values.yaml

# -- Global chart configuration
nameOverride: ""
fullnameOverride: "rag-server"
# subsection: rag-server
# RAG Orchestrator Service
# -- Kubernetes scheduling
nodeSelector: {}
affinity: {}
tolerations: []

# -- Common service account for rag-server
serviceAccount:
  create: true
  name: ""
  automount: true
  annotations: {}

# -- Replicas for rag-server
replicaCount: 1

# -- Namespace for documentation/reference; not actively used in templates
namespace: "nv-nvidia-blueprint-rag"

# -- Image pull secret for all images used by this chart
imagePullSecret:
  name: "ngc-secret"
  registry: "nvcr.io"
  username: "$oauthtoken"
  password: ""
  create: true

# -- Secret containing API keys for NVIDIA NGC model registry
ngcApiSecret:
  name: "ngc-api"
  password: ""
  create: true

# -- Secret containing service-specific API keys for RAG services
apiKeysSecret:
  # -- Enable automatic secret creation from values below
  create: true
  # -- Name of the secret to create or reference
  name: "rag-api-keys"
  # -- Reference to existing secret (if set, chart won't create secret, will reference this instead)
  existingSecret: ""
  # -- Service-specific API keys (optional, empty means use NVIDIA_API_KEY fallback)
  # Note: Non-empty values act as enable flags - only non-empty keys are injected as env vars
  llmApiKey: ""
  embeddingsApiKey: ""
  rankingApiKey: ""
  queryRewriterApiKey: ""
  filterExpressionGeneratorApiKey: ""
  vlmApiKey: ""
  summaryLlmApiKey: ""
  reflectionLlmApiKey: ""

# -- RAG server container image
image:
  repository: nvcr.io/nvidia/blueprint/rag-server
  tag: "2.4.0"
  pullPolicy: Always

# -- RAG server service configuration
service:
  type: ClusterIP
  port: 8081

# -- RAG server container resources
resources:
  limits:
    memory: "64Gi"
  requests:
    memory: "8Gi"

# -- Probes for rag-server (optional)
livenessProbe:
  httpGet:
    path: /health
    port: 8081
  initialDelaySeconds: 10
  periodSeconds: 15
  timeoutSeconds: 5
  failureThreshold: 3
readinessProbe:
  httpGet:
    path: /health
    port: 8081
  initialDelaySeconds: 5
  periodSeconds: 10
  timeoutSeconds: 5
  failureThreshold: 3

# -- RAG server runtime configuration
server:
  workers: 8

# -- Enable/disable creation of prompt ConfigMap
promptConfig:
  enabled: true

# -- Environment variables for rag-server
envVars:
  EXAMPLE_PATH: "./nvidia_rag/rag_server"
  PROMPT_CONFIG_FILE: "/prompt.yaml"
  PROMETHEUS_MULTIPROC_DIR: "/tmp-data/prom_data"

  ##===Authentication===
  ## Service-specific API keys are now managed via apiKeysSecret section above.
  ## See apiKeysSecret configuration for details on setting service-specific keys.

  ##===MINIO specific configurations used to store multimodal base64 content===
  MINIO_ENDPOINT: "rag-minio:9000"
  MINIO_ACCESSKEY: "minioadmin"
  MINIO_SECRETKEY: "minioadmin"

  ##===Redis configurations for summary status tracking===
  REDIS_HOST: "rag-redis-master"
  REDIS_PORT: "6379"
  REDIS_DB: "0"

  ##===Vector DB specific configurations===
  # URL on which vectorstore is hosted
  APP_VECTORSTORE_URL: "http://milvus:19530" # Use "http://rag-eck-elasticsearch-es-http:9200" for elasticsearch
  # Type of vectordb used to store embedding supported type "milvus" or "elasticsearch"
  APP_VECTORSTORE_NAME: "milvus"
  # Index type (e.g., GPU_CAGRA)
  APP_VECTORSTORE_INDEXTYPE: "GPU_CAGRA"
  # Type of vectordb search to be used
  APP_VECTORSTORE_SEARCHTYPE: "dense"
  # Type of ranker to use for vector store in case of Hybrid Search
  APP_VECTORSTORE_RANKER_TYPE: "rrf" # Can be "rrf" or "weighted"
  # Weight for dense vector search in case of "weighted" Hybrid Search
  APP_VECTORSTORE_DENSE_WEIGHT: "0.5"
  # Weight for sparse vector search in case of "weighted" Hybrid Search
  APP_VECTORSTORE_SPARSE_WEIGHT: "0.5"
  # Boolean to control GPU search for milvus vectorstore specific to rag-server
  APP_VECTORSTORE_ENABLEGPUSEARCH: "True"
  # ef: Parameter controlling query time/accuracy trade-off. Higher ef leads to more accurate but slower search.
  APP_VECTORSTORE_EF: "100"
  # Username for vector store authentication
  APP_VECTORSTORE_USERNAME: ""
  # Password for vector store authentication
  APP_VECTORSTORE_PASSWORD: ""
  # Elasticsearch API key auth (optional). Prefer these over username/password when set.
  # Provide either base64 APP_VECTORSTORE_APIKEY or split ID/SECRET.
  APP_VECTORSTORE_APIKEY_ID: ""
  APP_VECTORSTORE_APIKEY_SECRET: ""
  APP_VECTORSTORE_APIKEY: ""

