values.yaml

# =============================================================================
# nemo-retriever Helm chart — default values
# =============================================================================
#
# This chart deploys the "service" run mode of nemo-retriever and, optionally,
# the supporting NIM microservices via the NVIDIA NIM Operator (NIMCache /
# NIMService custom resources from the apps.nvidia.com/v1alpha1 API group).
# The NIM templates are gated on the operator CRDs being installed in the
# cluster, so the chart degrades gracefully when the operator is absent.
# See README.md for a full walkthrough of every knob below.
# =============================================================================

# -----------------------------------------------------------------------------
# Naming overrides
# -----------------------------------------------------------------------------
# nameOverride replaces the chart name (`nemo-retriever`) in resource names.
# fullnameOverride entirely replaces the generated `<release>-<chart>` prefix.
nameOverride: ""
fullnameOverride: ""

# -----------------------------------------------------------------------------
# NGC pull + API secrets (consumed by the service Deployment AND every
# operator-managed NIMService).
# -----------------------------------------------------------------------------
# The NIM Operator references its image pull secret and NGC auth secret by
# **name only** — those names are hard-coded into each NIMCache/NIMService
# template (`ngc-secret` and `ngc-api` by default, matching upstream NIM
# Operator examples). Set `create: true` to have this chart provision the
# secrets, or `create: false` + an externally managed Secret with the same
# name.
#
# Additional pre-existing pull secrets can be listed under `imagePullSecrets`
# at the bottom of this section.
ngcImagePullSecret:
  create: false
  # Name used by every Pod's `imagePullSecrets` and by each NIMService's
  # `image.pullSecrets`. Keep at "ngc-secret" unless you have a strong
  # reason to rename — every per-model template references this name by
  # default.
  name: "ngc-secret"
  registry: "nvcr.io"
  username: "$oauthtoken"
  # NGC API key. Required when `create: true`. May be left blank if you
  # supply `dockerconfigjson` directly.
  password: ""
  # Optional pre-assembled docker-config JSON (base64-encoded). When set,
  # overrides registry/username/password.
  dockerconfigjson: ""

ngcApiSecret:
  create: false
  # Name used as the NIMCache/NIMService `authSecret` and as the source for
  # the service container's NGC_API_KEY / NVIDIA_API_KEY env vars.
  name: "ngc-api"
  # NGC API key. Both `NGC_API_KEY` and `NGC_CLI_API_KEY` keys are
  # populated with this value so NIM containers and ad-hoc CLI users
  # both work.
  password: ""

# Extra pre-existing image pull secrets to append to every Pod in addition
# to the chart-managed `ngcImagePullSecret` above.
imagePullSecrets: []
#  - name: my-private-registry

# =============================================================================
# Service (the FastAPI nemo-retriever ingest server)
# =============================================================================
service:
  image:
    # Default points at the GA image published to NGC. Override
    # `repository` / `tag` to pin a different build, e.g. one produced by:
    #   docker build -f Dockerfile --target service \
    #       -t <your-registry>/nemo-retriever-service:<tag> .
    repository: nvcr.io/nvidia/nemo-microservices/nrl-service
    tag: "26.5.0"
    pullPolicy: IfNotPresent

  # Optional image for in-pod Hugging Face models (service-gpu Docker target).
  # Used when serviceConfig.localModels.enabled=true. When repository is
  # empty, the chart falls back to service.image above.
  #   docker build -f Dockerfile --target service-gpu \
  #       -t <your-registry>/nemo-retriever-service-gpu:<tag> .
  gpuImage:
    repository: ""
    tag: ""
    pullPolicy: ""

  # Number of standalone pod replicas. Keep at 1 because job and scheduler
  # state are process-local.
  replicas: 1

  # Pod-level scheduling controls.
  podAnnotations: {}
  podLabels: {}
  nodeSelector: {}
  tolerations: []
  # Prefer co-locating the retriever pod on the same node as NIM pods to
  # minimize network latency for inference calls. This is a soft preference
  # (preferredDuringScheduling) so the pod still schedules if no NIM node
  # has capacity.
  affinity:
    podAffinity:
      preferredDuringSchedulingIgnoredDuringExecution:
        - weight: 100
          podAffinityTerm:
            labelSelector:
              matchExpressions:
                - key: app.kubernetes.io/component
                  operator: In
                  values:
                    - nim
            topologyKey: kubernetes.io/hostname
  topologySpreadConstraints: []
  priorityClassName: ""
  terminationGracePeriodSeconds: 60

  podSecurityContext:
    runAsNonRoot: true
    runAsUser: 1000
    runAsGroup: 1000
    fsGroup: 1000
  securityContext: {}

  serviceAccount:
    # When `create` is true, a ServiceAccount named after the release is
    # created and bound to the Deployment. Set `create: false` and supply
    # `name` to bind to an existing one.
    create: true
    name: ""
    annotations: {}

  # Container resource requests/limits. Tune `processing.numWorkers` below in
  # tandem; each worker process consumes RAM proportional to typical batch
  # size.
  resources:
    requests:
      cpu: "16"
      memory: "16Gi"
    limits:
      cpu: "96"
      memory: "96Gi"

  # GPU is *not* required for service mode (the pipeline calls remote NIM
  # endpoints over HTTP). Leave `gpu.enabled` false unless you intentionally
  # want to schedule onto a GPU node for some custom workload.
  gpu:
    enabled: false
    count: 0
    productClass: ""

  # Liveness, readiness, and startup probes.  All use the dedicated
  # GET /v1/health endpoint which returns {"status":"ok"} with no auth.
  livenessProbe:
    enabled: true
    httpGet:
      path: /v1/health
      port: http
    initialDelaySeconds: 30
    periodSeconds: 30
    timeoutSeconds: 10
    failureThreshold: 6

  readinessProbe:
    enabled: true
    httpGet:
      path: /v1/health
      port: http
    initialDelaySeconds: 5
    periodSeconds: 10
    timeoutSeconds: 5
    failureThreshold: 6

  startupProbe:
    enabled: true
    httpGet:
      path: /v1/health
      port: http
    initialDelaySeconds: 5
    periodSeconds: 5
    timeoutSeconds: 5
    failureThreshold: 60

