Turning text into vectors (Effect's EmbeddingModel).
make(embed_many=...) builds the service from a provider's batch call.
embed goes through a RequestResolver, so concurrent embed calls
become one provider call (grouped by batch_key of the caller's
context); embed_many calls the provider directly.
Values#
| Attribute | Type | Description |
|---|---|---|
EMBEDDING_USAGE | Bottom[EmbeddingUsage, Mapping[str, object], Never, Never] | |
EMBED_RESPONSE | Bottom[EmbedResponse, Mapping[str, object], Never, Never] | |
EMBED_MANY_RESPONSE | Bottom[EmbedManyResponse, Mapping[str, object], Never, Never] |
Types#
| Type | Definition | Description |
|---|---|---|
EmbedMany | type EmbedMany = Callable[[ProviderOptions], Effect[ProviderResponse, E.AiError]] |
Classes#
DimensionsEmbeddingUsageEmbedResponseEmbedManyResponseProviderOptionsProviderResponseEmbeddingRequestServiceEmbeddingModel
Functions#
make#
make(*, embed_many, batch_key=None) -> Effect[Service]An embedding model from a provider's batch call.
| Parameter | Type | Description |
|---|---|---|
embed_many | EmbedMany | |
batch_key | Callable[[Context], Hashable] | None = None |
Returns Effect[Service]
def make(
*, embed_many: EmbedMany, batch_key: Callable[[Context], Hashable] | None = None
) -> Effect[Service]:
"""An embedding model from a provider's batch call."""
def resolve(entries: Sequence[Entry[EmbeddingRequest]], _: object) -> Effect[None, E.AiError]:
options = ProviderOptions(tuple(entry.request.input for entry in entries))
def complete(embeddings: tuple[EmbedResponse, ...]) -> None:
for entry, embedding in zip(entries, embeddings, strict=True):
entry.complete_unsafe(fx.Success(embedding))
return (
fx.suspend(lambda: embed_many(options))
.flat_map(lambda response: _vectors(len(entries), response))
.map(complete)
)
def group(entry: Entry[EmbeddingRequest]) -> Hashable:
return None if batch_key is None else batch_key(entry.context)
def build() -> Service:
resolver: Any = RequestResolver[EmbeddingRequest, Any].make_grouped(
key=group, resolver=resolve
)
return Service(resolver.with_span("EmbeddingModel.resolver"), embed_many)
return fx.sync(build)