  # vectorstore collection name to store embeddings
  COLLECTION_NAME: "multimodal_data"
  APP_RETRIEVER_SCORETHRESHOLD: "0.25"
  # Top K from vector DB, which goes as input to reranker model - not applicable if ENABLE_RERANKER is set to False
  VECTOR_DB_TOPK: "100"
  # Number of document chunks to insert in LLM prompt
  APP_RETRIEVER_TOPK: "10"

  ##===LLM Model specific configurations===
  APP_LLM_MODELNAME: "nvidia/llama-3.3-nemotron-super-49b-v1.5"
  # URL on which LLM model is hosted. If "", Nvidia hosted API is used
  APP_LLM_SERVERURL: "nim-llm:8000"
  # LLM model parameters
  LLM_MAX_TOKENS: "32768"
  LLM_TEMPERATURE: "0"
  LLM_TOP_P: "1.0"

  # Enable/disable thinking/reasoning for nemotron-3-nano models (30b variant)
  # Set to "true" to enable reasoning mode with reasoning_budget
  # Set to "false" to disable reasoning and get direct answers
  ENABLE_NEMOTRON_3_NANO_THINKING: "true"

  ##===Query Rewriter Model specific configurations===
  APP_QUERYREWRITER_MODELNAME: "nvidia/llama-3.3-nemotron-super-49b-v1.5"
  # URL on which query rewriter model is hosted. If "", Nvidia hosted API is used
  APP_QUERYREWRITER_SERVERURL: "nim-llm:8000"

  ##===Filter Expression Generator Model specific configurations===
  APP_FILTEREXPRESSIONGENERATOR_MODELNAME: "nvidia/llama-3.3-nemotron-super-49b-v1.5"
  # URL on which filter expression generator model is hosted. If "", Nvidia hosted API is used
  APP_FILTEREXPRESSIONGENERATOR_SERVERURL: "nim-llm:8000"
  # enable filter expression generator for natural language to filter expression conversion
  ENABLE_FILTER_GENERATOR: "False"

  ##===Embedding Model specific configurations===
  # URL on which embedding model is hosted. If "", Nvidia hosted API is used
  APP_EMBEDDINGS_SERVERURL: "nemoretriever-embedding-ms:8000/v1"
  APP_EMBEDDINGS_MODELNAME: "nvidia/llama-3.2-nv-embedqa-1b-v2"
  APP_EMBEDDINGS_DIMENSIONS: "2048"

  ##===Reranking Model specific configurations===
  # URL on which ranking model is hosted. If "", Nvidia hosted API is used
  APP_RANKING_SERVERURL: "nemoretriever-ranking-ms:8000"
  APP_RANKING_MODELNAME: "nvidia/llama-3.2-nv-rerankqa-1b-v2"
  ENABLE_RERANKER: "True"
  # Default confidence threshold for filtering documents by reranker relevance scores (0.0 to 1.0)
  RERANKER_CONFIDENCE_THRESHOLD: "0.0"

  ##===VLM Model specific configurations===
  ENABLE_VLM_INFERENCE: "false"
  # When true, allow falling back to LLM flow if no images are present in query/messages/context.
  # When false, always route generation through VLM when ENABLE_VLM_INFERENCE is true.
  VLM_TO_LLM_FALLBACK: "true"
  # Max images sent to VLM per request (query + context)
  APP_VLM_MAX_TOTAL_IMAGES: "5"
  # VLM generation parameters
  APP_VLM_MAX_TOKENS: "8192"
  APP_VLM_TEMPERATURE: "0.1"
  APP_VLM_TOP_P: "1.0"
  # VLM server URL
  APP_VLM_SERVERURL: "http://nim-vlm:8000/v1"
  # VLM model name
  APP_VLM_MODELNAME: "nvidia/nemotron-nano-12b-v2-vl"

  # === Text Splitter ===
  APP_TEXTSPLITTER_CHUNKSIZE: "2000"
  APP_TEXTSPLITTER_CHUNKOVERLAP: "200"

  # === General ===
  # Choose whether to enable citations in the response
  ENABLE_CITATIONS: "True"
  # Choose whether to enable/disable guardrails
  ENABLE_GUARDRAILS: "False"
  # Log level for server, supported level NOTSET, DEBUG, INFO, WARN, ERROR, CRITICAL
  LOGLEVEL: "INFO"
  # enable multi-turn conversation in the rag chain - this controls conversation history usage
  # while doing query rewriting and in LLM prompt
  ENABLE_MULTITURN: "True"
  # enable query rewriting for multiturn conversation in the rag chain.
  # This will improve accuracy of the retrieiver pipeline but increase latency due to an additional LLM call
  ENABLE_QUERYREWRITER: "False"
  # enable concatenating conversation history with current query for retrieval (when query rewriter is disabled)
  MULTITURN_RETRIEVER_SIMPLE: "False"
  # number of last n chat messages to consider from the provided conversation history (0 = no history)
  CONVERSATION_HISTORY: "0"

  # === Tracing ===
  APP_TRACING_ENABLED: "False"
  # HTTP endpoint
  APP_TRACING_OTLPHTTPENDPOINT: "http://rag-opentelemetry-collector:4318/v1/traces"
  # GRPC endpoint
  APP_TRACING_OTLPGRPCENDPOINT: "grpc://rag-opentelemetry-collector:4317"

  # === Reflection ===
  # enable reflection (context relevance and response groundedness checking) in the rag chain
  ENABLE_REFLECTION: "false"
  # Maximum number of context relevance loop iterations
  MAX_REFLECTION_LOOP: "3"
  # Minimum relevance score threshold (0-2)
  CONTEXT_RELEVANCE_THRESHOLD: "1"
  # Minimum groundedness score threshold (0-2)
  RESPONSE_GROUNDEDNESS_THRESHOLD: "1"
  # reflection llm
  REFLECTION_LLM: "nvidia/llama-3.3-nemotron-super-49b-v1.5"
  # reflection llm server url. If "", Nvidia hosted API is used
  REFLECTION_LLM_SERVERURL: "nim-llm:8000"

