API Reference

Packages

• vllm.ai/v1alpha1

vllm.ai/v1alpha1

Package v1alpha1 contains API Schema definitions for the vllm v1alpha1 API group

Resource Types

• SemanticRouter
• SemanticRouterList

APIConfig

APIConfig defines API configuration

Appears in:

• ConfigSpec

Field	Description	Default	Validation
batch_classification BatchClassificationConfig			Optional: {}

AutoscalingSpec

AutoscalingSpec defines autoscaling configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
enabled boolean	Enabled indicates if HPA is enabled	false	Optional: {}
minReplicas integer	MinReplicas is the minimum number of replicas	1	Optional: {}
maxReplicas integer	MaxReplicas is the maximum number of replicas	10	Optional: {}
targetCPUUtilizationPercentage integer	TargetCPUUtilizationPercentage is the target CPU percentage	80	Optional: {}
targetMemoryUtilizationPercentage integer	TargetMemoryUtilizationPercentage is the target memory percentage		Optional: {}

BatchClassificationConfig

BatchClassificationConfig defines batch classification configuration

Appears in:

• APIConfig

Field	Description	Default	Validation
max_batch_size integer		100	Optional: {}
concurrency_threshold integer		5	Optional: {}
max_concurrency integer		8	Optional: {}
metrics BatchMetricsConfig			Optional: {}

BatchMetricsConfig

BatchMetricsConfig defines batch classification metrics configuration

Appears in:

• BatchClassificationConfig

Field	Description	Default	Validation
enabled boolean		true	Optional: {}
detailed_goroutine_tracking boolean		true	Optional: {}
high_resolution_timing boolean		false	Optional: {}
sample_rate string	Sample rate for metrics (0.0-1.0). Stored as string to avoid float precision issues.	1.0	Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
duration_buckets string array	Duration buckets for histograms. Stored as strings to avoid float precision issues. Example: ["0.001", "0.005", "0.01", "0.025", "0.05", "0.1", "0.25", "0.5", "1", "2.5", "5", "10", "30"]		Optional: {}
size_buckets integer array			Optional: {}

CategoryModelConfig

CategoryModelConfig defines category model configuration

Appears in:

• ClassifierConfig

Field	Description	Default	Validation
model_id string			Optional: {}
threshold string	Classification threshold (0.0-1.0). Stored as string to avoid float precision issues.		Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
use_cpu boolean			Optional: {}
category_mapping_path string			Optional: {}

ClassifierConfig

ClassifierConfig defines classifier configuration

Appears in:

• ConfigSpec

Field	Description	Default	Validation
category_model CategoryModelConfig			Optional: {}
pii_model PIIModelConfig			Optional: {}

ComplexityCandidates

ComplexityCandidates defines candidate examples for complexity classification

Appears in:

• ComplexityRulesConfig

Field	Description	Default	Validation
candidates string array	List of candidate phrases or examples

ComplexityModelConfig

ComplexityModelConfig configures how the complexity signal produces its score. It mirrors global.model_catalog.modules.complexity in the router config and is passed through field for field.

The contract requirement sits here rather than on RemoteClassifierBackendConfig because it is a property of this consumer, not of the block: complexity reads two response shapes, so guessing wrong would surface per request instead of at admission. A consumer that reads one shape

• categories does - keeps the field optional and defaults it.

Appears in:

• ConfigSpec

Field	Description	Default	Validation
backend RemoteClassifierBackendConfig	Backend names a remote scoring model. Its absence keeps local prototype scoring; when set, the signal never reads the rules' hard/easy candidates. It sits on the module rather than on a rule because routing signals are replaced wholesale per recipe, so a per-rule backend would vanish under any recipe that did not repeat it.		Optional: {}

ComplexityRulesConfig

ComplexityRulesConfig defines complexity-based signal classification.

The CEL rules below reject at admission the boundary combinations the Router refuses at config load. Without them the API server accepts the object and the Router crashloops on it, which turns a typo into an outage instead of a rejected write. They are per-object and static; anything needing the model catalog - whether backend.model resolves, for instance - stays with the Router's validator, which remains the single source of truth for the rest.

Appears in:

• ConfigSpec

Field	Description	Default	Validation
prototype_scoring PrototypeScoringConfig	PrototypeScoring replaces the family prototype-scoring configuration for this rule. Absence inherits the family; a present object is a complete override, including an empty object. Defaults remain Router-owned.		Optional: {}
name string	Name of the complexity rule (e.g., "code-complexity", "reasoning-complexity")
description string	Description of what this rule classifies		Optional: {}
threshold string	Threshold for the local prototype-scoring path (0.0-1.0), stored as a string to avoid float precision issues. The local margin is hard-minus-easy and centred on zero, so the threshold is symmetric: a margin above it is "hard", below its negative is "easy", and in between is "medium". It does not apply under a score.v1 backend, whose score is in the model's own units; state a boundary pair instead.		Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
hard_above string	HardAbove and EasyBelow are the two cut points for a score where a higher value is harder, in the scoring model's own units - so no [0,1] pattern applies and negative values are valid. Both are required together, and the pair is mutually exclusive with Threshold and with HardBelow/EasyAbove. Stored as strings to avoid float precision issues.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}
easy_below string	EasyBelow is the lower cut point of the harder-when-higher pair: a score below it is "easy", and anything between EasyBelow and HardAbove is "medium". It must be below HardAbove, and both are required together.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}
hard_below string	HardBelow and EasyAbove are the pair for a score where a lower value is harder - a model predicting the chance of a correct answer, say. They require a score.v1 backend: the local margin is harder-when-higher by construction, and inverting it locally means swapping the candidate lists. Stored as strings to avoid float precision issues.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}
easy_above string	EasyAbove is the upper cut point of the harder-when-lower pair: a score above it is "easy", and anything between HardBelow and EasyAbove is "medium". It must be above HardBelow, and both are required together.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}
hard ComplexityCandidates	Hard candidates represent complex/difficult examples. Read only by the local path; a remote backend never consults them, so they are optional.		Optional: {}
easy ComplexityCandidates	Easy candidates represent simple/easy examples. Read only by the local path; a remote backend never consults them, so they are optional.		Optional: {}
composer RuleComposition	Composer allows filtering based on other signals (e.g., only apply this rule if domain:medical)		Optional: {}

CompositionCondition

CompositionCondition defines a single composition condition

Appears in:

• RuleComposition

Field	Description	Default	Validation
type string	Type of signal to check (e.g., "domain", "language", "category")
name string	Name of the specific signal/rule value to match

