dimension verification

Author: jsimnz

client/index.go

@@ -121,8 +121,7 @@ type VectorIndexDescription struct {
 	Algorithm VectorAlgorithm
 	// Metric is the distance metric used to compare vectors.
 	Metric DistanceMetric
-	// Dimensions is the length of the vectors being indexed. It must be set, except on an @embedding
-	// field, where the embedding model fixes the length and Dimensions may be left 0.
+	// Dimensions is the length of the vectors being indexed. It must be greater than zero.
 	Dimensions uint32
 	// HNSW holds HNSW-specific parameters. Non-nil when Algorithm == VectorAlgorithmHNSW.
 	HNSW *HNSWParams

internal/db/collection_index.go

@@ -519,10 +519,9 @@ func processNewIndexRequest(
 }
 
 // validateVectorIndexDescription checks and fills in the vector-specific parts of an index request.
-// The field must hold a float32 array, and dimensions must be set unless the field is an @embedding,
-// whose model fixes the vector length. It also defaults the algorithm, metric, and any missing
-// params (mutating desc.Vector), so a request made through the index API works the same as one from
-// the @index directive's vector configuration.
+// The field must hold a float32 array, and dimensions must be greater than zero. It also defaults
+// the algorithm, metric, and any missing params (mutating desc.Vector), so a request made through
+// the index API works the same as one from the @index directive's vector configuration.
 //
 // The field is guaranteed to exist here because validateIndexDescription and
 // checkExistingFieldsAndAdjustRelFieldNames run before this and already check that.
@@ -543,10 +542,13 @@ func validateVectorIndexDescription(def client.CollectionVersion, desc client.Ne
 	if !client.IsVectorEmbeddingCompatible(field.Kind) {
 		return NewErrUnsupportedVectorIndexFieldType(field.Kind)
 	}
+	if desc.Vector.Dimensions == 0 {
+		return NewErrVectorIndexMissingDimensions(fieldName)
+	}
 
-	// The config object present is what picks the algorithm, so the caller never sets one directly.
-	// Fill in the algorithm, metric, and any missing params with defaults. A nil config means an empty
-	// one, so a caller can leave it out and still get a working HNSW index.
+	// The algorithm config object present is what picks the algorithm, so the caller never sets one
+	// directly. Fill in the algorithm, metric, and any missing params with defaults. A nil HNSW config
+	// means an empty one, so a caller can leave it out and still get a working HNSW index.
 	if desc.Vector.HNSW == nil {
 		desc.Vector.HNSW = &client.HNSWParams{}
 	}
@@ -572,19 +574,7 @@ func validateVectorIndexDescription(def client.CollectionVersion, desc client.Ne
 		return err
 	}
 
-	if desc.Vector.Dimensions > 0 {
-		return nil
-	}
-
-	// No dimensions were given. That is only allowed when the field is an @embedding, since the
-	// model then fixes the dimensions. The value itself is filled in later.
-	for _, embedding := range def.VectorEmbeddings {
-		if embedding.FieldName == fieldName {
-			return nil
-		}
-	}
-
-	return NewErrVectorIndexMissingDimensions(fieldName)
+	return nil
 }
 
 // validateNoConflictingVectorIndexMetric rejects creating a vector index on a field that another

internal/db/errors.go

@@ -62,7 +62,7 @@ const (
 	errInvalidFieldValue                         string = "invalid field value"
 	errUnsupportedIndexFieldType                 string = "unsupported index field type"
 	errUnsupportedVectorIndexFieldType           string = "unsupported field type for vector index"
-	errVectorIndexMissingDimensions              string = "vector index requires dimensions unless field is an embedding"
+	errVectorIndexMissingDimensions              string = "vector index dimensions must be greater than zero"
 	errCannotIndexAccumulatedCRDTField           string = "indexing accumulated CRDT fields is not yet supported"
 	errIndexDescriptionHasNoFields               string = "index description has no fields"
 	errCreateFile                                string = "failed to create file"
@@ -606,8 +606,8 @@ func NewErrUnsupportedVectorIndexFieldType(kind client.FieldKind) error {
 	)
 }
 