  # Choose whether to enable source metadata in document content during generation
  ENABLE_SOURCE_METADATA: "true"

  # Whether to filter content within <think></think> tags in model responses
  FILTER_THINK_TOKENS: "true"

  NEMO_GUARDRAILS_URL: "nemo-guardrails:7331"

  # enable iterative query decomposition
  ENABLE_QUERY_DECOMPOSITION: "false"
  # maximum recursion depth for iterative query decomposition
  MAX_RECURSION_DEPTH: "3"

# -- Ingestor Server
# subsection: ingestor-server
# Ingestor API Service
ingestor-server:
  enabled: true
  appName: ingestor-server

  # -- Pod scheduling
  nodeSelector: {}
  affinity: {}
  tolerations: []

  replicaCount: 1

  imagePullSecret:
    create: false
    name: "ngc-secret"
    registry: "nvcr.io"
    username: "$oauthtoken"
    password: ""

  image:
    repository: nvcr.io/nvidia/blueprint/ingestor-server
    tag: "2.4.0"
    pullPolicy: Always

  # -- Service config for ingestor-server
  service:
    type: ClusterIP
    port: 8082

  server:
    workers: 1

  # -- Probes for ingestor-server (optional)
  livenessProbe: {}
  readinessProbe: {}

  resources:
    limits:
      memory: "25Gi"
    requests:
      memory: "25Gi"

  envVars:
    # Path to example directory relative to repo root
    EXAMPLE_PATH: "src/nvidia_rag/ingestor_server"
    # Absolute path to custom prompt.yaml file
    PROMPT_CONFIG_FILE: "/prompt.yaml"
    # === Vector Store Configurations ===
    APP_VECTORSTORE_URL: "http://milvus:19530" # Use "http://rag-eck-elasticsearch-es-http:9200" for elasticsearch
    APP_VECTORSTORE_NAME: "milvus" # supported values: "milvus" or "elasticsearch"
    APP_VECTORSTORE_SEARCHTYPE: "dense"
    # Type of ranker to use for vector store in case of Hybrid Search
    APP_VECTORSTORE_RANKER_TYPE: "rrf" # Can be "rrf" or "weighted"
    # Weight for dense vector search in case of "weighted" Hybrid Search
    APP_VECTORSTORE_DENSE_WEIGHT: "0.5"
    # Weight for sparse vector search in case of "weighted" Hybrid Search
    APP_VECTORSTORE_SPARSE_WEIGHT: "0.5"
    APP_VECTORSTORE_ENABLEGPUINDEX: "True"
    APP_VECTORSTORE_ENABLEGPUSEARCH: "True"
    # Username  for vector store authentication
    APP_VECTORSTORE_USERNAME: ""
    # Password for vector store authentication
    APP_VECTORSTORE_PASSWORD: ""
    # Elasticsearch API key auth (optional). Prefer these over username/password when set.
    # Provide either base64 APP_VECTORSTORE_APIKEY or split ID/SECRET.
    APP_VECTORSTORE_APIKEY_ID: ""
    APP_VECTORSTORE_APIKEY_SECRET: ""
    APP_VECTORSTORE_APIKEY: ""
    COLLECTION_NAME: "multimodal_data"

    # === MinIO Configurations ===
    MINIO_ENDPOINT: "rag-minio:9000"
    MINIO_ACCESSKEY: "minioadmin"
    MINIO_SECRETKEY: "minioadmin"

    # === Authentication ===
    ## Service-specific API keys are managed via apiKeysSecret section (see top-level config).
    ## APP_EMBEDDINGS_APIKEY and SUMMARY_LLM_APIKEY are loaded from secrets automatically.

    # === Embeddings Configurations ===
    APP_EMBEDDINGS_SERVERURL: "nemoretriever-embedding-ms:8000/v1"
    APP_EMBEDDINGS_MODELNAME: "nvidia/llama-3.2-nv-embedqa-1b-v2"
    APP_EMBEDDINGS_DIMENSIONS: "2048"

    # === NV-Ingest Configurations ===
    APP_NVINGEST_MESSAGECLIENTHOSTNAME: "rag-nv-ingest"
    APP_NVINGEST_MESSAGECLIENTPORT: "7670"

    # === NV-Ingest extraction configurations ===
    APP_NVINGEST_PDFEXTRACTMETHOD: "None"  # Method used for text extraction from "None", "pdfium", "nemotron_parse"
    APP_NVINGEST_EXTRACTTEXT: "True"  # Enable text extraction
    APP_NVINGEST_EXTRACTINFOGRAPHICS: "False"  # Enable infographic extraction
    APP_NVINGEST_EXTRACTTABLES: "True"  # Enable table extraction
    APP_NVINGEST_EXTRACTCHARTS: "True"  # Enable chart extraction
    APP_NVINGEST_EXTRACTIMAGES: "False"  # Enable image extraction
    APP_NVINGEST_EXTRACTPAGEASIMAGE: "False"  # Extracts each page as image if enabled
    APP_NVINGEST_STRUCTURED_ELEMENTS_MODALITY: ""  # "image", "text_image"
    APP_NVINGEST_IMAGE_ELEMENTS_MODALITY: ""  # "image"
    APP_NVINGEST_TEXTDEPTH: "page"  # Extract text by "page" or "document"

