platform/dbops/binaries/weaviate-src/usecases/modules/vectorizer.go

//                           _       _
// __      _____  __ ___   ___  __ _| |_ ___
// \ \ /\ / / _ \/ _` \ \ / / |/ _` | __/ _ \
//  \ V  V /  __/ (_| |\ V /| | (_| | ||  __/
//   \_/\_/ \___|\__,_| \_/ |_|\__,_|\__\___|
//
//  Copyright © 2016 - 2025 Weaviate B.V. All rights reserved.
//
//  CONTACT: hello@weaviate.io
//

package modules

import (
	"context"
	"errors"
	"fmt"
	"runtime"

	"github.com/weaviate/weaviate/entities/dto"
	enterrors "github.com/weaviate/weaviate/entities/errors"
	"github.com/weaviate/weaviate/entities/modelsext"

	"github.com/sirupsen/logrus"
	"github.com/weaviate/weaviate/entities/models"
	"github.com/weaviate/weaviate/entities/modulecapabilities"
	"github.com/weaviate/weaviate/entities/moduletools"
	"github.com/weaviate/weaviate/entities/vectorindex/dynamic"
	"github.com/weaviate/weaviate/entities/vectorindex/flat"
	"github.com/weaviate/weaviate/entities/vectorindex/hnsw"
	"github.com/weaviate/weaviate/usecases/config"
)

var _NUMCPU = runtime.NumCPU()

const (
	errorVectorizerCapability = "module %q exists, but does not provide the " +
		"Vectorizer or ReferenceVectorizer capability"

	errorVectorIndexType = "vector index config (%T) is not of type HNSW, " +
		"but objects manager is restricted to HNSW"

	warningVectorIgnored = "This vector will be ignored. If you meant to index " +
		"the vector, make sure to set vectorIndexConfig.skip to 'false'. If the previous " +
		"setting is correct, make sure you set vectorizer to 'none' in the schema and " +
		"provide a null-vector (i.e. no vector) at import time."

	warningSkipVectorGenerated = "this class is configured to skip vector indexing, " +
		"but a vector was generated by the %q vectorizer. " + warningVectorIgnored

	warningSkipVectorProvided = "this class is configured to skip vector indexing, " +
		"but a vector was explicitly provided. " + warningVectorIgnored
)

func (p *Provider) ValidateVectorizer(moduleName string) error {
	mod := p.GetByName(moduleName)
	if mod == nil {
		return fmt.Errorf("no module with name %q present", moduleName)
	}

	okVec := p.implementsVectorizer(mod)
	okRefVec := p.implementsReferenceVectorizer(mod)
	if !okVec && !okRefVec {
		return fmt.Errorf(errorVectorizerCapability, moduleName)
	}

	return nil
}

func (p *Provider) UsingRef2Vec(className string) bool {
	class, err := p.getClass(className)
	if err != nil {
		return false
	}

	cfg := class.ModuleConfig
	if cfg == nil {
		return false
	}

	for modName := range cfg.(map[string]interface{}) {
		if p.implementsReferenceVectorizer(p.GetByName(modName)) {
			return true
		}
	}

	return false
}

func (p *Provider) BatchUpdateVector(ctx context.Context, class *models.Class, objects []*models.Object,
	findObjectFn modulecapabilities.FindObjectFn,
	logger logrus.FieldLogger,
) (map[int]error, error) {
	modConfigs, err := p.getModuleConfigs(class)
	if err != nil {
		return nil, err
	}

	if len(modConfigs) == 0 {
		// short-circuit collections without vector index
		if class.Vectorizer == config.VectorizerModuleNone {
			return nil, nil
		}

		return nil, fmt.Errorf("no vectorizer configs for class %s", class.Class)
	}

	var (
		vecErrorsList = make([]map[int]error, len(modConfigs))
		errorList     = make([]error, len(modConfigs))
		counter       = 0
	)

	eg := enterrors.NewErrorGroupWrapper(logger)
	eg.SetLimit(_NUMCPU)
	for targetVector, modConfig := range modConfigs {
		shouldVectorizeClass, err := p.shouldVectorizeClass(class, targetVector, logger)
		if err != nil {
			errorList[counter] = err
			continue
		}
		if shouldVectorizeClass {
			targetVector := targetVector
			modConfig := modConfig
			counter := counter