  # Set service.installFfmpeg=true to install ffmpeg/ffprobe in the service
  # container at startup by setting INSTALL_FFMPEG=true for the entrypoint.
  # This requires package-repository network egress, a writable root
  # filesystem, and security policy that allows the image's scoped sudo use.
  # Do not also set INSTALL_FFMPEG manually in service.env.
  installFfmpeg: true

  # OpenTelemetry env defaults for the retriever service container.
  otel:
    enabled: true
    serviceName: nemo-retriever-service
    env: {}

  # Extra env vars (after the chart-managed ones). Use `envFrom` to pull
  # whole Secrets/ConfigMaps in.
  env: []
  #  - name: PYTHONFAULTHANDLER
  #    value: "1"
  envFrom: []

  # Extra arbitrary volumes / mounts (e.g. for shipping CA bundles).
  extraVolumes: []
  extraVolumeMounts: []

# =============================================================================
# Kubernetes Service (network)
# =============================================================================
networkService:
  type: NodePort
  port: 7670
  # Optional: override the targetPort name. Should match container port name.
  targetPort: http
  annotations: {}
  labels: {}
  # NodePort assignment (only used when type=NodePort).
  # Leave empty to let Kubernetes auto-assign from the 30000-32767 range.
  nodePort: 30670

# =============================================================================
# Topology — single-pod vs split gateway/worker architecture
# =============================================================================
# When mode is "standalone" (default), the chart creates a single Deployment
# that runs both realtime and batch pools in-process. This is the simplest
#
# When mode is "split", the chart creates three Deployments:
#   1. gateway  — lightweight proxy that routes uploads to the correct backend
#   2. realtime — worker pod(s) for low-latency page processing
#   3. batch    — worker pod(s) for throughput-oriented document processing
#
# External traffic always enters through the gateway Service; realtime and
# batch Services are ClusterIP-only and not directly exposed.
topology:
  mode: standalone   # "standalone" or "split"

  gateway:
    replicas: 1
    resources:
      requests:
        cpu: "4"
        memory: "12Gi"
      limits:
        cpu: "16"
        memory: "48Gi"
    nodeSelector: {}
    tolerations: []
    affinity: {}

  realtime:
    replicas: 1
    resources:
      requests:
        cpu: "8"
        memory: "24Gi"
      limits:
        cpu: "16"
        memory: "96Gi"
    gpu:
      enabled: false
      count: 0
    nodeSelector: {}
    tolerations: []
    affinity: {}
    # -------------------------------------------------------------------
    # Horizontal Pod Autoscaler — realtime
    # -------------------------------------------------------------------
    # The HPA combines three independent signals (any can be disabled):
    #   • queueBacklog          — gateway backlog per replica
    #   • processingLatencyP95  — 95th-percentile pipeline duration (s)
    #   • cpu                   — utilization fallback, plain Resource metric
    #
    # Metrics that depend on Prometheus only render when
    # `autoscaling.queueDepth.backend` is set to "prometheus-adapter"
    # (the default) AND the per-metric `enabled` flag below is true.
    # If neither queue-depth nor latency is enabled, the HPA degrades
    # to the CPU fallback so the chart still scales on a vanilla
    # cluster.
    #
    # The deprecated `targetCPUUtilizationPercentage` knob is honoured
    # as an alias for `metrics.cpu.targetUtilizationPercentage` so
    # existing values files keep working.
    hpa:
      enabled: true
      minReplicas: 1
      maxReplicas: 8
      metrics:
        queueBacklog:
          enabled: true
          # AverageValue across replicas; one pod normally has 24 slots.
          target: "24"
        processingLatencyP95:
          enabled: true
          # Seconds of per-item p95 latency that triggers scale-up.
          # Realtime pages typically finish in < 5s; 30s is the slow-stage
          # threshold (OCR + chart inference on a hot path).
          targetSeconds: "30"
        cpu:
          enabled: false
          targetUtilizationPercentage: 60
      # k8s HPA `v2` behavior block — passed through verbatim. Realtime
      # scales out aggressively (double or +2 pods per minute) and
      # scales down slowly to avoid flapping during burst-then-drain
      # workloads.
      behavior:
        scaleUp:
          stabilizationWindowSeconds: 30
          policies:
            - type: Percent
              value: 100
              periodSeconds: 60
            - type: Pods
              value: 2
              periodSeconds: 60
        scaleDown:
          stabilizationWindowSeconds: 300
          policies:
            - type: Percent
              value: 50
              periodSeconds: 120
      # Legacy alias — kept so values files written before the metrics
      # block existed still scale on CPU. When set AND
      # `metrics.cpu.enabled` is false, the chart auto-flips CPU on
      # with this value. Remove once you migrate to `metrics.cpu`.
      targetCPUUtilizationPercentage: null