ConfigSpec

ConfigSpec defines the semantic router configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
routing JSON	Routing contains canonical v0.3 routing configuration under config.routing. It is intentionally preserved as an object so the operator can pass through the router-owned signal, projection, decision, and algorithm contract without lagging behind every router schema addition.		Type: object Optional: {}
decision_model DecisionModelBinding	DecisionModel selects the declared deployment that answers default judgment tasks. Omitted uses the Router's primary deployment. Artifact identity belongs in model_deployments; the reference never infers a model family or alias.		Optional: {}
model_deployments JSON	ModelDeployments contains canonical global.model_catalog.deployments. The router validates provider, device, precision and task compatibility.		Type: object Optional: {}
model_admission JSON	ModelAdmission contains canonical global.model_catalog.admission budgets. Keys name deployments or the router's existing admission consumers.		Type: object Optional: {}
embedding_models EmbeddingModelsConfig	Embedding models configuration (qwen3, gemma, mmbert)		Optional: {}
response_cache SemanticCacheConfig	Response cache configuration.		Optional: {}
semantic_cache SemanticCacheConfig	SemanticCache is the deprecated response-cache field.		Optional: {}
tools ToolsConfig	Tools configuration		Optional: {}
prompt_guard PromptGuardConfig	Prompt guard configuration		Optional: {}
classifier ClassifierConfig	Classifier configuration		Optional: {}
complexity_rules ComplexityRulesConfig array	Complexity rules for complexity-aware routing		Optional: {}
complexity_model ComplexityModelConfig	ComplexityModel says how the complexity signal produces its score. Absent, the signal scores locally against each rule's hard/easy candidates. With a backend, a remote model produces the score and the candidates are never read. Mirrors global.model_catalog.modules.complexity in the router config.		Optional: {}
external_models ExternalModelConfig array	ExternalModels declares the remote models that classifier backends (classifier.pii.backend.model, complexity_model.backend.model) and the prompt guard protocol refer to by name. Mirrors global.model_catalog.external[] in the router config field for field; the router's own validator decides whether a backend resolves against it.		Optional: {}
strategy string	Decision routing strategy ("priority" for priority-based matching)		Enum: [priority] Optional: {}
decisions DecisionConfig array	Routing decisions based on signals (domain, complexity, etc.)		Optional: {}
reasoning_effort string	ReasoningEffort is the default reasoning effort for model bindings that do not select a different effort. The selected model family validates the value because built-in and custom families may expose different ladders.		Optional: {}
api APIConfig	API configuration		Optional: {}
observability ObservabilityConfig	Observability configuration		Optional: {}
streamed_body StreamedBodyConfig	StreamedBody enables streamed request body handling. Mirrors global.router.streamed_body; the gateway must send bodies to ExtProc in STREAMED or FullDuplexStreamed mode for it to take effect.		Optional: {}

DecisionConfig

DecisionConfig defines a routing decision

Appears in:

• ConfigSpec

Field	Description	Default	Validation
name string	Name is the unique identifier for this decision
description string	Description provides information about what this decision handles		Optional: {}
priority integer	Priority is used for decision ordering - higher priority decisions are evaluated first		Optional: {}
rules RuleCombinationConfig	Rules defines the combination of conditions using AND/OR logic
modelRefs ModelRefConfig array	ModelRefs contains model references for this decision		Optional: {}
preferred_endpoints string array	PreferredEndpoints specifies which vLLM endpoints to prefer for this decision		Optional: {}
plugins RawExtension array	Plugins contains policy configurations applied after rule matching		Optional: {}
algorithm JSON	Algorithm configures base model selection for this decision. It is preserved as a router-owned object so supported algorithms can evolve without requiring the operator CRD to duplicate every nested field.		Type: object Optional: {}
reliability DecisionReliabilityConfig	Reliability overrides the timeouts and retries of the provider model that serves this decision's requests		Optional: {}
fallback DecisionFallbackConfig	Fallback overrides the cross-model fallback policy for this decision's requests, over the recipe's and the router's		Optional: {}

DecisionFallbackConfig

DecisionFallbackConfig is a decision's fallback block, with the router configuration's field names and meaning. A field left out keeps the recipe's or the router's value.

Appears in:

• DecisionConfig

Field	Description	Default	Validation
enabled boolean	Enabled turns cross-model fallback on or off for this decision		Optional: {}
max_attempts integer	MaxAttempts bounds the candidates tried, the first one included		Minimum: 0 Optional: {}
total_timeout string	TotalTimeout bounds the whole fallback chain		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
per_attempt_timeout string	PerAttemptTimeout bounds each candidate's attempt		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
retryable_status_codes integer array	RetryableStatusCodes are the statuses that move on to the next candidate		items:Maximum: 599 items:Minimum: 100 Optional: {}

DecisionModelBinding

DecisionModelBinding selects one canonical model deployment.

Appears in:

• ConfigSpec

Field	Description	Default	Validation
deployment string	Deployment is an exact key in model_deployments.		MinLength: 1

DecisionReliabilityConfig

DecisionReliabilityConfig is a decision's reliability block, with the router configuration's field names and meaning. Durations are Go durations such as "30s"; "0s" turns a timeout off.

Appears in:

• DecisionConfig

Field	Description	Default	Validation
total_timeout string	TotalTimeout bounds the whole call: every attempt and the response		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
per_try_timeout string	PerTryTimeout bounds each attempt until its response starts		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
idle_timeout string	IdleTimeout bounds the wait for more of a streamed response (standalone mode only)		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
first_byte_timeout string	FirstByteTimeout bounds the wait for the first response byte (standalone mode only)		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
retry_count integer	RetryCount is the number of retries after the first attempt		Maximum: 5 Minimum: 0 Optional: {}
retry_on string	RetryOn adds retry conditions, with Envoy's names (5xx, reset, ...)		Optional: {}
retriable_status_codes integer array	RetriableStatusCodes adds statuses retried under retriable-status-codes		items:Maximum: 599 items:Minimum: 100 Optional: {}
retry_back_off_base string	RetryBackOffBase is the base of the randomized exponential wait between retries (standalone mode only)		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
retry_back_off_max string	RetryBackOffMax caps that wait (standalone mode only)		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}
retry_after_max string	RetryAfterMax honors a response's Retry-After up to this bound (standalone mode only)		Pattern: ^([0-9]+(\.[0-9]+)?(ns|us|ms|s|m|h))+$ Optional: {}

EmbeddingEndpointConfig

EmbeddingEndpointConfig defines an external OpenAI-compatible embedding endpoint.

Appears in:

• EmbeddingModelsConfig

Field	Description	Default	Validation
base_url string	BaseURL is the base URL for the embedding endpoint, typically ending in /v1.		Optional: {}
model string	Model is the embedding model name sent to the external provider.		Optional: {}
api_key_env string	APIKeyEnv names the environment variable containing the provider API key.		Optional: {}
timeout_seconds integer	TimeoutSeconds is the request timeout for embedding calls.		Minimum: 0 Optional: {}
max_retries integer	MaxRetries is the maximum number of retry attempts for embedding calls.		Minimum: 0 Optional: {}
max_response_bytes integer	MaxResponseBytes caps the size of each embedding response body.		Minimum: 0 Optional: {}
dimensions integer	Dimensions requests a provider-side output dimension when supported.		Minimum: 1 Optional: {}

EmbeddingModelsConfig

EmbeddingModelsConfig defines configuration for embedding models

Appears in:

• ConfigSpec

Field	Description	Default	Validation
qwen3_model_path string	Path to Qwen3-Embedding-0.6B model directory Qwen3 provides 32K context and high quality embeddings (1024 dimensions)		Optional: {}
mmbert_model_path string	Path to mmBERT 2D Matryoshka embedding model directory Supports layer early exit (3/6/11/22) and dimension reduction (64-768)		Optional: {}
use_cpu boolean	Use CPU for inference (default: true)	true	Optional: {}
embedding_config HNSWEmbeddingConfig	Embedding configuration for embedding-based classification		Optional: {}
endpoint EmbeddingEndpointConfig	Endpoint configures an external embedding provider endpoint. The API key should be injected into the semantic router pod environment and referenced by APIKeyEnv rather than stored directly in the CR.		Optional: {}

ExporterConfig

ExporterConfig defines exporter configuration

Appears in:

• TracingConfig

Field	Description	Default	Validation
type string		otlp	Optional: {}
endpoint string		jaeger:4317	Optional: {}
insecure boolean		true	Optional: {}

ExternalModelConfig

ExternalModelConfig is one entry of global.model_catalog.external[]: a remote model a classifier backend or the prompt guard can name. Field names are the router's YAML keys so the generic typed conversion carries them unchanged.