			fun := func() error {
				vecErrors, err := p.batchUpdateVector(ctx, objects, class, findObjectFn, targetVector, modConfig)
				errorList[counter] = err
				vecErrorsList[counter] = vecErrors
				return nil // to use error group
			}
			eg.Go(fun)
		}

		counter += 1
	}
	if err := eg.Wait(); err != nil {
		return nil, err
	}

	// combine errors from different runs
	combinedErrors := make(map[int]error, 0)
	for _, vecErrors := range vecErrorsList {
		for i, vecError := range vecErrors {
			if existingErr, ok := combinedErrors[i]; ok {
				vecError = errors.Join(existingErr, vecError)
			}
			combinedErrors[i] = vecError
		}
	}

	return combinedErrors, errors.Join(errorList...)
}

func (p *Provider) shouldVectorizeClass(class *models.Class, targetVector string, logger logrus.FieldLogger) (bool, error) {
	hnswConfig, err := p.getVectorIndexConfig(class, targetVector)
	if err != nil {
		return false, err
	}

	vectorizer := p.getVectorizer(class, targetVector)
	if vectorizer == config.VectorizerModuleNone {
		return false, nil
	}

	if hnswConfig.Skip {
		logger.WithField("className", class.Class).
			WithField("vectorizer", vectorizer).
			Warningf(warningSkipVectorGenerated, vectorizer)
	}
	return true, nil
}

func (p *Provider) batchUpdateVector(ctx context.Context, objects []*models.Object, class *models.Class,
	findObjectFn modulecapabilities.FindObjectFn,
	targetVector string, modConfig map[string]interface{},
) (map[int]error, error) {
	found := p.getModule(modConfig)
	if found == nil {
		return nil, fmt.Errorf("no vectorizer found for class %q", class.Class)
	}
	cfg := NewClassBasedModuleConfig(class, found.Name(), "", targetVector, &p.cfg)

	if vectorizer, ok := found.(modulecapabilities.Vectorizer[[]float32]); ok {
		// each target vector can have its own associated properties, and we need to determine for each one if we should
		// skip it or not. To simplify things, we create a boolean slice that indicates for each object if the given
		// vectorizer needs to act on it or not. This allows us to use the same objects slice for all vectorizers and
		// simplifies the mapping of the returned vectors to the objects.
		skipRevectorization := make([]bool, len(objects))
		for i, obj := range objects {
			if !p.shouldVectorizeObject(obj, cfg) {
				skipRevectorization[i] = true
				continue
			}
			reVectorize, addProps, vector, err := reVectorize(ctx, cfg, vectorizer, obj,
				class, nil, targetVector, findObjectFn, p.cfg.RevectorizeCheckDisabled.Get())
			if err != nil {
				return nil, fmt.Errorf("cannot vectorize class %q: %w", class.Class, err)
			}
			if !reVectorize {
				skipRevectorization[i] = true
				p.lockGuard(func() {
					p.addVectorToObject(obj, vector, nil, addProps, cfg)
				})
			}
		}
		vectors, addProps, vecErrors := vectorizer.VectorizeBatch(ctx, objects, skipRevectorization, cfg)
		for i := range objects {
			if _, ok := vecErrors[i]; ok || skipRevectorization[i] {
				continue
			}

			var addProp models.AdditionalProperties = nil
			if addProps != nil { // only present for contextionary and probably nobody is using this
				addProp = addProps[i]
			}

			p.lockGuard(func() {
				p.addVectorToObject(objects[i], vectors[i], nil, addProp, cfg)
			})
		}