  batch:
    replicas: 1
    resources:
      requests:
        cpu: "16"
        memory: "48Gi"
      limits:
        cpu: "32"
        memory: "128Gi"
    gpu:
      enabled: true
      count: 1
    nodeSelector: {}
    tolerations: []
    affinity: {}
    # Batch HPA tuning differs from realtime: a deeper queue is normal
    # (we *want* batching), and replicas usually carry expensive GPUs
    # that we'd rather not churn. So:
    #   • Higher queueBacklog target (allow steady-state fill)
    #   • Higher p95 latency target (batch jobs can legitimately take
    #     minutes per item)
    #   • Longer scaleDown stabilization window (no GPU thrash)
    hpa:
      enabled: true
      minReplicas: 1
      maxReplicas: 4
      metrics:
        queueBacklog:
          enabled: true
          target: "48"
        processingLatencyP95:
          enabled: true
          targetSeconds: "120"
        cpu:
          enabled: false
          targetUtilizationPercentage: 80
      behavior:
        scaleUp:
          stabilizationWindowSeconds: 60
          policies:
            - type: Pods
              value: 1
              periodSeconds: 120
        scaleDown:
          stabilizationWindowSeconds: 600
          policies:
            - type: Pods
              value: 1
              periodSeconds: 300
      targetCPUUtilizationPercentage: null

  vectordb:
    replicas: 1
    port: 7671
    # Request a GPU on the vectordb pod when using in-pod query embedding
    # (serviceConfig.localModels.embed.enabled=true without a remote embed URL).
    gpu:
      enabled: false
      count: 1
    resources:
      requests:
        cpu: "8"
        memory: "16Gi"
      limits:
        cpu: "16"
        memory: "96Gi"
    nodeSelector: {}
    tolerations: []
    affinity: {}
    persistence:
      enabled: true
      storageClass: ""
      accessModes:
        - ReadWriteOnce
      size: "50Gi"

  # ---------------------------------------------------------------------------
  # OpenTelemetry Collector
  # ---------------------------------------------------------------------------
  # Deploys an in-cluster OpenTelemetry Collector for traces, metrics, and logs.
  # Disable with `topology.otel.enabled: false`.  The collector exposes OTLP
  # gRPC (4317) and OTLP HTTP (4318) receivers by default.
  otel:
    enabled: true
    image:
      repository: otel/opentelemetry-collector-contrib
      tag: "0.127.0"
      pullPolicy: IfNotPresent
    replicas: 1
    ports:
      otlpGrpc: 4317
      otlpHttp: 4318
      prometheus: 8889
    resources:
      requests:
        cpu: "250m"
        memory: "256Mi"
      limits:
        cpu: "1"
        memory: "512Mi"
    nodeSelector: {}
    tolerations: []
    affinity: {}
    config:
      extensions:
        health_check:
          endpoint: "0.0.0.0:13133"
      receivers:
        otlp:
          protocols:
            grpc:
              endpoint: "0.0.0.0:4317"
            http:
              endpoint: "0.0.0.0:4318"
      processors:
        batch:
          timeout: 5s
          send_batch_size: 1024
      exporters:
        debug:
          verbosity: basic
      service:
        extensions: [health_check]
        pipelines:
          traces:
            receivers: [otlp]
            processors: [batch]
            exporters: [debug]
          metrics:
            receivers: [otlp]
            processors: [batch]
            exporters: [debug]
          logs:
            receivers: [otlp]
            processors: [batch]
            exporters: [debug]

  # ---------------------------------------------------------------------------
  # Zipkin
  # ---------------------------------------------------------------------------
  # Enabled by default to preserve managed trace-backend parity with the legacy
  # 26.1.2 Helm chart. Set false to avoid deploying chart-owned Zipkin.
  zipkin:
    enabled: true
    image:
      repository: openzipkin/zipkin
      tag: "3.5.0"
      pullPolicy: IfNotPresent
    port: 9411
    javaOpts: "-Xms128m -Xmx512m -XX:+ExitOnOutOfMemoryError"
    livenessProbe:
      enabled: true
      tcpSocket:
        port: http
      initialDelaySeconds: 30
      periodSeconds: 30
      timeoutSeconds: 5
      failureThreshold: 6
    readinessProbe:
      enabled: true
      tcpSocket:
        port: http
      initialDelaySeconds: 5
      periodSeconds: 10
      timeoutSeconds: 5
      failureThreshold: 6
    startupProbe:
      enabled: true
      tcpSocket:
        port: http
      initialDelaySeconds: 5
      periodSeconds: 5
      timeoutSeconds: 5
      failureThreshold: 60
    resources:
      requests:
        cpu: "250m"
        memory: "512Mi"
      limits:
        cpu: "1"
        memory: "1Gi"
    nodeSelector: {}
    tolerations: []
    affinity: {}
    podSecurityContext: {}
    securityContext: {}
    exporter:
      enabled: true
      endpoint: ""

# =============================================================================
# retriever-service.yaml configuration
# =============================================================================
# Everything in `serviceConfig` is rendered into a ConfigMap and mounted as
# /etc/nemo-retriever/retriever-service.yaml inside the container, replacing
# the bundled default that ships in the image.
#
# When a NIM is enabled under `nimOperator.<key>.enabled: true` and the NIM
# Operator CRDs are installed in the cluster, the chart auto-injects the
# in-cluster NIMService URL below into the matching
# `serviceConfig.nimEndpoints.*` field. An explicit value supplied here
# always wins (use it to point at an external NIM, e.g. build.nvidia.com).
serviceConfig:
  server:
    host: "0.0.0.0"
    port: 7670

  logging:
    level: "INFO"
    file: "/var/lib/nemo-retriever/retriever-service.log"
    format: "%(asctime)s | %(levelname)s | %(name)s | %(message)s"