Appears in:

• ConfigSpec

Field	Description	Default	Validation
name string	Name is the catalog name a backend block refers to in its model field.		MinLength: 1
model_role string	ModelRole is what the model is used for; classifier backends require "classification", the prompt guard protocol requires "guardrail".		MinLength: 1
llm_model_name string	ModelName is the model identifier the remote service expects, and the value a token_spans.v1 envelope's model member must equal.		MinLength: 1
llm_endpoint ExternalModelEndpoint	Endpoint is where the remote model is reached.
llm_timeout_seconds integer	TimeoutSeconds bounds one call when the backend block sets no deadline.		Minimum: 1 Optional: {}

ExternalModelEndpoint

ExternalModelEndpoint is the address of a remote classification model.

Appears in:

• ExternalModelConfig

Field	Description	Default	Validation
address string			MinLength: 1
port integer			Maximum: 65535 Minimum: 1
protocol string			Enum: [http https] Optional: {}

GatewayReference

GatewayReference references an existing Gateway

Appears in:

• GatewaySpec

Field	Description	Default	Validation
name string	Name of the Gateway		MinLength: 1
namespace string	Namespace of the Gateway		MinLength: 1

GatewaySpec

GatewaySpec defines Gateway API integration configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
existingRef GatewayReference	ExistingRef references an existing Gateway that calls the Router over ext_proc. Setting it selects extproc mode.		Optional: {}

HNSWCacheConfig

HNSWCacheConfig defines HNSW index configuration for hybrid/in-memory backends.

Appears in:

• SemanticCacheConfig

Field	Description	Default	Validation
use_hnsw boolean	UseHNSW enables HNSW indexing for faster similarity search	false	Optional: {}
hnsw_m integer	M is the number of bi-directional links per node	16	Minimum: 2 Optional: {}
hnsw_ef_construction integer	EfConstruction is the size of dynamic candidate list during construction	200	Minimum: 1 Optional: {}
max_memory_entries integer	MaxMemoryEntries limits in-memory entries for hybrid backend	1000	Minimum: 0 Optional: {}

HNSWEmbeddingConfig

HNSWEmbeddingConfig contains settings for embedding classification with HNSW indexing

Appears in:

• EmbeddingModelsConfig

Field	Description	Default	Validation
backend string	Backend selects the embedding provider backend: the built-in model runtime (the default) or an external OpenAI-compatible endpoint.		Enum: [model_runtime openai_compatible] Optional: {}
model_type string	ModelType specifies which embedding model to use Options: "qwen3" (1024-dim, 32K context), "gemma" (768-dim, 8K context), "mmbert" (64-768-dim, multilingual), "remote" (external provider)		Enum: [qwen3 gemma mmbert remote] Optional: {}
preload_embeddings boolean	PreloadEmbeddings enables precomputing candidate embeddings at startup	true	Optional: {}
target_dimension integer	TargetDimension is the embedding dimension to use (default: 768) For mmBERT, supported local dimensions are 64, 128, 256, 512, 768. External providers may use other positive dimensions such as 1024, 1536, or 3072.		Minimum: 1 Optional: {}
target_layer integer	TargetLayer controls mmBERT early exit and is used only when ModelType is "mmbert". Lower layers reduce encoder work but may reduce quality; layer 22 uses the full encoder depth. Evaluate the latency and quality trade-off on representative deployment data.		Enum: [3 6 11 22] Optional: {}
enable_soft_matching boolean	EnableSoftMatching allows below-threshold matches when no rule meets its threshold.	false	Optional: {}
min_score_threshold string	MinScoreThreshold for matching (0.0-1.0). Stored as string to avoid float precision issues.	0.5	Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}

ImageSpec

ImageSpec defines the container image configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
repository string	Repository is the container image repository	ghcr.io/vllm-project/semantic-router/vllm-sr	Optional: {}
tag string	Tag is the container image tag	latest	Optional: {}
pullPolicy PullPolicy	PullPolicy is the image pull policy	IfNotPresent	Enum: [Always Never IfNotPresent] Optional: {}
imageRegistry string	ImageRegistry is an optional registry prefix		Optional: {}

IngressHost

IngressHost defines an ingress host

Appears in:

• IngressSpec

Field	Description	Default	Validation
host string			Optional: {}
paths IngressPath array			Optional: {}

IngressPath

IngressPath defines an ingress path

Appears in:

• IngressHost

Field	Description	Default	Validation
path string			Optional: {}
pathType string			Optional: {}
servicePort integer			Optional: {}

IngressSpec

IngressSpec defines ingress configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
enabled boolean	Enabled indicates if ingress is enabled	false	Optional: {}
className string	ClassName is the ingress class name		Optional: {}
annotations object (keys:string, values:string)	Annotations for ingress		Optional: {}
hosts IngressHost array	Hosts configuration		Optional: {}
tls IngressTLS array	TLS configuration		Optional: {}

IngressTLS

IngressTLS defines ingress TLS configuration

Appears in:

• IngressSpec

Field	Description	Default	Validation
secretName string			Optional: {}
hosts string array			Optional: {}

LoRAAdapterSpec

LoRAAdapterSpec defines one LoRA adapter exposed by a VLLMEndpoint model.

Appears in:

• VLLMEndpointSpec

Field	Description	Default	Validation
name string	Name is the unique adapter identifier referenced by decision.modelRefs[].lora_name.		MaxLength: 100 MinLength: 1
description string	Description provides a short human-readable summary for UI and docs surfaces.		MaxLength: 500 Optional: {}

MetricsPortSpec

MetricsPortSpec extends PortSpec with enable flag

Appears in:

• ServiceSpec

Field	Description	Default	Validation
port integer	Port is the service port		Maximum: 65535 Minimum: 1 Optional: {}
targetPort integer	TargetPort is the container port		Maximum: 65535 Minimum: 1 Optional: {}
protocol Protocol	Protocol is the port protocol	TCP	Optional: {}
enabled boolean	Enabled indicates if metrics should be exposed	true	Optional: {}

MilvusCacheAuth

MilvusCacheAuth defines Milvus authentication.

Appears in:

• MilvusCacheConnection

Field	Description	Default	Validation
enabled boolean	Enabled controls whether to use authentication	false	Optional: {}
username string	Username for Milvus authentication		Optional: {}
password string	Password for Milvus authentication (plaintext - consider using PasswordSecretRef instead)		Optional: {}
password_secret_ref SecretKeySelector	PasswordSecretRef references a Secret containing the Milvus password Preferred over plaintext Password field for security		Optional: {}

MilvusCacheCollection

MilvusCacheCollection defines Milvus collection configuration.

