embed

type Provider ¶

type Provider interface {
	// Embed generates embeddings for one or more texts. Returns one
	// vector per input text, in the same order. Implementations should
	// support batching where the underlying provider allows it.
	Embed(ctx context.Context, texts []string) ([][]float32, error)

	// ModelID returns the identifier of the model being used, for
	// tracking in the embedding_model property on nodes.
	ModelID() string

	// ContextWindow returns the model's context window in tokens.
	// Used to size chunks before embedding. Implementations should
	// auto-detect where possible (e.g., Ollama /api/show) and fall
	// back to config or DefaultContextWindow.
	ContextWindow() int
}

Provider generates vector embeddings from text. This is the shared interface with multiple implementations -- the correct Go reason to define an interface at the provider (D29).

func New(cfg config.EmbeddingConfig) (Provider, error)

type QueryEmbedder ¶

type QueryEmbedder interface {
	EmbedQuery(ctx context.Context, text string) ([]float32, error)
}

QueryEmbedder is implemented by providers that distinguish between document-time and query-time embeddings at the underlying API level. Cohere on Bedrock accepts an `input_type` field that's "search_document" when embedding indexed content and "search_query" when embedding a retrieval query; using the document type for queries degrades cosine similarity measurably.

Search-time code should type-assert its embedder for QueryEmbedder and prefer EmbedQuery when available; non-implementing providers (OpenAI, Ollama, Titan) treat queries and documents identically and can fall back to Embed.

type SetupResult ¶

type SetupResult struct {
	Configured bool
	Provider   string
	Model      string
	Messages   []string
}

SetupResult describes what the setup process found and configured.

func SetupEmbedding(ctx context.Context, cfg *config.Config) SetupResult