  # External NIM endpoints. Used as-is when the operator NIM is disabled
  # or when the NIM Operator CRDs are absent. When the corresponding
  # `nimOperator.<key>.enabled` is true (and the CRDs exist), the
  # in-cluster NIMService URL is used unless explicitly set here.
  nimEndpoints:
    pageElementsInvokeUrl: ""
    tableStructureInvokeUrl: ""
    ocrInvokeUrl: ""
    embedInvokeUrl: ""
    # Optional remote VLM endpoint for image captioning (Nemotron 3 Nano
    # Omni). Auto-wired from the in-cluster Service when
    # `nimOperator.nemotron_3_nano_omni_30b_a3b_reasoning.enabled=true`
    # (and the NIM Operator CRDs are present). Set explicitly to point
    # at a hosted endpoint.
    captionInvokeUrl: ""
    # Model identifier passed to the remote caption endpoint. Auto-set
    # to the Omni reasoning model id (matching
    # `nemo_retriever.caption.model_profiles.OMNI_REMOTE_MODEL_ID`) when
    # the operator-managed Omni NIM is enabled. Override to point at a
    # different VLM SKU.
    captionModelName: ""
    # gRPC endpoint for the Parakeet ASR NIM (e.g. "parakeet-nim:50051").
    # Required for audio/video ingestion in service mode (without torch).
    audioGrpcEndpoint: ""

  # In-pod Hugging Face models (alternative to nimEndpoints above).
  # Requires a GPU-enabled service image with nemo-retriever[local] installed.
  # NIM URLs always take precedence when both are configured for a stage.
  localModels:
    enabled: false
    hfCacheDir: ""
    device: ""
    warmupOnStartup: false
    maxTasksPerChild: 0
    maxProcessPoolWorkers: 1
    extract:
      enabled: true
      useTableStructure: true
      ocrVersion: v2
      ocrLang: ""
    embed:
      enabled: true
      modelName: nvidia/llama-nemotron-embed-vl-1b-v2
      localIngestEmbedBackend: hf
      gpuMemoryUtilization: 0.45
    asr:
      enabled: true

  # FastMCP agent integration mounted at serviceConfig.mcp.path (default /mcp).
  mcp:
    enabled: false
    path: "/mcp"
    baseUrl: ""
    enableWriteTools: true
    maxConcurrency: 8
    requestTimeoutS: 60.0
    ingestTimeoutS: 1800.0
    pollIntervalS: 2.0

  # Remote LLM endpoint used by POST /v1/answer. When the generic
  # `nimOperator.answer_llm.enabled` block is enabled (and the NIM
  # Operator CRDs exist), the chart auto-wires apiBase to that
  # in-cluster OpenAI-compatible /v1 endpoint and flips answering on.
  # Explicit apiBase/model values below always win, which lets operators
  # point /v1/answer at an external OpenAI-compatible endpoint instead.
  llm:
    enabled: false
    # Optional explicit LiteLLM/OpenAI model id. Leave empty to inherit
    # `nimOperator.answer_llm.model` when the operator-managed answer LLM
    # is enabled; set this when using an external LLM endpoint.
    model: ""
    apiBase: ""
    # Optional Secret reference for external LLM credentials. The Secret is
    # mounted as NEMO_RETRIEVER_LLM_API_KEY and read by the service CLI, so
    # the key is not rendered into the retriever-service ConfigMap. When this
    # is empty and nimOperator.answer_llm is enabled, the service mounts that
    # NIM's authSecret instead because LiteLLM/OpenAI-compatible clients
    # require a credential value even for local NIM endpoints.
    apiKeySecret:
      name: ""
      key: api_key
      optional: false
    temperature: 0.0
    topP: null
    maxTokens: 512
    extraParams: {}
    numRetries: 3
    timeout: 180.0
    ragSystemPrompt: ""
    # Optional prefix prepended to the RAG system prompt. Leave empty
    # unless an endpoint needs model-specific prompt directives.
    ragSystemPromptPrefix: ""
    # Unified request-level reasoning control for /v1/answer. Defaults to true
    # so external OpenAI-compatible providers receive only standard request
    # parameters. Set false for Nemotron endpoints that should receive the
    # portable no-reasoning controls.
    reasoningEnabled: true

  # Pipeline worker pools.  Workers are abstract dispatchers — sizing
  # depends on whether they do local GPU work or fan out to remote NIMs.
  # For CPU-only NIM-forwarding nodes, higher worker counts are fine.
  pipeline:
    realtimeWorkers: 24
    realtimeQueueSize: 2048
    batchWorkers: 48
    batchQueueSize: 8192

  workQueue:
    spoolDirectory: /tmp/nemo-retriever-work
    spoolLimitBytes: 21474836480
    claimTimeoutS: 30.0
    leaseTtlS: 60.0
    heartbeatIntervalS: 20.0
    maxDeliveryAttempts: 3
    # Cluster-wide budgets; do not infer these from worker replica counts.
    maxActiveLeases:
      realtime: 8
      batch: 48

  resources:
    maxMemoryMb: null
    maxCpuCores: null
    gpuDevices: []
    maxUploadBytes: 500000000

  # VectorDB pod configuration.  When enabled, a dedicated LanceDB pod
  # is deployed.  Workers POST embeddings to it; the gateway proxies
  # /v1/query to it for retrieval.
  vectordb:
    enabled: true
    lancedbUri: "/data/vectordb"
    tableName: "nemo_retriever"
    embedModel: "nvidia/llama-nemotron-embed-vl-1b-v2"
    embedModelProviderPrefix: ""

  # Optional bearer-token authentication. When apiToken is set, every
  # request must carry "Authorization: Bearer <token>".
  auth:
    apiToken: null
    headerName: "Authorization"
    bypassPaths:
      - "/v1/health"
      - "/docs"
      - "/openapi.json"
      - "/redoc"


# =============================================================================
# Retriever Results volume
# =============================================================================
# A PVC mounted at `mountPath` for storing per-job page results as JSON files.
# The gateway writes worker handoffs here and serves retained rows from this
# directory instead of retaining bulky rows in process memory.
# In split mode, only the gateway mounts this claim. Workers hand retained
# rows back to the gateway after processing, so realtime and batch replicas
# remain independently schedulable across nodes with ReadWriteOnce storage.
#
# Use local/host-path storage for results to avoid network I/O overhead.
# The default ReadWriteOnce access mode supports the default single gateway;
# configuring multiple gateway replicas requires storage and job-state
# coordination beyond the current split-topology contract.
retrieverResults:
  enabled: true
  # Immutable result generations and interrupted temporary files are removed
  # opportunistically after this age. Keep this at least as long as the job
  # tracker stale and terminal retention windows.
  ttlSeconds: 28800
  existingClaim: ""
  storageClass: ""
  accessModes:
    - ReadWriteOnce
  size: 50Gi
  mountPath: /retriever_results
  annotations: {}