Appears in:

• MilvusCacheConfig

Field	Description	Default	Validation
name string	Name of the Milvus collection	semantic_cache	Optional: {}
description string	Description of the collection	Semantic cache for LLM request-response pairs	Optional: {}
vector_field MilvusCacheVectorField	VectorField configuration for embeddings		Optional: {}
index MilvusCacheCollectionIndex	Index configuration for the collection		Optional: {}

MilvusCacheCollectionIndex

MilvusCacheCollectionIndex defines collection index settings.

Appears in:

• MilvusCacheCollection

Field	Description	Default	Validation
type string	Type of index algorithm	HNSW	Enum: [HNSW IVF_FLAT IVF_SQ8 IVF_PQ] Optional: {}
params MilvusCacheIndexParams	Params for the index		Optional: {}

MilvusCacheConfig

MilvusCacheConfig defines Milvus cache backend configuration. Configure these settings when using Milvus as the semantic cache backend.

Appears in:

• SemanticCacheConfig

Field	Description	Default	Validation
connection MilvusCacheConnection	Connection settings for Milvus server		Optional: {}
collection MilvusCacheCollection	Collection settings for Milvus		Optional: {}
search MilvusCacheSearch	Search settings for Milvus queries		Optional: {}
development MilvusCacheDevelopment	Development settings for Milvus cache		Optional: {}

MilvusCacheConnection

MilvusCacheConnection defines Milvus connection parameters.

Appears in:

• MilvusCacheConfig

Field	Description	Default	Validation
host string	Host is the Milvus server hostname or IP address		Optional: {}
port integer	Port is the Milvus server port	19530	Maximum: 65535 Minimum: 1 Optional: {}
database string	Database name in Milvus	semantic_router_cache	Optional: {}
timeout integer	Timeout for Milvus operations in seconds	30	Minimum: 0 Optional: {}
auth MilvusCacheAuth	Auth configuration for Milvus authentication		Optional: {}
tls MilvusCacheTLS	TLS configuration for secure Milvus connections		Optional: {}

MilvusCacheDevelopment

MilvusCacheDevelopment defines development-mode settings.

Appears in:

• MilvusCacheConfig

Field	Description	Default	Validation
drop_collection_on_startup boolean	DropCollectionOnStartup clears the collection when router starts (for testing)	false	Optional: {}
auto_create_collection boolean	AutoCreateCollection automatically creates the collection if it doesn't exist	true	Optional: {}

MilvusCacheIndexParams

MilvusCacheIndexParams defines index parameters.

Appears in:

• MilvusCacheCollectionIndex

Field	Description	Default	Validation
M integer	M is the number of bi-directional links for HNSW	16	Minimum: 2 Optional: {}
efConstruction integer	EfConstruction for HNSW index building	64	Minimum: 1 Optional: {}

MilvusCacheSearch

MilvusCacheSearch defines Milvus search parameters.

Appears in:

• MilvusCacheConfig

Field	Description	Default	Validation
params MilvusCacheSearchParams	Params for search operations		Optional: {}
topk integer	TopK is the number of results to return	10	Minimum: 1 Optional: {}
consistency_level string	ConsistencyLevel for search operations Options: "Strong", "Session", "Bounded", "Eventually"	Session	Enum: [Strong Session Bounded Eventually] Optional: {}

MilvusCacheSearchParams

MilvusCacheSearchParams defines search-time parameters.

Appears in:

• MilvusCacheSearch

Field	Description	Default	Validation
ef integer	Ef is the search-time HNSW parameter	64	Minimum: 1 Optional: {}

MilvusCacheTLS

MilvusCacheTLS defines TLS settings for Milvus connections.

Appears in:

• MilvusCacheConnection

Field	Description	Default	Validation
enabled boolean	Enabled controls whether to use TLS	false	Optional: {}
cert_file string	CertFile is the path to client certificate file		Optional: {}
key_file string	KeyFile is the path to client key file		Optional: {}
ca_file string	CAFile is the path to CA certificate file		Optional: {}

MilvusCacheVectorField

MilvusCacheVectorField defines vector field configuration.

Appears in:

• MilvusCacheCollection

Field	Description	Default	Validation
name string	Name of the vector field	embedding	Optional: {}
dimension integer	Dimension of the embedding vectors		Minimum: 1 Optional: {}
metric_type string	MetricType for vector similarity Options: "IP" (inner product), "L2", "COSINE"	IP	Enum: [IP L2 COSINE] Optional: {}

ModelReasoningSpec

ModelReasoningSpec selects a catalog reasoning family or defines the request projection for a custom self-hosted model. Family and inline fields are mutually exclusive and are validated by the Router's canonical compiler.

Appears in:

• VLLMEndpointSpec

Field	Description	Default	Validation
family string			Optional: {}
type string			Enum: [chat_template_kwargs reasoning_effort reasoning_mode top_level_reasoning_effort] Optional: {}
parameter string			Optional: {}
activationParameter string			Optional: {}
effortFlags object (keys:string, values:string)	EffortFlags maps a logical effort to a boolean chat-template parameter.		Optional: {}
levels string array			Optional: {}
default string			Optional: {}
modes string array			items:Enum: [enabled disabled adaptive] Optional: {}
defaultMode string			Enum: [enabled disabled adaptive] Optional: {}
disabled string			Optional: {}

ModelRefConfig

ModelRefConfig defines a model reference for routing

Appears in:

• DecisionConfig

Field	Description	Default	Validation
model string	Model name to route to
lora_name string	LoRAName is the optional LoRA adapter name		Optional: {}
use_reasoning boolean	UseReasoning enables reasoning mode for this model		Optional: {}
reasoning_mode string	ReasoningMode selects the model's reasoning activation mode when the family supports more than a boolean switch.		Enum: [enabled disabled adaptive] Optional: {}
reasoning_effort string	ReasoningEffort selects one of the model family's declared effort levels.		Optional: {}

ObservabilityConfig

ObservabilityConfig defines observability configuration

Appears in:

• ConfigSpec

Field	Description	Default	Validation
tracing TracingConfig			Optional: {}

OpenShiftFeaturesStatus

OpenShiftFeaturesStatus tracks OpenShift-specific feature status

Appears in:

• SemanticRouterStatus

Field	Description	Default	Validation
routesEnabled boolean	RoutesEnabled indicates if OpenShift Routes are enabled
routeHostname string	RouteHostname is the hostname of the created Route		Optional: {}

OpenShiftSpec

OpenShiftSpec defines OpenShift-specific configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
routes RouteConfig	Routes configuration for OpenShift Routes		Optional: {}

PIIModelConfig

PIIModelConfig defines PII model configuration.

The contract rule sits on the consumer, as on ComplexityModelConfig: the shared backend block lists every contract any consumer reads, and each consumer narrows it to what it can parse, so a mismatch is refused at admission instead of by the router at load.

