embedding

type Adapter ¶

type Adapter interface {
	// Embed converts text to a normalised float32 vector.
	Embed(ctx context.Context, text string) ([]float32, error)
}

Adapter defines the contract for all embedding backends. Implementations: voyage.go, ollama.go.

Implementations MUST return L2-normalised vectors (‖v‖₂ = 1). The search layer assumes this when interpreting distance scores — an L2 distance threshold of 1.0 corresponds to cos_sim ≈ 0.5 only for unit vectors. Both Voyage AI and Ollama nomic-embed-text return normalised vectors by default; any new adapter that does not must normalise before returning.

All embeddings in the DB must use the same model and dimensions.

type OllamaAdapter ¶

type OllamaAdapter struct {
	// contains filtered or unexported fields
}

OllamaAdapter implements EmbeddingAdapter using a local Ollama server.

func NewOllamaEmbeddingAdapter(baseURL, modelName string) *OllamaAdapter

func (o *OllamaAdapter) Embed(ctx context.Context, text string) ([]float32, error)

type VoyageAdapter ¶

type VoyageAdapter struct {
	// contains filtered or unexported fields
}

VoyageAdapter implements EmbeddingAdapter using the Voyage AI REST API.

func NewVoyageAdapter(apiKey, model string) *VoyageAdapter

func (v *VoyageAdapter) Embed(ctx context.Context, text string) ([]float32, error)