    # === NV-Ingest caption configurations ===
    APP_NVINGEST_CAPTIONMODELNAME: "nvidia/nemotron-nano-12b-v2-vl"  # Model name for captioning
    APP_NVINGEST_CAPTIONENDPOINTURL: ""  # Endpoint URL for captioning model

    # === NV-Ingest save to disk configurations ===
    APP_NVINGEST_SAVETODISK: "False"
    NVINGEST_MINIO_BUCKET: "nv-ingest" # If this value is modified, ensure the corresponding Milvus bucketName is also updated

    # === NV-Ingest performance configurations ===
    APP_NVINGEST_ENABLE_PDF_SPLIT_PROCESSING: "False"
    APP_NVINGEST_PAGES_PER_CHUNK: "16"

    # === General ===
    # Summary Model Configurations
    SUMMARY_LLM: "nvidia/llama-3.3-nemotron-super-49b-v1.5"
    SUMMARY_LLM_SERVERURL: "nim-llm:8000"
    SUMMARY_LLM_MAX_CHUNK_LENGTH: "9000"
    SUMMARY_CHUNK_OVERLAP: "400"
    SUMMARY_LLM_TEMPERATURE: "0.0"
    SUMMARY_LLM_TOP_P: "1.0"
    SUMMARY_MAX_PARALLELIZATION: "20"

    # === General ===
    ENABLE_CITATIONS: "True"
    LOGLEVEL: "INFO"

    # === NV-Ingest splitting configurations ===
    APP_NVINGEST_CHUNKSIZE: "512"  # Size of chunks for splitting
    APP_NVINGEST_CHUNKOVERLAP: "150"  # Overlap size for chunks
    APP_NVINGEST_ENABLEPDFSPLITTER: "True"  # Enable PDF splitter
    APP_NVINGEST_SEGMENTAUDIO: "False"  # Enable audio segmentation for NV Ingest

    # === Redis configurations ===
    REDIS_HOST: "rag-redis-master"
    REDIS_PORT: "6379"
    REDIS_DB: "0"
    ENABLE_REDIS_BACKEND: "False"

    # === Bulk upload to MinIO ===
    ENABLE_MINIO_BULK_UPLOAD: "True"
    TEMP_DIR: "/tmp-data"
    INGESTOR_SERVER_DATA_DIR: "/data/"

    # === NV-Ingest Batch Mode Configurations ===
    NV_INGEST_FILES_PER_BATCH: "16"
    NV_INGEST_CONCURRENT_BATCHES: "4"
    ENABLE_NV_INGEST_DYNAMIC_BATCHING: "True"

    # === Tracing ===
    APP_TRACING_ENABLED: "False"
    # HTTP endpoint
    APP_TRACING_OTLPHTTPENDPOINT: "http://otel-collector:4318/v1/traces"
    # GRPC endpoint
    APP_TRACING_OTLPGRPCENDPOINT: "grpc://otel-collector:4317"

  # -- Persistent storage for ingestor-server data directory
  # If enabled, mounts a PVC at the path used by INGESTOR_SERVER_DATA_DIR
  persistence:
    enabled: true
    # If set, use an existing PVC by name; when empty and enabled, a PVC is created
    existingClaim: ""
    # StorageClass to use for the PVC; keep empty to use the cluster default
    storageClass: ""
    # Access modes for the PVC
    accessModes:
      - ReadWriteOnce
    # Requested size for the PVC
    size: 50Gi
    # Mount path inside the container; defaults to envVars.INGESTOR_SERVER_DATA_DIR if empty
    mountPath: "/data/"
    # Optional subPath within the PVC
    subPath: ""

# -- Frontend
# subsection: frontend
# rag frontend Frontend
frontend:
  enabled: true
  appName: "rag-frontend"

  replicaCount: 1

  image:
    repository: nvcr.io/nvidia/blueprint/rag-frontend
    pullPolicy: IfNotPresent
    tag: "2.4.0"

  imagePullSecret:
    name: "ngc-secret"
    registry: "nvcr.io"
    username: "$oauthtoken"
    password: ""

  service:
    type: NodePort
    port: 3000

  # -- Probes for frontend (optional)
  livenessProbe: {}
  readinessProbe: {}

  envVars:
    # Runtime environment variables for Vite frontend
    # Note: Model names are now managed by frontend settings store
    - name: VITE_API_CHAT_URL
      value: "http://rag-server:8081/v1"
    - name: VITE_API_VDB_URL
      value: "http://ingestor-server:8082/v1"
    - name: VITE_MILVUS_URL
      value: "http://milvus:19530"

# -- Elasticsearch dependency toggle
eck-elasticsearch:
  enabled: false
  http:
    tls:
      selfSignedCertificate:
        disabled: true
  nodeSets:
  - name: default
    count: 1
    config:
      node.store.allow_mmap: false
      xpack.security.enabled: false
      xpack.security.http.ssl.enabled: false
      xpack.security.transport.ssl.enabled: false
    podTemplate:
      spec:
        containers:
        - name: elasticsearch
          # Override readiness probe to not require authentication
          readinessProbe:
            exec:
              command:
              - bash
              - -c
              - |
                curl -s http://localhost:9200/_cluster/health | grep -q '"status":"green"\|"status":"yellow"'
            initialDelaySeconds: 10
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 3