# =============================================================================
# General persistence (logs and other non-scheduler files)
# =============================================================================
# A pre-existing ReadWriteOnce PVC mounted at `mountPath`. The gateway scheduler
# never stores job state or payloads here; its spool is always under the
# size-limited /tmp emptyDir.
persistence:
  enabled: true
  # When `existingClaim` is set the chart will skip PVC creation and mount
  # the named claim instead. Useful for promotion across releases.
  existingClaim: ""
  storageClass: ""
  accessModes:
    - ReadWriteOnce
  size: 50Gi
  mountPath: /var/lib/nemo-retriever
  annotations: {}

# =============================================================================
# Ingress (on by default)
# =============================================================================
ingress:
  enabled: true
  className: ""
  annotations: {}
  #   nginx.ingress.kubernetes.io/proxy-body-size: "200m"
  #   nginx.ingress.kubernetes.io/proxy-read-timeout: "600"
  hosts:
    - host: nemo-retriever.local
      paths:
        - path: /
          pathType: Prefix
  tls: []
  #  - secretName: nemo-retriever-tls
  #    hosts:
  #      - nemo-retriever.example.com

# =============================================================================
# HorizontalPodAutoscaler — standalone mode
# =============================================================================
# Standalone job and scheduler state are process-local. Keep maxReplicas=1
# unless an external coordination design is introduced. The template remains
# available for future coordinated deployments.
#
# For split mode (`topology.mode: split`), HPA configuration lives under
# `topology.realtime.hpa` and `topology.batch.hpa` instead — those pods
# are stateless and can scale on queue depth + latency. See the
# `autoscaling.queueDepth` block below for the backend wiring.
autoscaling:
  enabled: false
  minReplicas: 1
  maxReplicas: 1
  targetCPUUtilizationPercentage: 70
  targetMemoryUtilizationPercentage: ""
  behavior: {}

  # ---------------------------------------------------------------------------
  # Queue-depth + latency autoscaling backend (split mode only)
  # ---------------------------------------------------------------------------
  # Chooses how the HPA gets at `nemo_retriever_pool_queue_depth_ratio`
  # and `nemo_retriever_pool_processing_duration_seconds`. The metrics
  # themselves are *always* published (see H1 — instrumented in
  # services/pipeline_pool.py); this block only controls the *delivery
  # path* into the Kubernetes metrics API.
  #
  # Supported values:
  #
  #   prometheus-adapter (recommended for production)
  #     • Requires: Prometheus + prometheus-adapter installed in-cluster.
  #     • Renders: `prometheus-adapter-rules.yaml` ConfigMap with the
  #       PromQL needed for the External Metrics API.
  #     • Chart hands the ConfigMap name to the operator; operator
  #       points its prometheus-adapter at it. No subchart bundling.
  #
  #   cpu
  #     • No Prometheus needed. The realtime/batch HPAs degrade to
  #       a Resource-typed CPU metric. Convenient for clusters
  #       without an observability stack — but ignores queue depth,
  #       so scaling lags behind queue saturation by minutes.
  #
  #   keda
  #     • Reserved hook. The chart does NOT install or render KEDA
  #       resources today; setting this disables the chart-managed
  #       HPAs so you can apply your own ScaledObject snippet.
  #       See README "Queue-depth autoscaling" for the template.
  #
  # The recommendation in this codebase is `prometheus-adapter`.
  # Reasoning is documented in helm/README.md.
  queueDepth:
    backend: "prometheus-adapter"

    prometheusAdapter:
      # External-metrics API names exposed by the adapter for our two
      # PromQL queries. These are what the HPA `metric.name` field
      # references; the adapter ConfigMap's `name.as` field has to match.
      queueBacklogMetric: "nemo_retriever_gateway_work_queue_backlog"
      # Deprecated for one release; when set, overrides queueBacklogMetric.
      queueDepthRatioMetric: ""
      processingLatencyP95Metric: "nemo_retriever_pool_processing_duration_p95_seconds"
      # When `rules.create` is true the chart renders a ConfigMap with
      # the PromQL rules below; the operator wires their existing
      # prometheus-adapter at this ConfigMap via:
      #   helm upgrade prometheus-adapter ... \
      #     --set rules.existing=<release>-prom-adapter-rules
      # Set to false to bring-your-own rules (e.g. a cluster-wide ones).
      rules:
        create: true
        # Window for the histogram_quantile() PromQL — short enough to
        # react within a single HPA cycle (15s default), long enough to
        # smooth out noise. 2m matches the prometheus-adapter docs'
        # recommended default for short-lived workloads.
        latencyWindow: "2m"

# =============================================================================
# Prometheus integration (off by default; enable only when the cluster has
# the ServiceMonitor CRD and Prometheus Operator)
# =============================================================================
# When `topology.mode: split` is active, queue-depth-driven autoscaling
# needs the worker pods' `/metrics` endpoint to be scraped by Prometheus.
# Stock `retriever-dev` clusters do not install the ServiceMonitor CRD, so
# split mode does not render one unless an operator explicitly opts in.
#
# Operators on standalone (or operators who scrape via another route —
# Prometheus Pushgateway, an external agent, OpenTelemetry, etc.) can
# leave both flags false and no ServiceMonitor will be produced.
serviceMonitor:
  enabled: false
  autoEnableInSplitMode: false
  namespace: ""
  interval: 15s
  scrapeTimeout: 10s
  path: /metrics
  labels: {}
  # Labels copied from the Service onto each scraped sample. We always
  # propagate `app.kubernetes.io/component` so prometheus-adapter (and
  # human Grafana queries) can filter on `component="realtime"` /
  # `component="batch"` even when our publisher's `pool` label happens
  # to disagree (e.g. during the standalone→split migration window).
  targetLabels:
    - app.kubernetes.io/component
  relabelings: []
  metricRelabelings: []

