down. An address identifies a tenant and a resource, so a failure that named one would put it in every log line; the alias is a name the operator chose, which is exactly what ADR 0003 permits a message to carry.

RoutedModelName is separate from the identity's model identifier because the two answer different questions. The identity says which model produced a vector and is what makes two vectors comparable; the routed name says which string this endpoint recognizes, and a cloud deployment names that after itself. Keeping them apart is what lets one vector space be reached through a first-party API and a cloud deployment of the same model, which is the case a fallback chain exists for.

SupportsRequestedDimension is declared rather than inferred from the model name, for the reason ADR 0006 gives for the model never being a compile-time constant: a table of which models accept the parameter would be a list of model names in code, and it would be wrong the week after it was written.