# -- Observability
# subsection: serviceMonitor
serviceMonitor:
  enabled: false

# subsection: opentelemetry-collector
opentelemetry-collector:
  enabled: false
  mode: deployment
  image:
    repository: docker.io/otel/opentelemetry-collector-contrib
    tag: "0.140.0"
  command:
    name: otelcol-contrib
  config:
    receivers:
      otlp:
        protocols:
          grpc:
            endpoint: '${env:MY_POD_IP}:4317'
          http:
            cors:
              allowed_origins:
                - "*"
    exporters:
      # NOTE: Prior to v0.86.0 use `logging` instead of `debug`.
      zipkin:
        endpoint: "http://rag-zipkin:9411/api/v2/spans"
      debug:
        verbosity: detailed
      prometheus:
        endpoint: ${env:MY_POD_IP}:8889
    extensions:
      health_check: {}
      zpages:
        endpoint: 0.0.0.0:55679
    processors:
      batch: {}
      tail_sampling:
        # filter out health checks
        # https://github.com/open-telemetry/opentelemetry-collector/issues/2310#issuecomment-1268157484
        policies:
          - name: drop_noisy_traces_url
            type: string_attribute
            string_attribute:
              key: http.target
              values:
                - \/health
              enabled_regex_matching: true
              invert_match: true
      transform:
        trace_statements:
          - context: span
            statements:
              - set(status.code, 1) where attributes["http.path"] == "/health"

              # after the http target has been anonymized, replace other aspects of the span
              - replace_match(attributes["http.route"], "/v1", attributes["http.target"]) where attributes["http.target"] != nil

              # replace the title of the span with the route to be more descriptive
              - replace_pattern(name, "/v1", attributes["http.route"]) where attributes["http.route"] != nil

              # set the route to equal the URL if it's nondescriptive (for the embedding case)
              - set(name, Concat([name, attributes["http.url"]], " ")) where name == "POST"
    service:
      extensions: [zpages, health_check]
      pipelines:
        traces:
          receivers: [otlp]
          exporters: [debug, zipkin]
          processors: [tail_sampling, transform]
        metrics:
          exporters:
            - debug
            - prometheus
          processors:
            - memory_limiter
            - batch
          receivers:
            - otlp
            - prometheus
        logs:
          receivers: [otlp]
          exporters: [debug]
          processors: [batch]
  ports:
    metrics:
      enabled: true
      containerPort: 8889
      servicePort: 8889
      protocol: TCP

# subsection: zipkin
zipkin:
  enabled: false
  image:
    repository: docker.io/openzipkin/zipkin
    tag: 3.5.0

# subsection: kube-prometheus-stack
kube-prometheus-stack:
  enabled: false
  prometheus-node-exporter:
    service:
      port: 9100
      targetPort: 9100
    hostNetwork: false
  prometheus:
    serviceMonitor:
      interval: "1s"
    prometheusSpec:
      scrapeInterval: "1s"
      evaluationInterval: "1s"
      additionalScrapeConfigs:
        - job_name: rag-app
          scrape_interval: 5s
          metrics_path: /metrics
          static_configs:
            - targets:
                # If the app is in another namespace, use rag-server.<namespace>.svc:8081
                - rag-server:8081
  grafana:
    adminUser: admin
    adminPassword: "admin"
  prometheusOperator:
    crds:
      enabled: false

# -- NIMs (dependencies) configuration
# subsection: nim-llm
# NIM LLM
nimOperator:       
  draResources:
    enabled: false
    name: "rag-claim"
    type: "TimeSlicing"
  nim-llm:
    enabled: true
    replicas: 1
    service:
      name: "nim-llm"
    image:
      repository: nvcr.io/nim/nvidia/llama-3.3-nemotron-super-49b-v1.5
      pullPolicy: IfNotPresent
      tag: "1.14.0"
    resources:
      limits:
        nvidia.com/gpu: 1
      requests:
        nvidia.com/gpu: 1
    nodeSelector: {}
    tolerations: []
    model:
      engine: tensorrt_llm
# -- Uncomment this section to enable FP8 precision and throughput profile for the NIM LLM, change the GPU product according to the GPU you are using
#      precision: "fp8"
#      qosProfile: "throughput"
#      tensorParallelism: "1"
#      gpus:
#        - product: "rtx6000_blackwell_sv"
    storage:
      pvc:
        create: true
        size: "120Gi"
        volumeAccessMode: ReadWriteOnce
        storageClass: ""
      sharedMemorySizeLimit: "16Gi"
#    draResources:
#      - claimCreationSpec:
#          devices:
#            - name: nimllm-gpu
#              count: 1
#              deviceClassName: gpu.nvidia.com
#              driverName: gpu.nvidia.com
#              attributeSelectors:
#                - key: driverVersion
#                  op: GreaterThanOrEqual
#                  value:
#                    versionValue: "550.127.8"
#              capacitySelectors:
#                - key: memory
#                  op: GreaterThanOrEqual
#                  value: 40Gi
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
      - name: NIM_SERVED_MODEL_NAME
        value: "nvidia/llama-3.3-nemotron-super-49b-v1.5"
#      - name: CUDA_VISIBLE_DEVICES
#        value: "0"
    expose:
      service:
        name: http
        type: ClusterIP
        port: 8000
        grpcPort: 8001
# subsection: nvidia-nim-llama-32-nv-embedqa-1b-v2
# NIM Text Embedding
  nvidia-nim-llama-32-nv-embedqa-1b-v2:
    enabled: true
    replicas: 1
    service:
      name: "nemoretriever-embedding-ms"
    image:
      repository: nvcr.io/nim/nvidia/llama-3.2-nv-embedqa-1b-v2
      tag: "1.10.1"
      pullPolicy: IfNotPresent
    resources:
      limits:
        nvidia.com/gpu: 1
      requests:
        nvidia.com/gpu: 1
    storage:
      pvc:
        create: true
        size: "50Gi"
        volumeAccessMode: ReadWriteOnce
        storageClass: ""
    nodeSelector: {}
    tolerations: []
#    draResources:
#      - resourceClaimName: rag-claim
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
      - name: OMP_NUM_THREADS
        value: "1"
      - name: NIM_TRITON_PERFORMANCE_MODE
        value: "throughput"
    expose:
      service:
        name: http
        type: ClusterIP
        port: 8000
        grpcPort: 8001