		return vecErrors, nil
	} else if vectorizer, ok := found.(modulecapabilities.Vectorizer[[][]float32]); ok {
		// each target vector can have its own associated properties, and we need to determine for each one if we should
		// skip it or not. To simplify things, we create a boolean slice that indicates for each object if the given
		// vectorizer needs to act on it or not. This allows us to use the same objects slice for all vectorizers and
		// simplifies the mapping of the returned vectors to the objects.
		skipRevectorization := make([]bool, len(objects))
		for i, obj := range objects {
			if !p.shouldVectorizeObject(obj, cfg) {
				skipRevectorization[i] = true
				continue
			}
			reVectorize, addProps, multiVector, err := reVectorizeMulti(ctx, cfg,
				vectorizer, obj, class, nil, targetVector, findObjectFn,
				p.cfg.RevectorizeCheckDisabled.Get())
			if err != nil {
				return nil, fmt.Errorf("cannot vectorize class %q: %w", class.Class, err)
			}
			if !reVectorize {
				skipRevectorization[i] = true
				p.lockGuard(func() {
					p.addVectorToObject(obj, nil, multiVector, addProps, cfg)
				})
			}
		}
		multiVectors, addProps, vecErrors := vectorizer.VectorizeBatch(ctx, objects, skipRevectorization, cfg)
		for i := range objects {
			if _, ok := vecErrors[i]; ok || skipRevectorization[i] {
				continue
			}

			var addProp models.AdditionalProperties = nil
			if addProps != nil { // only present for contextionary and probably nobody is using this
				addProp = addProps[i]
			}

			p.lockGuard(func() {
				p.addVectorToObject(objects[i], nil, multiVectors[i], addProp, cfg)
			})
		}

		return vecErrors, nil
	} else {
		refVectorizer := found.(modulecapabilities.ReferenceVectorizer[[]float32])
		errs := make(map[int]error, 0)
		for i, obj := range objects {
			vector, err := refVectorizer.VectorizeObject(ctx, obj, cfg, findObjectFn)
			if err != nil {
				errs[i] = fmt.Errorf("update reference vector: %w", err)
			}
			p.lockGuard(func() {
				p.addVectorToObject(obj, vector, nil, nil, cfg)
			})
		}
		return errs, nil
	}
}

func (p *Provider) UpdateVector(ctx context.Context, object *models.Object, class *models.Class,
	findObjectFn modulecapabilities.FindObjectFn,
	logger logrus.FieldLogger,
) error {
	eg := enterrors.NewErrorGroupWrapper(logger)
	eg.SetLimit(_NUMCPU)

	modConfigs, err := p.getModuleConfigs(class)
	if err != nil {
		return err
	}

	for targetVector, modConfig := range modConfigs {
		targetVector := targetVector // https://golang.org/doc/faq#closures_and_goroutines
		modConfig := modConfig       // https://golang.org/doc/faq#closures_and_goroutines
		eg.Go(func() error {
			return p.vectorizeOne(ctx, object, class, findObjectFn, targetVector, modConfig, logger)
		}, targetVector)
	}
	if err = eg.Wait(); err != nil {
		return err
	}
	return nil
}

func (p *Provider) lockGuard(mutate func()) {
	p.vectorsLock.Lock()
	defer p.vectorsLock.Unlock()
	mutate()
}

func (p *Provider) addVectorToObject(object *models.Object,
	vector []float32, multiVector [][]float32,
	additional models.AdditionalProperties, cfg moduletools.ClassConfig,
) {
	if len(additional) > 0 {
		if object.Additional == nil {
			object.Additional = models.AdditionalProperties{}
		}
		for additionalName, additionalValue := range additional {
			object.Additional[additionalName] = additionalValue
		}
	}
	if cfg.TargetVector() == "" {
		object.Vector = vector
		return
	}
	if object.Vectors == nil {
		object.Vectors = models.Vectors{}
	}
	if multiVector != nil {
		object.Vectors[cfg.TargetVector()] = multiVector
	} else {
		object.Vectors[cfg.TargetVector()] = vector
	}
}

func (p *Provider) vectorizeOne(ctx context.Context, object *models.Object, class *models.Class,
	findObjectFn modulecapabilities.FindObjectFn,
	targetVector string, modConfig map[string]interface{},
	logger logrus.FieldLogger,
) error {
	vectorize, err := p.shouldVectorize(object, class, targetVector, logger)
	if err != nil {
		return fmt.Errorf("vectorize check for target vector %s: %w", targetVector, err)
	}
	if vectorize {
		if err := p.vectorize(ctx, object, class, findObjectFn, targetVector, modConfig); err != nil {
			return fmt.Errorf("vectorize target vector %s: %w", targetVector, err)
		}
	}
	return nil
}