# =============================================================================
# NIMs (NVIDIA NIM Operator integration)
# =============================================================================
# Master switch — set `nims.enabled: false` to disable the entire NIM
# sub-stack regardless of per-NIM flags below. Handy for installs that
# only want the retriever service (talking to external NIM URLs) and
# nothing GPU-shaped in this release.
nims:
  enabled: true

# Each `nimOperator.<key>` block renders a pair of `NIMCache` + `NIMService`
# custom resources (apiVersion: apps.nvidia.com/v1alpha1) that the NIM
# Operator reconciles into a Deployment + Service per model.
#
# The per-model templates are gated on ALL of:
#   1. `.Capabilities.APIVersions.Has "apps.nvidia.com/v1alpha1"` — the
#      NIM Operator CRDs must be installed in the cluster.
#   2. `nims.enabled: true` — master switch above.
#   3. `nimOperator.<key>.enabled: true` — per-model toggle below.
#
# When auto-resolution is in effect, the chart populates
# `serviceConfig.nimEndpoints.*` with the NIMService's in-cluster URL
# (e.g. http://nemotron-page-elements-v3:8000/v1/infer). An explicit value
# in `serviceConfig.nimEndpoints.*` always wins.
#
# NIMCache resources may carry `helm.sh/resource-policy: keep` so model
# downloads survive `helm uninstall` (see nimOperator.nimCache.keepOnUninstall).
nimOperator:
  # ---------------------------------------------------------------------------
  # NIMCache storage defaults (currently informational — each per-NIM block
  # carries its own storage.pvc settings).
  # ---------------------------------------------------------------------------
  nimCache:
    # When true, every NIMCache is annotated `helm.sh/resource-policy: keep`
    # and Helm will NOT delete it on uninstall (PVC + pulled weights remain).
    # NIMService CRs are always removed by Helm uninstall. Set false on dev
    # clusters when you want `helm uninstall` to delete caches too.
    keepOnUninstall: true
    # NIMCache/NIMService names retired from the chart. deploy.sh deletes these
    # after each successful helm reconcile so they do not linger with keep-on-uninstall.
    pruneRetiredNimResources:
      - nemotron-ocr-v1
    pvc:
      create: true
      # If set, applies to every per-NIM PVC that doesn't override
      # storageClass explicitly.
      storageClass: ""
      size: "25Gi"
      volumeAccessMode: ReadWriteOnce

  # ---------------------------------------------------------------------------
  # NIMCache GPU / profile filter (chart-wide default)
  # ---------------------------------------------------------------------------
  # Renders directly under `spec.source.ngc.model` on every NIMCache the
  # chart provisions; the NIM Operator uses it to restrict which model
  # profiles each cache job downloads. Without a filter the operator
  # caches every profile applicable to the GPUs it detects, which on
  # heterogeneous clusters (or any cluster with ≥ 3 NIMs) wastes tens of
  # GiB of PVC storage and NGC bandwidth.
  #
  # Two filter dimensions are supported (matching the NIMCache CRD —
  # set whichever fits your cluster; ``gpus`` is the common case):
  #
  #   gpus:                 list of GPU selectors. Each selector is
  #     - ids: ["<pci>"]    ``{ids: [<PCI device IDs>], product:
  #       product: "..."    "<NVIDIA marketing name>"}``. Profiles that
  #                         do not match at least one selector are
  #                         skipped.
  #   profiles:             list of profile UUIDs. Only those exact
  #     - "<uuid>"          profiles are downloaded.
  #
  # Default: ``{}`` — no filter, preserves the operator's "cache all
  # profiles applicable to detected GPUs" behaviour. Override examples:
  #
  #   # Homogeneous H100 80 GB SXM cluster:
  #   modelProfile:
  #     gpus:
  #       - ids: ["26B5"]
  #         product: "NVIDIA-H100-80GB-HBM3"
  #
  # Per-NIM overrides live next to each NIM's ``enabled:`` flag below
  # (``nimOperator.<key>.modelProfile``) and REPLACE this chart-wide
  # default when non-empty (they do not merge). Documented in
  # helm/README.md §"Filtering cached GPU profiles".
  modelProfile: {}

  # Default GPU limit rendered on every NIMService when a per-NIM
  # ``resources`` block is empty (``{}``). The NIM Operator does not
  # reliably set ``spec.resources.limits.nvidia.com/gpu`` from the model
  # profile on all supported versions; without this, NIM pods can start
  # without GPU access. Set to ``null`` to omit the resources block and
  # rely on operator reconciliation only (see helm/README.md). Per-NIM
  # ``resources`` overrides the entire block when non-empty.
  nimServiceGpuLimit: 1

  # OpenTelemetry defaults applied to NIMService env blocks.
  otel:
    enabled: true
    endpoint: ""
    tritonPath: "/v1/traces"
    env:
      NIM_ENABLE_OTEL: "true"
      NIM_OTEL_TRACES_EXPORTER: "otlp"
      NIM_OTEL_METRICS_EXPORTER: "console"
      TRITON_OTEL_RATE: "1"