# subsection: nvidia-nim-llama-32-nemoretriever-1b-vlm-embed-v1
# NIM VLM Embedding
  nvidia-nim-llama-32-nemoretriever-1b-vlm-embed-v1:
    enabled: false
    service:
      name: "nemoretriever-vlm-embedding-ms"
    image:
      repository: nvcr.io/nvidia/nemo-microservices/llama-3.2-nemoretriever-1b-vlm-embed-v1
      tag: "1.7.0"
      pullPolicy: IfNotPresent
    replicas: 1
    resources:
      limits:
        nvidia.com/gpu: 1
      requests:
        nvidia.com/gpu: 1
    nodeSelector: {}
    tolerations: []
#    draResources:
#      - resourceClaimName: rag-claim
    storage:
      pvc:
        create: true
        size: "50Gi"
        volumeAccessMode: ReadWriteOnce
        storageClass: ""
    env:
      - name: NIM_HTTP_API_PORT
        value: "8000"
      - name: NIM_TRITON_LOG_VERBOSE
        value: "1"
    expose:
      service:
        name: http
        type: ClusterIP
        port: 8000
        grpcPort: 8001
# subsection: text-reranking-nim
# NIM Text Reranking
  nvidia-nim-llama-32-nv-rerankqa-1b-v2:
    enabled: true
    replicas: 1
    service:
      name: "nemoretriever-ranking-ms"
    image:
      repository: nvcr.io/nim/nvidia/llama-3.2-nv-rerankqa-1b-v2
      tag: "1.8.0"
      pullPolicy: IfNotPresent
    resources:
      limits:
        nvidia.com/gpu: 1
      requests:
        nvidia.com/gpu: 1
    nodeSelector: {}
    tolerations: []
#    draResources:
#      - resourceClaimName: rag-claim
    storage:
      pvc:
        create: true
        size: "50Gi"
        volumeAccessMode: ReadWriteOnce
        storageClass: ""
    env: []
    expose:
      service:
        name: http
        type: ClusterIP
        port: 8000
        grpcPort: 8001

  # subsection: nim-vlm
  # NIM Vision-Language (VLM)
  nim-vlm:
    enabled: false
    replicas: 1
    service:
      name: "nim-vlm"
    image:
      repository: nvcr.io/nim/nvidia/nemotron-nano-12b-v2-vl
      tag: "1.5.0"
      pullPolicy: IfNotPresent
    resources:
      limits:
        nvidia.com/gpu: 1
      requests:
        nvidia.com/gpu: 1
    nodeSelector: {}
    tolerations: []
#    draResources:
#      - resourceClaimName: rag-claim
    storage:
      pvc:
        create: true
        size: "50Gi"
        volumeAccessMode: ReadWriteOnce
        storageClass: ""
    env: []
    expose:
      service:
        name: http
        type: ClusterIP
        port: 8000
        grpcPort: 8001

# -- NV-Ingest dependency configuration
# subsection: nv-ingest
# NV-Ingest Service
nv-ingest:
  enabled: true
  imagePullSecrets:
    - name: "ngc-secret"
  ngcApiSecret:
    create: false
  ngcImagePullSecret:
    create: false
  image:
    repository: "nvcr.io/nvidia/nemo-microservices/nv-ingest"
    tag: "26.1.1"
  resources:
    limits:
      nvidia.com/gpu: 0
  envVars:
    ARROW_DEFAULT_MEMORY_POOL: "system"
    OMP_NUM_THREADS: "1"
    INGEST_LOG_LEVEL: WARNING
    INGEST_RAY_LOG_LEVEL: "PRODUCTION"
    INGEST_DYNAMIC_MEMORY_THRESHOLD: "0.80"
    INGEST_DISABLE_DYNAMIC_SCALING: "True"
    NV_INGEST_MAX_UTIL: 48
    INGEST_EDGE_BUFFER_SIZE: 64
    MRC_IGNORE_NUMA_CHECK: 1
    READY_CHECK_ALL_COMPONENTS: "False"
    REDIS_MORPHEUS_TASK_QUEUE: morpheus_task_queue
    REDIS_INGEST_TASK_QUEUE: "ingest_task_queue"
    NV_INGEST_DEFAULT_TIMEOUT_MS: "1234"
    MAX_INGEST_PROCESS_WORKERS: 16
    VLM_CAPTION_PROMPT: "Caption the content of this image"
    
    # Ray internals configuration
    RAY_num_grpc_threads: "1"
    RAY_num_server_call_thread: "1"
    RAY_worker_num_grpc_internal_threads: "1"
    