Appears in:

• ClassifierConfig

Field	Description	Default	Validation
max_sequence_length integer	MaxSequenceLength is the total tokenized input budget, including special tokens. Omission or zero preserves the 512-token legacy limit.		Minimum: 0 Optional: {}
window PromptGuardWindowConfig	Window scans original content tokens with explicit overlap. Omission or null leaves window selection unchanged; no CRD defaults are injected.		Optional: {}
model_id string			Optional: {}
threshold string	Detection threshold (0.0-1.0). Stored as string to avoid float precision issues.		Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
use_cpu boolean			Optional: {}
pii_mapping_path string			Optional: {}
backend RemoteClassifierBackendConfig	Backend names a remote token classifier speaking token_spans.v1. Its absence keeps local PII inference. The local selectors this replaces are model_id and use_cpu above. Explicit token windows are only supported by the local model.		Optional: {}
on_error string	OnError selects what a PII backend failure, or a provider-declared truncation, does to the rule that consumed it: allow (default) treats the content as not matching, block matches it as classification_error.		Enum: [allow block] Optional: {}

PersistenceSpec

PersistenceSpec defines persistence configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
enabled boolean	Enabled indicates if persistence is enabled	true	Optional: {}
storageClassName string	StorageClassName is the storage class name	standard	Optional: {}
accessMode PersistentVolumeAccessMode	AccessMode is the access mode	ReadWriteOnce	Optional: {}
size string	Size is the storage size	10Gi	Optional: {}
existingClaim string	ExistingClaim is an existing PVC to use		Optional: {}
annotations object (keys:string, values:string)	Annotations for the PVC		Optional: {}

PortSpec

PortSpec defines a service port configuration

Appears in:

• MetricsPortSpec
• ServiceSpec

Field	Description	Default	Validation
port integer	Port is the service port		Maximum: 65535 Minimum: 1 Optional: {}
targetPort integer	TargetPort is the container port		Maximum: 65535 Minimum: 1 Optional: {}
protocol Protocol	Protocol is the port protocol	TCP	Optional: {}

ProbeSpec

ProbeSpec defines probe configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
enabled boolean	Enabled indicates if the probe is enabled	true	Optional: {}
initialDelaySeconds integer	InitialDelaySeconds before probe starts		Optional: {}
periodSeconds integer	PeriodSeconds between probes		Optional: {}
timeoutSeconds integer	TimeoutSeconds for probe		Optional: {}
failureThreshold integer	FailureThreshold for probe		Optional: {}

PromptGuardConfig

PromptGuardConfig defines prompt guard configuration.

Appears in:

• ConfigSpec

Field	Description	Default	Validation
backend RemoteClassifierBackendConfig	Backend selects a named external classifier and its typed result contract.		Optional: {}
max_sequence_length integer	MaxSequenceLength limits the total tokenized input, including special tokens. Omission or zero retains the 512-token budget. The model loader validates the requested budget against the loaded model's capacity.		Minimum: 0 Optional: {}
window PromptGuardWindowConfig	Window enables explicit scanning of all input tokens. Omission or null keeps whole-input inference. Only the local model supports it.		Optional: {}
enabled boolean		true	Optional: {}
model_id string	ModelID binds the guard to one model; empty runs the decision model.		Optional: {}
threshold string	Jailbreak detection threshold (0.0-1.0). Stored as string to avoid float precision issues; empty takes the threshold calibrated for the model.		Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
use_cpu boolean		true	Optional: {}
jailbreak_mapping_path string			Optional: {}
positive_labels string array	PositiveLabels lists the jailbreak_mapping labels that count as unsafe, for a custom backend whose positive class isn't named "jailbreak" (e.g. "INJECTION", "malicious"). Defaults to ["jailbreak"] when unset.		Optional: {}
on_error string	OnError selects what a prompt-guard classifier failure does to the rule that failed to evaluate. "allow" (the default) tolerates the failure and treats the content as not matching; "block" treats it as a positive detection, because an inference failure means the content could not be verified safe. Without this field on the CRD the setting is pruned by the API server and an operator-managed deployment silently fails open.		Enum: [allow block] Optional: {}

PromptGuardWindowConfig

PromptGuardWindowConfig scans original content tokens with overlap. The native tokenizer also checks that special tokens leave enough content room.

Appears in:

• PIIModelConfig
• PromptGuardConfig

Field	Description	Default	Validation
size integer	Size is the inference window budget, including special tokens.		Minimum: 1
overlap integer	Overlap counts content tokens shared by consecutive windows.		Minimum: 0 Optional: {}

PrototypeScoringConfig

PrototypeScoringConfig overrides prototype-bank construction and scoring for one embedding-backed signal rule. Router-owned defaults apply only after translation; the operator preserves an absent override and an empty object.

Appears in:

• ComplexityRulesConfig

Field	Description	Default	Validation
enabled boolean	Enabled controls prototype clustering. False retains every candidate.		Optional: {}
cluster_similarity_threshold string	ClusterSimilarityThreshold is the clustering similarity threshold. Stored as a numeric string, like other fractional operator config fields.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}
max_prototypes integer	MaxPrototypes caps the number of cluster representatives.		Optional: {}
best_weight string	BestWeight weights the best prototype against the top-M mean.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}
top_m integer	TopM is the number of highest-scoring prototypes included in the mean.		Optional: {}
margin_threshold string	MarginThreshold is the minimum winner-versus-runner-up score margin.		Pattern: ^-?[0-9]+(\.[0-9]+)?$ Optional: {}

QdrantCacheConfig

QdrantCacheConfig defines Qdrant cache backend configuration. Configure these settings when using Qdrant as the semantic cache backend.

Appears in:

• SemanticCacheConfig

Field	Description	Default	Validation
host string	Host is the Qdrant server hostname or IP address		Optional: {}
port integer	Port is the Qdrant gRPC port	6334	Maximum: 65535 Minimum: 1 Optional: {}
api_key string	APIKey for Qdrant authentication		Optional: {}
use_tls boolean	UseTLS enables TLS for the Qdrant connection	false	Optional: {}
collection_name string	CollectionName is the Qdrant collection to use for semantic cache	semantic_cache	Optional: {}
connect_timeout integer	ConnectTimeout is the timeout in seconds for Qdrant connection	10	Optional: {}

RedisCacheConfig

RedisCacheConfig defines Redis cache backend configuration. Configure these settings when using Redis as the semantic cache backend.

Appears in:

• SemanticCacheConfig

Field	Description	Default	Validation
connection RedisCacheConnection	Connection settings for Redis server		Optional: {}
index RedisCacheIndex	Index settings for Redis vector search		Optional: {}
search RedisCacheSearch	Search settings for Redis queries		Optional: {}
development RedisCacheDevelopment	Development settings for Redis cache		Optional: {}

RedisCacheConnection

RedisCacheConnection defines Redis connection parameters.

Appears in:

• RedisCacheConfig

Field	Description	Default	Validation
host string	Host is the Redis server hostname or IP address Example: "redis.default.svc.cluster.local"		Optional: {}
port integer	Port is the Redis server port	6379	Maximum: 65535 Minimum: 1 Optional: {}
database integer	Database is the Redis database number to use	0	Minimum: 0 Optional: {}
password string	Password for Redis authentication (plaintext - consider using PasswordSecretRef instead)		Optional: {}
password_secret_ref SecretKeySelector	PasswordSecretRef references a Secret containing the Redis password Preferred over plaintext Password field for security		Optional: {}
timeout integer	Timeout for Redis operations in seconds	30	Minimum: 0 Optional: {}
tls RedisCacheTLS	TLS configuration for secure Redis connections		Optional: {}

RedisCacheDevelopment

RedisCacheDevelopment defines development-mode settings.