func (p *Provider) vectorize(ctx context.Context, object *models.Object, class *models.Class,
	findObjectFn modulecapabilities.FindObjectFn,
	targetVector string, modConfig map[string]interface{},
) error {
	found := p.getModule(modConfig)
	if found == nil {
		return fmt.Errorf(
			"no vectorizer found for class %q", object.Class)
	}

	cfg := NewClassBasedModuleConfig(class, found.Name(), "", targetVector, &p.cfg)

	if vectorizer, ok := found.(modulecapabilities.Vectorizer[[]float32]); ok {
		if p.shouldVectorizeObject(object, cfg) {
			var targetProperties []string
			vecConfig, ok := modConfig[found.Name()]
			if ok {
				if properties, ok := vecConfig.(map[string]interface{})["properties"]; ok {
					if propSlice, ok := properties.([]string); ok {
						targetProperties = propSlice
					}
				}
			}
			needsRevectorization, additionalProperties, vector, err := reVectorize(ctx,
				cfg, vectorizer, object, class, targetProperties, targetVector, findObjectFn,
				p.cfg.RevectorizeCheckDisabled.Get())
			if err != nil {
				return fmt.Errorf("cannot revectorize class %q: %w", object.Class, err)
			}
			if needsRevectorization {
				var err error
				vector, additionalProperties, err = vectorizer.VectorizeObject(ctx, object, cfg)
				if err != nil {
					return fmt.Errorf("update vector: %w", err)
				}
			}

			p.lockGuard(func() {
				p.addVectorToObject(object, vector, nil, additionalProperties, cfg)
			})
			return nil
		}
	} else if vectorizer, ok := found.(modulecapabilities.Vectorizer[[][]float32]); ok {
		if p.shouldVectorizeObject(object, cfg) {
			var targetProperties []string
			vecConfig, ok := modConfig[found.Name()]
			if ok {
				if properties, ok := vecConfig.(map[string]interface{})["properties"]; ok {
					if propSlice, ok := properties.([]string); ok {
						targetProperties = propSlice
					}
				}
			}
			needsRevectorization, additionalProperties, multiVector, err := reVectorizeMulti(ctx,
				cfg, vectorizer, object, class, targetProperties, targetVector, findObjectFn,
				p.cfg.RevectorizeCheckDisabled.Get())
			if err != nil {
				return fmt.Errorf("cannot revectorize class %q: %w", object.Class, err)
			}
			if needsRevectorization {
				var err error
				multiVector, additionalProperties, err = vectorizer.VectorizeObject(ctx, object, cfg)
				if err != nil {
					return fmt.Errorf("update vector: %w", err)
				}
			}

			p.lockGuard(func() {
				p.addVectorToObject(object, nil, multiVector, additionalProperties, cfg)
			})
			return nil
		}
	} else {
		refVectorizer := found.(modulecapabilities.ReferenceVectorizer[[]float32])
		vector, err := refVectorizer.VectorizeObject(ctx, object, cfg, findObjectFn)
		if err != nil {
			return fmt.Errorf("update reference vector: %w", err)
		}
		p.lockGuard(func() {
			p.addVectorToObject(object, vector, nil, nil, cfg)
		})
	}
	return nil
}

func (p *Provider) shouldVectorizeObject(object *models.Object, cfg moduletools.ClassConfig) bool {
	if cfg.TargetVector() == "" {
		return object.Vector == nil
	}

	targetVectorExists := false
	p.lockGuard(func() {
		vec, ok := object.Vectors[cfg.TargetVector()]
		isVectorEmpty, _ := dto.IsVectorEmpty(vec)
		targetVectorExists = ok && !isVectorEmpty
	})
	return !targetVectorExists
}

func (p *Provider) shouldVectorize(object *models.Object, class *models.Class,
	targetVector string, logger logrus.FieldLogger,
) (bool, error) {
	hnswConfig, err := p.getVectorIndexConfig(class, targetVector)
	if err != nil {
		return false, err
	}

	vectorizer := p.getVectorizer(class, targetVector)
	if vectorizer == config.VectorizerModuleNone {
		vector := p.getVector(object, targetVector)
		isEmpty, err := dto.IsVectorEmpty(vector)
		if err != nil {
			return false, fmt.Errorf("should vectorize: is vector empty: %w", err)
		}
		if hnswConfig.Skip && !isEmpty {
			logger.WithField("className", class.Class).
				Warningf(warningSkipVectorProvided)
		}
		return false, nil
	}