-// NewErrVectorIndexMissingDimensions returns a new error indicating that a vector index request is
-// missing its dimensions, and dimensions could not be inferred from a generated embedding.
+// NewErrVectorIndexMissingDimensions returns a new error indicating that a vector index request has
+// no dimensions.
 func NewErrVectorIndexMissingDimensions(fieldName string) error {
 	return errors.New(
 		errVectorIndexMissingDimensions,

internal/db/vector_index_test.go

@@ -82,6 +82,21 @@ func vectorIndexSearch(
 	return docIDs
 }
 
+func TestValidateVectorIndexDescription_EmbeddingRequiresDimensions(t *testing.T) {
+	const fieldName = "embedding"
+	def := client.CollectionVersion{
+		Fields:           []client.CollectionFieldDescription{{Name: fieldName, Kind: client.FieldKind_FLOAT32_ARRAY}},
+		VectorEmbeddings: []client.VectorEmbeddingDescription{{FieldName: fieldName}},
+	}
+	desc := client.NewIndexRequest{
+		Fields: []client.IndexedFieldDescription{{Name: fieldName}},
+		Vector: &client.VectorIndexDescription{},
+	}
+
+	err := validateVectorIndexDescription(def, desc)
+	require.ErrorContains(t, err, "vector index dimensions must be greater than zero")
+}
+
 func TestCollectionVectorIndex_Save_InsertsIntoGraphAndIsSearchable(t *testing.T) {
 	ctx, db, col := newVectorIndexTestDB(t, 3)
 

internal/request/graphql/schema/collection.go

@@ -751,8 +751,13 @@ func vectorIndexFromAST(
 		return client.NewIndexRequest{}, ErrIndexWithInvalidArg
 	}
 
-	if config == nil {
-		return client.NewIndexRequest{}, ErrIndexWithInvalidArg
+	obj := &ast.ObjectValue{}
+	if config != nil {
+		var ok bool
+		obj, ok = config.(*ast.ObjectValue)
+		if !ok {
+			return client.NewIndexRequest{}, ErrIndexWithInvalidArg
+		}
 	}
 
 	var dimensions uint32
@@ -763,11 +768,6 @@ func vectorIndexFromAST(
 		EfConstruction: client.DefaultHNSWEfConstruction,
 		EfSearch:       client.DefaultHNSWEfSearch,
 	}
-
-	obj, ok := config.(*ast.ObjectValue)
-	if !ok {
-		return client.NewIndexRequest{}, ErrIndexWithInvalidArg
-	}
 	for _, field := range obj.Fields {
 		switch field.Name.Value {
 		case types.VectorIndexPropDimensions:

internal/request/graphql/schema/testfixtures/schema.relatedmany.gen.graphql

@@ -3205,7 +3205,7 @@ enum VectorIndexAlgorithm {
 input VectorIndexConfig {
     "Selects the vector index algorithm using its default configuration."
     alg: VectorIndexAlgorithm
-    "Vector dimensions; required unless inferable from an @embedding."
+    "Vector dimensions; must be greater than zero."
     dimensions: Int
     "Configures the HNSW algorithm."
     hnsw: HNSWIndexConfig

internal/request/graphql/schema/testfixtures/schema.relatedone.gen.graphql

@@ -3148,7 +3148,7 @@ enum VectorIndexAlgorithm {
 input VectorIndexConfig {
     "Selects the vector index algorithm using its default configuration."
     alg: VectorIndexAlgorithm
-    "Vector dimensions; required unless inferable from an @embedding."
+    "Vector dimensions; must be greater than zero."
     dimensions: Int
     "Configures the HNSW algorithm."
     hnsw: HNSWIndexConfig

internal/request/graphql/schema/testfixtures/schema.simple.gen.graphql

@@ -2583,7 +2583,7 @@ enum VectorIndexAlgorithm {
 input VectorIndexConfig {
     "Selects the vector index algorithm using its default configuration."
     alg: VectorIndexAlgorithm
-    "Vector dimensions; required unless inferable from an @embedding."
+    "Vector dimensions; must be greater than zero."
     dimensions: Int
     "Configures the HNSW algorithm."
     hnsw: HNSWIndexConfig

internal/request/graphql/schema/types/types.go

@@ -522,7 +522,7 @@ func VectorIndexInputObject(
 		Description: "Configures an approximate-nearest-neighbour index over a vector field.",
 		Fields: gql.InputObjectConfigFieldMap{
 			VectorIndexPropDimensions: &gql.InputObjectFieldConfig{
-				Description: "Vector dimensions; required unless inferable from an @embedding.",
+				Description: "Vector dimensions; must be greater than zero.",
 				Type:        gql.Int,
 			},
 			VectorIndexPropAlgorithm: &gql.InputObjectFieldConfig{

tests/integration/collection_version/vector_index_test.go

@@ -67,7 +67,7 @@ func TestCollectionVersion_VectorIndexOnFloat32ArrayWithoutDimensionsOrEmbedding
 						embedding: [Float32!] @index(kind: vector)
 					}
 				`,
-				ExpectedError: "vector index requires dimensions unless field is an embedding",
+				ExpectedError: "vector index dimensions must be greater than zero",
 			},
 		},
 	}