Appears in:

• RedisCacheConfig

Field	Description	Default	Validation
drop_index_on_startup boolean	DropIndexOnStartup clears the index when router starts (for testing)	false	Optional: {}
auto_create_index boolean	AutoCreateIndex automatically creates the index if it doesn't exist	true	Optional: {}

RedisCacheIndex

RedisCacheIndex defines Redis vector index configuration.

Appears in:

• RedisCacheConfig

Field	Description	Default	Validation
name string	Name of the Redis index	semantic_cache_idx	Optional: {}
prefix string	Prefix for Redis keys	doc:	Optional: {}
vector_field RedisCacheVectorField	VectorField configuration for embeddings		Optional: {}
index_type string	IndexType specifies the index algorithm Options: "HNSW" (recommended), "FLAT"	HNSW	Enum: [HNSW FLAT] Optional: {}
params RedisCacheIndexParams	Params for HNSW index		Optional: {}

RedisCacheIndexParams

RedisCacheIndexParams defines HNSW index parameters.

Appears in:

• RedisCacheIndex

Field	Description	Default	Validation
M integer	M is the number of bi-directional links per node Higher values = better recall, more memory	16	Minimum: 2 Optional: {}
efConstruction integer	EfConstruction is the size of dynamic candidate list during construction Higher values = better quality, slower indexing	64	Minimum: 1 Optional: {}

RedisCacheSearch

RedisCacheSearch defines Redis search parameters.

Appears in:

• RedisCacheConfig

Field	Description	Default	Validation
topk integer	TopK is the number of results to return from vector search	1	Minimum: 1 Optional: {}

RedisCacheTLS

RedisCacheTLS defines TLS settings for Redis connections.

Appears in:

• RedisCacheConnection

Field	Description	Default	Validation
enabled boolean	Enabled controls whether to use TLS for Redis connection	false	Optional: {}
cert_file string	CertFile is the path to client certificate file		Optional: {}
key_file string	KeyFile is the path to client key file		Optional: {}
ca_file string	CAFile is the path to CA certificate file		Optional: {}

RedisCacheVectorField

RedisCacheVectorField defines vector field configuration.

Appears in:

• RedisCacheIndex

Field	Description	Default	Validation
name string	Name of the vector field	embedding	Optional: {}
dimension integer	Dimension of the embedding vectors For BERT: 384, for Qwen3: 1024, for Gemma: 768		Minimum: 1 Optional: {}
metric_type string	MetricType for vector similarity Options: "COSINE", "IP" (inner product), "L2" (Euclidean)	COSINE	Enum: [COSINE IP L2] Optional: {}

RemoteClassifierBackendConfig

RemoteClassifierBackendConfig is the shared remote-classifier block. How the remote is called (protocol), what shape it answers with (contract), which catalog entry it is (model) and how long to wait (deadline) are independent axes rather than one enumeration. It mirrors the router's backend block field for field so the operator passes it through unchanged.

Appears in:

• ComplexityModelConfig
• PIIModelConfig
• PromptGuardConfig

Field	Description	Default	Validation
protocol string	Protocol is how the remote is called.		Enum: [http_classify http_chat]
contract string	Contract is the response shape the signal reads. Complexity reads two - score.v1, one regression number interpreted through each rule's boundaries, and label_distribution.v1, hard/easy/medium probabilities - so the router requires it there rather than guessing per request. PII reads token_spans.v1, entity spans with code-point offsets. Prompt guard http_chat reads label_decision.v1, a verdict without invented probability.		Enum: [score.v1 label_distribution.v1 token_spans.v1 label_decision.v1] Optional: {}
model string	Model is the name of an entry in the external model catalog.		MinLength: 1
deadline_ms integer	DeadlineMs bounds one remote call. Defaults to the router's value.		Minimum: 1 Optional: {}

ResourceConfig

ResourceConfig defines resource configuration for tracing

Appears in:

• TracingConfig

Field	Description	Default	Validation
service_name string		vllm-semantic-router	Optional: {}
service_version string		v0.1.0	Optional: {}
deployment_environment string		development	Optional: {}

RouteConfig

RouteConfig defines OpenShift Route configuration

Appears in:

• OpenShiftSpec

Field	Description	Default	Validation
enabled boolean	Enabled specifies whether to create an OpenShift Route	false	Optional: {}
hostname string	Hostname for the Route (optional - OpenShift generates if empty)		Optional: {}
tls RouteTLSConfig	TLS configuration for the Route		Optional: {}

RouteTLSConfig

RouteTLSConfig defines TLS configuration for OpenShift Routes

Appears in:

• RouteConfig

Field	Description	Default	Validation
termination string	Termination type (edge, passthrough, reencrypt)	edge	Enum: [edge passthrough reencrypt] Optional: {}
insecureEdgeTerminationPolicy string	InsecureEdgeTerminationPolicy for HTTP traffic	Redirect	Enum: [Allow Redirect None] Optional: {}

RuleCombinationConfig

RuleCombinationConfig defines how to combine multiple rule conditions

Appears in:

• DecisionConfig

Field	Description	Default	Validation
operator string	Operator specifies how to combine conditions: "AND", "OR", or "NOT". NOT is strictly unary: it takes exactly one child condition and negates its result. Compose NOR/NAND by nesting NOT around OR/AND.		Enum: [AND OR NOT]
on_unknown string	OnUnknown resolves a terminal unknown result after the rule tree is evaluated.		Enum: [no_match match fail_request] Optional: {}
conditions RuleConditionConfig array	Conditions is the list of rule references to evaluate

RuleComposition

RuleComposition defines how to compose/filter rules based on other signals

Appears in:

• ComplexityRulesConfig

Field	Description	Default	Validation
operator string	Operator for combining conditions (AND, OR, NOT). NOT is strictly unary and negates its single child.		Enum: [AND OR NOT]
conditions CompositionCondition array	List of conditions that must be met

RuleConditionConfig

RuleConditionConfig references a specific rule by type and name

Appears in:

• RuleCombinationConfig

Field	Description	Default	Validation
type string	Type specifies the signal or projection type referenced by this condition.		Enum: [keyword embedding domain fact_check user_feedback reask preference language context structure complexity modality authz jailbreak pii kb conversation event projection]
name string	Name is the name of the rule to reference

SamplingConfig

SamplingConfig defines sampling configuration

Appears in:

• TracingConfig

Field	Description	Default	Validation
type string		always_on	Optional: {}
rate string	Sampling rate (0.0-1.0). Stored as string to avoid float precision issues.	1.0	Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}

SemanticCacheConfig

SemanticCacheConfig defines semantic cache configuration

Appears in:

