embed

type EmbedTask ¶

type EmbedTask uint8

EmbedTask classifies how the embedded text will be used. Different embedding models tune outputs differently for retrieval vs. classification vs. similarity, so the caller passes intent and the embedder maps it to whatever wire format the underlying model wants (a `task_type` API parameter on gemini-embedding-001; an in-band prompt prefix on gemini-embedding-2; ignored entirely on the fake embedder used in tests).

memmy itself only emits Unspecified, RetrievalDocument, and RetrievalQuery today — the other values are reserved for future use and listed here so the embedder layer doesn't need to grow when a future feature wants them.

const (
	EmbedTaskUnspecified EmbedTask = iota
	// EmbedTaskRetrievalDocument is used when persisting a chunk for
	// later retrieval (Service.Write). Tunes the embedding for the
	// "document being searched against" side of an asymmetric pair.
	EmbedTaskRetrievalDocument
	// EmbedTaskRetrievalQuery is used when embedding a recall query
	// (Service.Recall). Tunes the embedding for the "search input" side.
	EmbedTaskRetrievalQuery
	EmbedTaskSemanticSimilarity
	EmbedTaskClassification
	EmbedTaskClustering
	EmbedTaskCodeRetrievalQuery
	EmbedTaskQuestionAnswering
	EmbedTaskFactVerification
)

type Embedder ¶

type Embedder interface {
	Embed(ctx context.Context, task EmbedTask, texts []string) ([][]float32, error)
	Dim() int
}

Embedder produces vectors for input text. Returned vectors are NOT normalized — callers normalize at index/query boundaries.