batch

type Batch ¶

type Batch[T dto.Embedding] struct {
	Label string
	// contains filtered or unexported fields
}

func NewBatchVectorizer[T dto.Embedding](client BatchClient[T], maxBatchTime time.Duration, settings Settings, logger logrus.FieldLogger, label string) *Batch[T]

func (b *Batch[T]) SubmitBatchAndWait(ctx context.Context, cfg moduletools.ClassConfig, skipObject []bool, tokenCounts []int, texts []string) ([]T, map[int]error)

type BatchClient ¶

type BatchClient[T dto.Embedding] interface {
	Vectorize(ctx context.Context, input []string,
		config moduletools.ClassConfig) (*modulecomponents.VectorizationResult[T], *modulecomponents.RateLimits, int, error)
	GetVectorizerRateLimit(ctx context.Context, config moduletools.ClassConfig) *modulecomponents.RateLimits
	GetApiKeyHash(ctx context.Context, config moduletools.ClassConfig) [32]byte
}

type BatchJob ¶

type BatchJob[T dto.Embedding] struct {
	// contains filtered or unexported fields
}

type BatchSimple ¶ added in v1.35.23

type BatchSimple[T dto.Embedding] struct {
	// contains filtered or unexported fields
}

BatchSimple exposes the per-object (VectorizeBatch) and per-batch (VectorizeBatchObjects) vectorization helpers used by modules that do not need the full token-aware Batch[T] pipeline.

When a positive RPM is supplied to NewBatchSimple, requests are paced by a token-bucket limiter. One limiter token is reserved per input embedding (NOT per HTTP request) because some providers — e.g. Google Gemini — count each embedding as a separate request against the RPM quota even when multiple inputs are batched into a single HTTP call.

func NewBatchSimple[T dto.Embedding](logger logrus.FieldLogger, rpm int) *BatchSimple[T]

func (b *BatchSimple[T]) VectorizeBatch(
	ctx context.Context,
	objs []*models.Object,
	skipObject []bool,
	cfg moduletools.ClassConfig,
	vectorize objectVectorizer[T],
) ([]T, []models.AdditionalProperties, map[int]error)

func (b *BatchSimple[T]) VectorizeBatchObjects(
	ctx context.Context,
	objs []*models.Object,
	skipObject []bool,
	cfg moduletools.ClassConfig,
	vectorize batchObjectsVectorizer[T],
	batchSize int,
) ([]T, []models.AdditionalProperties, map[int]error)

type EncoderCache ¶ added in v1.27.18

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

func NewEncoderCache() *EncoderCache

func (e *EncoderCache) Get(model string) (*tiktoken.Tiktoken, bool)

func (e *EncoderCache) Set(model string, tk *tiktoken.Tiktoken)

type Settings ¶ added in v1.25.25

type Settings struct {
	TokenMultiplier    float32
	MaxTimePerBatch    float64
	MaxObjectsPerBatch int
	MaxTokensPerBatch  func(cfg moduletools.ClassConfig) int
	HasTokenLimit      bool
	ReturnsRateLimit   bool
}

type TokenizerFuncType ¶ added in v1.25.25

type TokenizerFuncType func(ctx context.Context, objects []*models.Object, skipObject []bool, cfg moduletools.ClassConfig, objectVectorizer *objectsvectorizer.ObjectVectorizer, encoderCache *EncoderCache) ([]string, []int, []bool, bool, error)

func ReturnBatchTokenizer(multiplier float32, moduleName string, lowerCaseInput bool) TokenizerFuncType

func ReturnBatchTokenizerWithAltNames(multiplier float32, moduleName string, altNames []string, lowerCaseInput bool) TokenizerFuncType