	if hnswConfig.Skip {
		logger.WithField("className", class.Class).
			WithField("vectorizer", vectorizer).
			Warningf(warningSkipVectorGenerated, vectorizer)
	}
	return true, nil
}

func (p *Provider) getVectorizer(class *models.Class, targetVector string) string {
	if targetVector != "" && len(class.VectorConfig) > 0 {
		if vectorConfig, ok := class.VectorConfig[targetVector]; ok {
			if vectorizer, ok := vectorConfig.Vectorizer.(map[string]interface{}); ok && len(vectorizer) == 1 {
				for vectorizerName := range vectorizer {
					return vectorizerName
				}
			}
		}
		return ""
	}
	return class.Vectorizer
}

func (p *Provider) getVector(object *models.Object, targetVector string) models.Vector {
	p.vectorsLock.Lock()
	defer p.vectorsLock.Unlock()
	if targetVector != "" {
		if len(object.Vectors) == 0 {
			return nil
		}
		return object.Vectors[targetVector]
	}
	return object.Vector
}

func (p *Provider) getVectorIndexConfig(class *models.Class, targetVector string) (hnsw.UserConfig, error) {
	vectorIndexConfig := class.VectorIndexConfig
	if targetVector != "" {
		vectorIndexConfig = class.VectorConfig[targetVector].VectorIndexConfig
	}
	hnswConfig, okHnsw := vectorIndexConfig.(hnsw.UserConfig)
	_, okFlat := vectorIndexConfig.(flat.UserConfig)
	_, okDynamic := vectorIndexConfig.(dynamic.UserConfig)
	if !(okHnsw || okFlat || okDynamic) {
		return hnsw.UserConfig{}, fmt.Errorf(errorVectorIndexType, vectorIndexConfig)
	}
	return hnswConfig, nil
}

func (p *Provider) getModuleConfigs(class *models.Class) (map[string]map[string]interface{}, error) {
	modConfigs := map[string]map[string]interface{}{}
	// get all named vectorizers for classs
	for name, vectorConfig := range class.VectorConfig {
		modConfig, ok := vectorConfig.Vectorizer.(map[string]interface{})
		if !ok {
			return nil, fmt.Errorf("class %v vectorizer %s not present", class.Class, name)
		}
		modConfigs[name] = modConfig
	}

	if modelsext.ClassHasLegacyVectorIndex(class) && class.Vectorizer != config.VectorizerModuleNone {
		if modConfig, ok := class.ModuleConfig.(map[string]interface{}); ok {
			modConfigs[""] = modConfig
		} else {
			return nil, fmt.Errorf("no moduleconfig for class %v present", class.Class)
		}
	}

	return modConfigs, nil
}

func (p *Provider) getModule(modConfig map[string]interface{}) (found modulecapabilities.Module) {
	for modName := range modConfig {
		if err := p.ValidateVectorizer(modName); err == nil {
			found = p.GetByName(modName)
			break
		}
	}
	return
}

func (p *Provider) VectorizerName(className string) (string, error) {
	name, _, err := p.getClassVectorizer(className)
	if err != nil {
		return "", err
	}
	return name, nil
}

func (p *Provider) getClassVectorizer(className string) (string, interface{}, error) {
	class := p.schemaGetter.ReadOnlyClass(className)
	if class == nil {
		// this should be impossible by the time this method gets called, but let's
		// be 100% certain
		return "", nil, fmt.Errorf("class %s not present", className)
	}

	return class.Vectorizer, class.VectorIndexConfig, nil
}

func (p *Provider) implementsVectorizer(mod modulecapabilities.Module) bool {
	switch mod.(type) {
	case modulecapabilities.Vectorizer[[]float32], modulecapabilities.Vectorizer[[][]float32]:
		return true
	default:
		return false
	}
}

func (p *Provider) implementsReferenceVectorizer(mod modulecapabilities.Module) bool {
	switch mod.(type) {
	case modulecapabilities.ReferenceVectorizer[[]float32], modulecapabilities.ReferenceVectorizer[[][]float32]:
		return true
	default:
		return false
	}
}