    EMBEDDING_NIM_ENDPOINT: "http://nemoretriever-embedding-ms:8000/v1"
    EMBEDDING_NIM_MODEL_NAME: "nvidia/llama-3.2-nv-embedqa-1b-v2"
    MESSAGE_CLIENT_HOST: "rag-redis-master"
    MESSAGE_CLIENT_PORT: 6379
    MESSAGE_CLIENT_TYPE: "redis"
    MINIO_INTERNAL_ADDRESS: "rag-minio:9000"
    MINIO_PUBLIC_ADDRESS: "http://localhost:9000"
    MINIO_BUCKET: "nv-ingest"
    MINIO_ACCESS_KEY: "minioadmin"
    MINIO_SECRET_KEY: "minioadmin"
    MILVUS_ENDPOINT: "http://milvus:19530"
    OTEL_EXPORTER_OTLP_ENDPOINT: "otel-collector:4317"
    MODEL_PREDOWNLOAD_PATH: "/workspace/models/"
    INSTALL_AUDIO_EXTRACTION_DEPS: "true"
    COMPONENTS_TO_READY_CHECK: "ALL"

    # OCR routing (defaults to NeMo Retriever OCR service)
    OCR_GRPC_ENDPOINT: nemoretriever-ocr-v1:8001
    OCR_HTTP_ENDPOINT: http://nemoretriever-ocr-v1:8000/v1/infer
    OCR_INFER_PROTOCOL: grpc
    OCR_MODEL_NAME: scene_text_ensemble

    # NeMo Retriever Parse (VLM text extraction)
    NEMOTRON_PARSE_HTTP_ENDPOINT: http://nemotron-parse:8000/v1/chat/completions
    NEMOTRON_PARSE_INFER_PROTOCOL: http
    NEMOTRON_PARSE_MODEL_NAME: nvidia/nemotron-parse

    # PDF splitting configuration
    PDF_SPLIT_PAGE_COUNT: "32"

    # YOLOX endpoints
    YOLOX_PAGE_IMAGE_FORMAT: "JPEG"
    YOLOX_GRPC_ENDPOINT: nemoretriever-page-elements-v3:8001
    YOLOX_HTTP_ENDPOINT: http://nemoretriever-page-elements-v3:8000/v1/infer
    YOLOX_INFER_PROTOCOL: grpc
    YOLOX_GRAPHIC_ELEMENTS_GRPC_ENDPOINT: nemoretriever-graphic-elements-v1:8001
    YOLOX_GRAPHIC_ELEMENTS_HTTP_ENDPOINT: http://nemoretriever-graphic-elements-v1:8000/v1/infer
    YOLOX_GRAPHIC_ELEMENTS_INFER_PROTOCOL: grpc
    YOLOX_TABLE_STRUCTURE_GRPC_ENDPOINT: nemoretriever-table-structure-v1:8001
    YOLOX_TABLE_STRUCTURE_HTTP_ENDPOINT: http://nemoretriever-table-structure-v1:8000/v1/infer
    YOLOX_TABLE_STRUCTURE_INFER_PROTOCOL: grpc

    # Captioning
    VLM_CAPTION_MODEL_NAME: nvidia/nemotron-nano-12b-v2-vl
    VLM_CAPTION_ENDPOINT: http://nim-vlm:8000/v1/chat/completions

    # Audio service
    AUDIO_GRPC_ENDPOINT: nv-ingest-riva-nim:50051
    AUDIO_INFER_PROTOCOL: grpc

  # Expose internal Milvus/MinIO config managed by nv-ingest subchart
  milvusDeployed: true
  milvus:
    # Uncomment this section to enable authentication for Milvus
    # extraConfigFiles:
    #   user.yaml: |+
    #     common:
    #       security:
    #         authorizationEnabled: true
    #         defaultRootPassword: Milvus
    image:
      all:
        repository: docker.io/milvusdb/milvus
        tag: v2.6.5-gpu
      tools:
        repository: docker.io/milvusdb/milvus-config-tool
        tag: v0.1.2
        pullPolicy: IfNotPresent
    etcd:
      image:
        repository: milvusdb/etcd
        tag: "3.5.23-r2"
    standalone:
      resources:
        limits:
          nvidia.com/gpu: 1
    minio:
      image:
        repository: docker.io/minio/minio
        tag: "RELEASE.2025-09-07T16-13-09Z"
      accessKey: minioadmin
      secretKey: minioadmin
      bucketName: nv-ingest
    fullnameOverride: milvus

  # Redis Master
  redis:
    image:
      repository: redis
      tag: 8.2.1


  # Ensure nv-ingest does not deploy its own observability components
  otelDeployed: false
  zipkinDeployed: false

  nimOperator:
    # Ensure nv-ingest does not deploy its own embedding NIM
    embedqa:
      enabled: false

    # NeMo Retriever OCR (default)
    # To use PaddleOCR instead, override the image:
    # nemoretriever_ocr_v1:
    #   enabled: true
    #   image:
    #     repository: nvcr.io/nim/baidu/paddleocr
    #     tag: 1.5.0
    # And set OCR_MODEL_NAME to "paddle" in envVars above
    nemoretriever_ocr_v1:
      enabled: true
      storage:
        pvc:
          volumeAccessMode: "ReadWriteOnce"
          storageClass: ""
      tolerations: []
      replicaCount: 1
      image:
        repository: nvcr.io/nim/nvidia/nemoretriever-ocr-v1
        tag: "1.2.0"
      imagePullSecrets:
        - name: ngc-secret
      env:
        - name: OMP_NUM_THREADS
          value: "8"
        - name: NIM_HTTP_API_PORT
          value: "8000"
        - name: NIM_TRITON_LOG_VERBOSE
          value: "1"
        - name: CUDA_VISIBLE_DEVICES
          value: "0"
        - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
          value: "8192"
        - name: NIM_TRITON_MAX_BATCH_SIZE
          value: "32"
        - name: NIM_TRITON_ENABLE_MODEL_CONTROL
          value: "1"
      resources:
        limits:
          nvidia.com/gpu: 1
        requests:
          nvidia.com/gpu: 1