  # ---------------------------------------------------------------------------
  # Per-NIM defaults
  # ---------------------------------------------------------------------------
  # Two enablement tiers, matching the docs' "core vs optional" contract
  # (see docs/extraction/deployment-options.md "Core NIMs for the default
  # extraction pipeline (26.05)"):
  #
  #   Core (enabled: true by default, auto-wired into the service config):
  #     - page_elements
  #     - table_structure
  #     - ocr
  #     - vlm_embed
  #
  #   Optional (enabled: false by default, NOT auto-wired):
  #     - rerankqa                                 (VL reranker,
  #                                                 ≈ 3.1 GiB GPU)
  #     - nemotron_parse                           (Parse v1.2, ≈ 3.5 GiB GPU)
  #     - nemotron_3_nano_omni_30b_a3b_reasoning   (Omni 30B caption NIM,
  #                                                 ≈ 62 GiB BF16 weights)
  #     - answer_llm                               (OpenAI-compatible answer LLM,
  #                                                 Super-49B defaults)
  #     - audio                                    (Parakeet ASR)
  #
  # Two flags gate every NIM, so each one is opt-in along two independent
  # axes:
  #   * `nims.enabled: true`              — chart-wide master switch.
  #   * `nimOperator.<key>.enabled: true` — per-NIM toggle below.
  #
  # The optional NIMs sit behind their per-NIM toggles to honor the
  # "optional and disabled by default" contract in 26.05 — turning them
  # on alongside the core stack at install time would silently pull tens
  # of gigabytes of model weights (Omni 30B ≈ 62 GiB BF16) and consume
  # an additional dedicated GPU per NIM with no opt-in from the operator.
  # See helm/README.md "Recommended minimal install (26.05)" for the
  # opt-in flags and "Image tag conventions" for what the
  # ``1.7.0-variant`` Parse / Omni tags mean.
  # ---------------------------------------------------------------------------

  # Page-elements detector (YOLOX-style). Used by the page-element extraction
  # stage of the ingest pipeline.
  page_elements:
    enabled: true
    image:
      repository: nvcr.io/nim/nvidia/nemotron-page-elements-v3
      tag: "1.8.0"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "25Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
      - name: NIM_TRITON_MAX_BATCH_SIZE
        value: "32"
      - name: NIM_TRITON_CPU_THREADS_PRE_PROCESSOR
        value: "2"
      - name: NIM_TRITON_CPU_THREADS_POST_PROCESSOR
        value: "1"
      - name: OMP_NUM_THREADS
        value: "2"

  # Table structure detector. Used by the table extraction stage.
  table_structure:
    enabled: true
    image:
      repository: nvcr.io/nim/nvidia/nemotron-table-structure-v1
      tag: "1.8.0"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "25Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
      - name: NIM_TRITON_RATE_LIMIT
        value: "3"
      - name: NIM_TRITON_MAX_BATCH_SIZE
        value: "32"
      - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
        value: "2048"
      - name: OMP_NUM_THREADS
        value: "1"

  # Nemotron OCR v2. Used by the OCR stage of the pipeline.
  ocr:
    enabled: true
    # NIMService / NIMCache resource name and in-cluster Service DNS label.
    nimServiceName: nemotron-ocr-v2
    image:
      repository: nvcr.io/nim/nvidia/nemotron-ocr-v2
      tag: "1.4.0"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "25Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: OMP_NUM_THREADS
        value: "8"
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
      - name: NIM_TRITON_MAX_BATCH_SIZE
        value: "32"

  # Llama Nemotron multimodal (VL) embed 1B v2. Used by the embedding stage,
  # the vectordb query path, and multimodal ingest.
  vlm_embed:
    enabled: true
    # NIMService / NIMCache resource name and in-cluster Service DNS label.
    nimServiceName: llama-nemotron-embed-vl-1b-v2
    image:
      repository: nvcr.io/nim/nvidia/llama-nemotron-embed-vl-1b-v2
      tag: "1.12.0"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "50Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
      - name: OMP_NUM_THREADS
        value: "1"
      # Optional: enable if your NIM build supports throughput mode for this SKU.
      # - name: NIM_TRITON_PERFORMANCE_MODE
      #   value: throughput

  # ---------------------------------------------------------------------------
  # Optional / advanced NIMs. Disabled by default unless noted below;
  # opt in to the ones your pipeline or service endpoints actually use.
  # ---------------------------------------------------------------------------

  # Generic OpenAI-compatible answer-generation LLM NIM.
  #
  # Disabled by default because answer LLMs are larger than the core
  # extraction NIMs. The default image/model is Super-49B and requests
  # two GPUs. Override `image.repository`, `image.tag`, `model`,
  # `nimServiceName`, `resources`, `modelProfile`, and `env` to swap in
  # another compatible LLM NIM such as Nemotron 3 Nano.
  answer_llm:
    enabled: false
    # NIMService / NIMCache resource name and in-cluster Service DNS label.
    nimServiceName: answer-llm
    # LiteLLM/OpenAI model id rendered into retriever-service.yaml when
    # this operator-managed LLM is enabled and serviceConfig.llm.model is
    # left empty.
    model: "openai/nvidia/llama-3.3-nemotron-super-49b-v1.5"
    # Optional prompt prefix to apply when this operator-managed LLM
    # auto-wires /v1/answer. Leave empty to keep the LLM deployment and
    # prompt defaults model-neutral; use serviceConfig.llm.reasoningEnabled
    # for the portable request-level reasoning toggle.
    ragSystemPromptPrefix: ""
    image:
      repository: nvcr.io/nim/nvidia/llama-3.3-nemotron-super-49b-v1.5
      tag: "2.0.5"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    # Cache only the default BF16 2-GPU profile used by the Super-49B
    # default resources. Override this when deploying a different model,
    # profile, or GPU topology.
    modelProfile:
      profiles:
        - "1146f49f84dff5dea09f5aa633cc70b92d7d972223d67878c841cd0fbccad4fb"
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "250Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # The Super-49B default uses a 2-GPU tensor-parallel profile. This
    # non-empty resources block intentionally replaces
    # nimOperator.nimServiceGpuLimit. Override for smaller LLM NIMs.
    resources:
      limits:
        nvidia.com/gpu: 2
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TENSOR_PARALLEL_SIZE
        value: "2"
      # Conservative tensor-parallel startup default; override when your
      # deployment fabric has been validated with custom all-reduce enabled.
      - name: NIM_PASSTHROUGH_ARGS
        value: "--disable-custom-all-reduce"
      # Conservative single-pod tensor-parallel defaults that avoid
      # transport assumptions across GPU fabrics.
      - name: NCCL_IB_DISABLE
        value: "1"
      - name: NCCL_P2P_DISABLE
        value: "1"