• ConfigSpec

Field	Description	Default	Validation
enabled boolean	Enabled controls whether semantic caching is active	true	Optional: {}
backend_type string	BackendType specifies the cache backend to use Options: "memory" (default), "redis", "valkey", "milvus", "qdrant", "hybrid"	memory	Enum: [memory redis valkey milvus qdrant hybrid] Optional: {}
similarity_threshold string	Similarity threshold for cache hits (0.0-1.0). Stored as string to avoid float precision issues.	0.8	Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
max_entries integer	MaxEntries is the maximum number of cache entries (for memory/hybrid backends)	1000	Optional: {}
ttl_seconds integer	TTLSeconds is the time-to-live for cache entries in seconds	3600	Optional: {}
eviction_policy string	EvictionPolicy for in-memory cache ("fifo", "lru", "lfu")	fifo	Enum: [fifo lru lfu] Optional: {}
redis RedisCacheConfig	Redis configuration (required when backend_type is "redis")		Optional: {}
valkey ValkeyCacheConfig	Valkey configuration (required when backend_type is "valkey")		Optional: {}
milvus MilvusCacheConfig	Milvus configuration (required when backend_type is "milvus")		Optional: {}
qdrant QdrantCacheConfig	Qdrant configuration (required when backend_type is "qdrant")		Optional: {}
embedding_model string	EmbeddingModel specifies which embedding model to use for semantic similarity Options: "mmbert" (default), "bert", "qwen3", "gemma"	mmbert	Enum: [bert qwen3 gemma mmbert] Optional: {}
hnsw HNSWCacheConfig	HNSW configuration for hybrid/in-memory backends		Optional: {}

SemanticRouter

SemanticRouter is the Schema for the semanticrouters API

Appears in:

• SemanticRouterList

Field	Description	Default	Validation
apiVersion string	vllm.ai/v1alpha1
kind string	SemanticRouter
metadata ObjectMeta	Refer to Kubernetes API documentation for fields of metadata.
spec SemanticRouterSpec
status SemanticRouterStatus

SemanticRouterList

SemanticRouterList contains a list of SemanticRouter

Field	Description	Default	Validation
apiVersion string	vllm.ai/v1alpha1
kind string	SemanticRouterList
metadata ListMeta	Refer to Kubernetes API documentation for fields of metadata.
items SemanticRouter array

SemanticRouterSpec

SemanticRouterSpec defines the desired state of SemanticRouter

Appears in:

• SemanticRouter

Field	Description	Default	Validation
image ImageSpec	Image configuration		Optional: {}
replicas integer	Number of replicas	1	Minimum: 0 Optional: {}
imagePullSecrets LocalObjectReference array	ImagePullSecrets for private registries		Optional: {}
serviceAccount ServiceAccountSpec	ServiceAccount configuration		Optional: {}
service ServiceSpec	Service configuration		Optional: {}
resources ResourceRequirements	Resource requirements		Optional: {}
persistence PersistenceSpec	Persistence configuration		Optional: {}
config ConfigSpec	Configuration overrides merged into the canonical v0.3 config.yaml. Router-wide runtime overrides land under config.global.router/services/stores/ integrations/model_catalog, with model-backed modules nested under config.global.model_catalog.modules. Provider defaults land under config.providers.defaults.		Optional: {}
toolsDb ToolEntry array	Tools database configuration		Optional: {}
vllmEndpoints VLLMEndpointSpec array	VLLMEndpoints is a Kubernetes-native backend discovery adapter. It generates canonical config.providers.models[].backend_refs and config.routing.modelCards entries.		Optional: {}
autoscaling AutoscalingSpec	Autoscaling configuration		Optional: {}
startupProbe ProbeSpec	Probes configuration		Optional: {}
livenessProbe ProbeSpec			Optional: {}
readinessProbe ProbeSpec			Optional: {}
securityContext SecurityContext	Security context		Optional: {}
podSecurityContext PodSecurityContext	Pod security context		Optional: {}
podAnnotations object (keys:string, values:string)	Pod annotations		Optional: {}
nodeSelector object (keys:string, values:string)	Node selector		Optional: {}
tolerations Toleration array	Tolerations		Optional: {}
affinity Affinity	Affinity		Optional: {}
env EnvVar array	Environment variables		Optional: {}
args string array	Router arguments, after the gateway mode flags the Operator passes (-gateway=standalone -listener-address=0.0.0.0, or -gateway=extproc). The gateway mode follows spec.gateway, so args may not set those flags.		MaxItems: 64 items:MaxLength: 4096 Optional: {}
gateway GatewaySpec	Gateway selects what serves client traffic. Omitted, the Router runs standalone: it serves the OpenAI-compatible API on port 8801 itself, with no Envoy, and the Service exposes that port. With existingRef, the Router serves ext_proc gRPC on port 50051 for that Gateway, whose ext_proc policy and routes you manage.		Optional: {}
openshift OpenShiftSpec	OpenShift-specific features		Optional: {}
ingress IngressSpec	Ingress configuration		Optional: {}

SemanticRouterStatus

SemanticRouterStatus defines the observed state of SemanticRouter

Appears in:

• SemanticRouter

Field	Description	Default	Validation
conditions Condition array	Conditions represent the latest available observations of the SemanticRouter's state		Optional: {}
observedGeneration integer	ObservedGeneration reflects the generation of the most recently observed SemanticRouter		Optional: {}
replicas integer	Replicas is the current number of replicas		Optional: {}
readyReplicas integer	ReadyReplicas is the number of ready replicas		Optional: {}
phase string	Phase represents the current phase of the SemanticRouter		Optional: {}
gatewayMode string	GatewayMode is what serves client traffic: standalone (the Router's own listener on port 8801) or gateway-integration (the Gateway in spec.gateway, with the Router serving ext_proc)		Optional: {}
openshiftFeatures OpenShiftFeaturesStatus	OpenShiftFeatures tracks OpenShift-specific feature status		Optional: {}

ServiceAccountSpec

ServiceAccountSpec defines service account configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
create boolean	Create specifies whether to create a service account	true	Optional: {}
name string	Name of the service account to use		Optional: {}
annotations object (keys:string, values:string)	Annotations for the service account		Optional: {}

ServiceBackend

ServiceBackend defines a direct Kubernetes service backend

Appears in:

• VLLMBackend

Field	Description	Default	Validation
name string	Service name		MinLength: 1
namespace string	Service namespace (defaults to same namespace)		Optional: {}
port integer	Service port		Minimum: 1

ServiceSpec

ServiceSpec defines the service configuration

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
type ServiceType	Type is the service type	ClusterIP	Enum: [ClusterIP NodePort LoadBalancer] Optional: {}
grpc PortSpec	GRPC port configuration		Optional: {}
api PortSpec	API port configuration		Optional: {}
metrics MetricsPortSpec	Metrics port configuration		Optional: {}

StreamedBodyConfig

StreamedBodyConfig defines streamed request body handling.