    # NIM Graphic Elements
    graphic_elements:
      enabled: true
      storage:
        pvc:
          volumeAccessMode: "ReadWriteOnce"
          storageClass: ""
      tolerations: []
      replicaCount: 1
      image:
        repository: nvcr.io/nim/nvidia/nemoretriever-graphic-elements-v1
        tag: "1.6.0"
      env:
        - name: NIM_HTTP_API_PORT
          value: "8000"
        - name: NIM_TRITON_LOG_VERBOSE
          value: "1"
        - name: NIM_TRITON_RATE_LIMIT
          value: "3"
        - name: CUDA_VISIBLE_DEVICES
          value: "0"
        - name: NIM_TRITON_MAX_BATCH_SIZE
          value: "32"
        - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
          value: "2048"
        - name: OMP_NUM_THREADS
          value: "1"
      resources:
        limits:
          nvidia.com/gpu: 1
        requests:
          nvidia.com/gpu: 1

    # NIM Page Elements
    page_elements:
      enabled: true
      storage:
        pvc:
          volumeAccessMode: "ReadWriteOnce"
          storageClass: ""
      tolerations: []
      replicaCount: 1
      image:
        repository: nvcr.io/nim/nvidia/nemoretriever-page-elements-v3
        tag: "1.7.0"
      env:
        - name: NIM_HTTP_API_PORT
          value: "8000"
        - name: NIM_TRITON_LOG_VERBOSE
          value: "1"
        - name: NIM_TRITON_RATE_LIMIT
          value: "3"
        - name: CUDA_VISIBLE_DEVICES
          value: "0"
        - name: NIM_TRITON_MAX_BATCH_SIZE
          value: "32"
        - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
          value: "2048"
        - name: NIM_TRITON_CPU_THREADS_PRE_PROCESSOR
          value: "2"
        - name: NIM_TRITON_CPU_THREADS_POST_PROCESSOR
          value: "1"
        - name: OMP_NUM_THREADS
          value: "2"
        - name: NIM_ENABLE_OTEL
          value: "true"
        - name: NIM_OTEL_SERVICE_NAME
          value: "page-elements"
        - name: NIM_OTEL_TRACES_EXPORTER
          value: "otlp"
        - name: NIM_OTEL_METRICS_EXPORTER
          value: "console"
        - name: NIM_OTEL_EXPORTER_OTLP_ENDPOINT
          value: "http://otel-collector:4318"
        - name: TRITON_OTEL_URL
          value: "http://otel-collector:4318/v1/traces"
        - name: TRITON_OTEL_RATE
          value: "1"
      resources:
        limits:
          nvidia.com/gpu: 1
        requests:
          nvidia.com/gpu: 1

    # NIM Table Structure
    table_structure:
      enabled: true
      storage:
        pvc:
          storageClass: ""
          volumeAccessMode: "ReadWriteOnce"
      tolerations: []
      replicaCount: 1
      image:
        repository: nvcr.io/nim/nvidia/nemoretriever-table-structure-v1
        tag: "1.6.0"
      env:
        - name: NIM_HTTP_API_PORT
          value: "8000"
        - name: NIM_TRITON_LOG_VERBOSE
          value: "1"
        - name: NIM_TRITON_RATE_LIMIT
          value: "3"
        - name: CUDA_VISIBLE_DEVICES
          value: "0"
        - name: NIM_TRITON_MAX_BATCH_SIZE
          value: "32"
        - name: NIM_TRITON_CUDA_MEMORY_POOL_MB
          value: "2048"
        - name: OMP_NUM_THREADS
          value: "1"
      resources:
        limits:
          nvidia.com/gpu: 1
        requests:
          nvidia.com/gpu: 1

    # Nemotron Parse (VLM text extraction)
    nemotron_parse:
      enabled: false
      storage:
        pvc:
          storageClass: ""
          volumeAccessMode: "ReadWriteOnce"
      nodeSelector: {}
      tolerations: []
      replicaCount: 1
      image:
        repository: nvcr.io/nim/nvidia/nemotron-parse
        tag: "1.5.0"
      env:
        - name: NIM_HTTP_API_PORT
          value: "8000"
        - name: NIM_TRITON_LOG_VERBOSE
          value: "1"
      resources:
        limits:
          nvidia.com/gpu: 1
        requests:
          nvidia.com/gpu: 1

    # Riva NIM for Audio Transcription
    audio:
      enabled: false
      authSecret: ngc-api
      storage:
        pvc:
          create: true
          storageClass: ""
          size: "25Gi"
          volumeAccessMode: "ReadWriteOnce"
      nodeSelector: {}
      tolerations: []
      replicas: 1
      image:
        repository: nvcr.io/nim/nvidia/parakeet-1-1b-ctc-en-us
        tag: "1.4.0"
        pullSecrets:
          - ngc-secret
      env:
        - name: NIM_TAGS_SELECTOR
          value: "name=parakeet-1-1b-ctc-en-us,mode=ofl"
        - name: NIM_TRITON_LOG_VERBOSE
          value: "1"
      expose:
        service:
          type: ClusterIP
          name: http
          port: 9000
          grpcPort: 50051
      resources:
        limits:
          nvidia.com/gpu: 1
        requests:
          nvidia.com/gpu: 1