  # Llama Nemotron rerank VL 1B v2 — multimodal reranking NIM.
  #
  # Disabled by default per the 26.05 "optional and disabled by default"
  # contract (see docs/extraction/deployment-options.md L21 and
  # docs/extraction/prerequisites-support-matrix.md L92 / L128). Opt in
  # only when the retrieval pipeline calls a reranker — enabling this NIM
  # consumes one GPU and ~3.1 GiB of GPU memory at runtime.
  #
  # IMPORTANT — VL vs text-only: 26.05 documents the **VL** reranker
  # (``llama-nemotron-rerank-vl-1b-v2``) as the supported NIM so the
  # ingest pipeline can score image + text passages together. The
  # text-only ``llama-nemotron-rerank-1b-v2`` is a different SKU that
  # silently degrades to text-only behaviour on multimodal queries.
  # Override the repository only when validated against the matching
  # service/Helm release.
  rerankqa:
    enabled: false
    image:
      repository: nvcr.io/nim/nvidia/llama-nemotron-rerank-vl-1b-v2
      tag: "1.11.0"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "50Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"

  # Nemotron Parse v1.2. Optional structured document parser.
  #
  # Disabled by default per the 26.05 "optional and disabled by default"
  # contract (see docs/extraction/deployment-options.md). Opt in only when
  # the pipeline runs `extract_method="nemotron_parse"`; enabling this
  # NIM consumes one GPU and ~3.5 GiB of GPU memory at runtime, plus
  # ~16 GB of on-disk NIM model cache (see prerequisites-support-matrix).
  #
  # The default image tag (`1.7.0-variant`) is the Nemotron Parse v1.2
  # NIM SKU shipped alongside the 26.05 release; the `-variant` suffix
  # marks it as the build that ships with the per-GPU TensorRT engine
  # variants required by NIM Operator profile reconciliation, matching
  # `nemotron-3-nano-omni-30b-a3b-reasoning:1.7.0-variant` and the row
  # for Nemotron Parse in
  # docs/extraction/prerequisites-support-matrix.md "Model hardware
  # requirements". Override to a different NGC tag only when validated
  # against the matching service/Helm release.
  nemotron_parse:
    enabled: false
    image:
      repository: nvcr.io/nim/nvidia/nemotron-parse-v1.2
      tag: "1.7.0-variant"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "100Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"

  # Nemotron 3 Nano Omni 30B A3B Reasoning. Multimodal reasoning VLM
  # used by the image-captioning stage.
  #
  # Disabled by default per the 26.05 "optional and disabled by default"
  # contract (see docs/extraction/deployment-options.md). This is the
  # canonical caption NIM — once enabled the chart auto-wires
  # `nim_endpoints.caption_invoke_url` + `caption_model_name` into the
  # retriever-service config (see helm/README.md "Image captioning
  # (Omni 30B)"), but it is intentionally opt-in because the model is
  # very large: ~62 GiB (BF16), ~33 GiB (FP8), ~21 GiB (NVFP4), plus
  # ~80 GB on-disk NIM cache. Bump `storage.pvc.size` below if your
  # weights need more than 300 GiB; GPU count is reconciled by the NIM
  # Operator from the model profile (override with `kubectl edit
  # nimservice` if necessary).
  #
  # The default image tag (`1.7.0-variant`) is the Omni NIM SKU shipped
  # alongside the 26.05 release; the `-variant` suffix marks it as the
  # build that ships with the per-GPU TensorRT engine variants required
  # by NIM Operator profile reconciliation, matching the row for the
  # Omni caption NIM in
  # docs/extraction/prerequisites-support-matrix.md "Model hardware
  # requirements" (footnote ³). Override to a different NGC tag only
  # when validated against the matching service/Helm release.
  nemotron_3_nano_omni_30b_a3b_reasoning:
    enabled: false
    image:
      repository: nvcr.io/nim/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
      tag: "1.7.0-variant"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "300Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 8000
        grpcPort: 8001
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"

  # Parakeet ASR. Optional audio transcription NIM.
  audio:
    enabled: false
    image:
      repository: nvcr.io/nim/nvidia/parakeet-1-1b-ctc-en-us
      tag: "1.5.0"
      pullPolicy: IfNotPresent
      pullSecrets:
        - ngc-secret
    authSecret: ngc-api
    # Per-NIM NIMCache GPU/profile filter. See `nimOperator.modelProfile`
    # above for the shape; non-empty REPLACES the chart-wide default.
    modelProfile: {}
    storage:
      pvc:
        create: true
        storageClass: ""
        size: "100Gi"
        volumeAccessMode: ReadWriteOnce
    replicas: 1
    # Empty (`{}`) uses `nimOperator.nimServiceGpuLimit` (default 1 GPU).
    # Set a non-empty map to replace the whole resources block.
    resources: {}
    nodeSelector: {}
    tolerations: []
    expose:
      service:
        type: ClusterIP
        port: 9000
        grpcPort: 50051
    otel:
      enabled: null
      serviceName: ""
      endpoint: ""
      env: {}
    env:
      - name: NIM_TAGS_SELECTOR
        value: "name=parakeet-1-1b-ctc-en-us,mode=ofl,vad=default,diarizer=disabled"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"