Appears in:

• ConfigSpec

Field	Description	Default	Validation
enabled boolean	Enabled accumulates request body chunks before routing at end-of-stream.		Optional: {}
max_bytes integer	MaxBytes caps the accumulated body size. A larger body is rejected and the ExtProc stream ends; the downstream response follows the gateway's ExtProc failure policy. Zero disables the limit.		Minimum: 0 Optional: {}
timeout_sec integer	TimeoutSec caps how long body accumulation may take. A slower body is rejected and the ExtProc stream ends; the downstream response follows the gateway's ExtProc failure policy. Zero disables the limit.		Minimum: 0 Optional: {}

Tool

Tool defines a tool function

Appears in:

• ToolEntry

Field	Description	Default	Validation
type string			Enum: [function] Optional: {}
function ToolFunction			Optional: {}

ToolEntry

ToolEntry defines a tool entry in the tools database

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
tool Tool			Optional: {}
description string			Optional: {}
category string			Optional: {}
tags string array			Optional: {}

ToolFunction

ToolFunction defines a tool function details

Appears in:

• Tool

Field	Description	Default	Validation
name string			Optional: {}
description string			Optional: {}
parameters ToolParameters			Optional: {}

ToolParameters

ToolParameters defines tool function parameters

Appears in:

• ToolFunction

Field	Description	Default	Validation
type string			Optional: {}
properties JSON			Type: object Optional: {}
required string array			Optional: {}

ToolsConfig

ToolsConfig defines tools configuration

Appears in:

• ConfigSpec

Field	Description	Default	Validation
enabled boolean		true	Optional: {}
top_k integer		3	Optional: {}
similarity_threshold string	Similarity threshold for tool selection (0.0-1.0). Stored as string to avoid float precision issues.	0.2	Pattern: ^0(\.[0-9]+)?$|^1(\.0+)?$ Optional: {}
tools_db_path string		config/tools_db.json	Optional: {}
fallback_to_empty boolean		true	Optional: {}

TracingConfig

TracingConfig defines tracing configuration

Appears in:

• ObservabilityConfig

Field	Description	Default	Validation
enabled boolean		false	Optional: {}
provider string		opentelemetry	Optional: {}
exporter ExporterConfig			Optional: {}
sampling SamplingConfig			Optional: {}
resource ResourceConfig			Optional: {}

VLLMBackend

VLLMBackend specifies how to reach the vLLM service

Appears in:

• VLLMEndpointSpec

Field	Description	Default	Validation
type string	Type of backend: kserve, llamastack, or service		Enum: [kserve llamastack service]
inferenceServiceName string	For type=kserve: InferenceService name for auto-discovery		Optional: {}
discoveryLabels object (keys:string, values:string)	For type=llamastack: Labels to match services		Optional: {}
service ServiceBackend	For type=service: Direct service configuration		Optional: {}

VLLMEndpointSpec

VLLMEndpointSpec defines a vLLM model backend endpoint

Appears in:

• SemanticRouterSpec

Field	Description	Default	Validation
name string	Name of the backend ref generated under config.providers.models[].backend_refs		MinLength: 1
model string	Model name as reported by vLLM (e.g., "Model-A", "llama3-8b")		MinLength: 1
catalog string	Catalog optionally selects a repository built-in Model Card. Model remains the request-facing alias.		Optional: {}
reasoning ModelReasoningSpec	Reasoning optionally selects a built-in family or defines inline wire behavior for this self-hosted model. Catalog-backed models normally omit it.		Optional: {}
loras LoRAAdapterSpec array	LoRAs declares the LoRA adapters exposed for this logical model in routing.modelCards.		MaxItems: 50 Optional: {}
backend VLLMBackend	Backend configuration
weight integer	Weight for load balancing (default: 1)	1	Optional: {}

ValkeyCacheConfig

ValkeyCacheConfig defines Valkey cache backend configuration. Configure these settings when using Valkey as the semantic cache backend.

Appears in:

• SemanticCacheConfig

Field	Description	Default	Validation
connection ValkeyCacheConnection	Connection settings for Valkey server		Optional: {}
index ValkeyCacheIndex	Index settings for Valkey vector search		Optional: {}
search ValkeyCacheSearch	Search settings for Valkey queries		Optional: {}
development ValkeyCacheDevelopment	Development settings for Valkey cache		Optional: {}

ValkeyCacheConnection

ValkeyCacheConnection defines Valkey connection parameters.

Appears in:

• ValkeyCacheConfig

Field	Description	Default	Validation
host string	Host is the Valkey server hostname or IP address Example: "valkey.default.svc.cluster.local"		Optional: {}
port integer	Port is the Valkey server port	6379	Maximum: 65535 Minimum: 1 Optional: {}
database integer	Database is the Valkey database number to use	0	Minimum: 0 Optional: {}
password string	Password for Valkey authentication (plaintext - consider using PasswordSecretRef instead)		Optional: {}
password_secret_ref SecretKeySelector	PasswordSecretRef references a Secret containing the Valkey password Preferred over plaintext Password field for security		Optional: {}
timeout integer	Timeout for Valkey operations in seconds	30	Minimum: 0 Optional: {}
tls ValkeyCacheTLS	TLS configuration for secure Valkey connections		Optional: {}

ValkeyCacheDevelopment

ValkeyCacheDevelopment defines development-mode settings.

Appears in:

• ValkeyCacheConfig

Field	Description	Default	Validation
drop_index_on_startup boolean	DropIndexOnStartup clears the index when router starts (for testing)	false	Optional: {}
auto_create_index boolean	AutoCreateIndex automatically creates the index if it doesn't exist	true	Optional: {}

ValkeyCacheIndex

ValkeyCacheIndex defines Valkey vector index configuration.

Appears in:

• ValkeyCacheConfig

Field	Description	Default	Validation
name string	Name of the Valkey index	semantic_cache_idx	Optional: {}
prefix string	Prefix for Valkey keys	doc:	Optional: {}
vector_field ValkeyCacheVectorField	VectorField configuration for embeddings		Optional: {}
index_type string	IndexType specifies the index algorithm Options: "HNSW" (recommended), "FLAT"	HNSW	Enum: [HNSW FLAT] Optional: {}
params ValkeyCacheIndexParams	Params for HNSW index		Optional: {}

ValkeyCacheIndexParams

ValkeyCacheIndexParams defines HNSW index parameters.

Appears in:

• ValkeyCacheIndex

Field	Description	Default	Validation
M integer	M is the number of bi-directional links per node Higher values = better recall, more memory	16	Minimum: 2 Optional: {}
efConstruction integer	EfConstruction is the size of dynamic candidate list during construction Higher values = better quality, slower indexing	64	Minimum: 1 Optional: {}

ValkeyCacheSearch

ValkeyCacheSearch defines Valkey search parameters.

Appears in:

• ValkeyCacheConfig

Field	Description	Default	Validation
topk integer	TopK is the number of results to return from vector search	1	Minimum: 1 Optional: {}

ValkeyCacheTLS

ValkeyCacheTLS defines TLS settings for Valkey connections.

Appears in:

• ValkeyCacheConnection

Field	Description	Default	Validation
enabled boolean	Enabled controls whether to use TLS for Valkey connection	false	Optional: {}
cert_file string	CertFile is the path to client certificate file		Optional: {}
key_file string	KeyFile is the path to client key file		Optional: {}
ca_file string	CAFile is the path to CA certificate file		Optional: {}

ValkeyCacheVectorField

ValkeyCacheVectorField defines vector field configuration.

Appears in:

• ValkeyCacheIndex

Field	Description	Default	Validation
name string	Name of the vector field	embedding	Optional: {}
dimension integer	Dimension of the embedding vectors For BERT: 384, for Qwen3: 1024, for Gemma: 768		Minimum: 1 Optional: {}
metric_type string	MetricType for vector similarity Options: "COSINE", "IP" (inner product), "L2" (Euclidean)	COSINE	Enum: [COSINE IP L2] Optional: {}