Skip to main content

Configuration reference

You configure VoisX Orchestrate through the Helm chart's values. Keep your settings in your own values file, for example orchestrate-values.yaml, and pass it on every install and change. This page covers the values a customer sets. Leave anything not listed here at its default.

Many features are off by default and stay off until you set their values: code agents, channels, the SIP gateway, hybrid sites, multi-tenancy, PostgreSQL high availability, backups and metrics. With them off, the chart renders nothing for them.

Apply a change​

Edit your values file, then apply it with voisctl from the bundle of the version you run:

./voisctl upgrade --bundle ./bundle --keys ./keys \
--registry registry.example.com \
--registry-user <push-account> --registry-password <password> \
--kube-context <your-context> --namespace vois-orchestrate \
-f orchestrate-values.yaml --force

voisctl keeps the values already on the install, including the credentials it generated, applies your file on top, pins every image to its signed digest again, and checks the result.

You can also use Helm directly with the chart from the same bundle:

helm upgrade vois-orchestrate ./bundle/chart/vois-orchestrate-<version>.tgz \
--kube-context <your-context> --namespace vois-orchestrate \
--reset-then-reuse-values -f orchestrate-values.yaml --wait
Never run a bare helm upgrade

A helm upgrade without --reset-then-reuse-values and without your full values file resets every value to the chart default. That empties your secrets and image pins. Services then fail authentication, often well after the upgrade reports success. Always keep the previous values, or pass your complete values file.

Security guard​

The chart refuses to install or upgrade while a required credential is empty, or while a well-known default password is still in place, and lists each one. security.allowInsecureDefaults: true turns the guard off. It is for development only.

Images and registries​

ValueDefaultWhat to set
image.registryempty, requiredYour registry host, for example registry.example.com.
image.repositoryvois-orchestrateYour registry project.
image.tagemptyEmpty means the release the chart came from.
image.digestsemptyLeave empty. voisctl sets a digest for each Orchestrate image.
image.pullPolicyIfNotPresentAlways or IfNotPresent.
imagePullSecretsemptyNames of docker-registry Secrets in the namespace, for a private registry project.
postgres.image, temporal.image, temporal.uiImage, opa.image, livekit.image, keycloak.image, qdrant.image, redis.image, minio.imagepublic image namesThe copies in your registry. voisctl sets and pins them. See Install.

Model registry​

Models you deploy from the catalog run from images in your registry.

ValueDefaultWhat to set
modelRegistry.urlemptyRegistry host, or host:port, with no scheme. It must be an address your nodes can pull from. Until you set it, model builds fail with a message that names it.
modelRegistry.projectvois-orchestrateThe registry project for model images.
modelRegistry.imagePullSecretsemptyPull Secret names for model images, if the project is private.
modelRegistry.apiUrlemptyThe address the deployment controller uses to look up image digests. Empty means https://<modelRegistry.url>.
modelRegistry.caCertSecretemptyA Secret with key ca.crt, for a registry whose certificate comes from your own CA.
modelRegistry.credentialsSecretemptyA pull Secret the deployment controller uses to look up image digests in a private model project. Needed once the project stops answering anonymous pulls.
modelRegistry.purgeCredentialsSecretemptyA registry account that may list and delete repositories, so that purging an organization also removes its images. Without it the purge leaves them and names each one.
modelRegistry.scanner.urlemptyYour registry's vulnerability scanner (Trivy) adapter. The System page reads the age of its database from it. Empty: the page says scanning isn't configured.
modelRegistry.scanner.maxDbAgeDays30A scanner database older than this is shown as stale.
modelIntegrity.enforcementwarnwarn deploys model images that have no build record from this platform and records them. enforce refuses them. An image whose digest doesn't match what this platform built is refused either way.
runtimeBases.registry, runtimeBases.tagemptyWhere model builds take their prebuilt base images from. voisctl sets these to the copies the bundle carries, so models from the local model store build with no internet.
buildImages.kaniko, buildImages.busybox, buildImages.mcemptyThe build tool images. voisctl sets them to the bundle's copies.

Secrets​

voisctl install generates every credential and keeps them in the Secret vois-orchestrate-install-values. With plain Helm, voisctl secrets generate --out values-secrets.yaml writes the same file. Set these yourself only if you manage credentials another way:

ValueUsed for
postgres.password, postgres.appPassword, postgres.keycloakPassword, postgres.auditPasswordThe bundled PostgreSQL. Applied only when PostgreSQL first creates its data. Changing them later breaks the services' connections.
keycloak.adminPasswordKeycloak's own administrator.
keycloak.consoleAdminPasswordThe first console user, admin.
keycloak.apiClientSecretThe API's client in Keycloak, used to search users, send invitations and manage identity providers.
orchestrationWorker.apiKey, realtimeAgentWorker.apiKey, ragIngestion.apiKey, eval.apiKeyThe workers' keys for the Model Gateway. Each must start with vcpk_.
internalAuth.ragIngestionSearchKey, internalAuth.auditReadKey, internalAuth.redisPassword, internalAuth.qdrantApiKeyCredentials between services inside the cluster.
minio.accessKey, minio.secretKeyMinIO, which holds documents, model files and recordings.
livekit.apiKey, livekit.apiSecretThe LiveKit key pair for voice sessions.
web.nextAuthSecretSigns console sessions. If it is empty, every upgrade signs everyone out.

Some credentials are generated by the chart on first install and kept across upgrades, so you don't set them: the policy bundle token, the code agent ingress key, the Temporal payload key, the audit keys, the site CA and the SIP gateway's key.

Rotating a worker key. Change its value and upgrade. On start-up the API rotates the stored key in place and keeps the old one working for 60 minutes while the workers restart. A value without the vcpk_ prefix is refused with an error in the log.

Internal database credentials

PostgreSQL is never published outside the cluster. Limit who can run kubectl exec and read Secrets in the vois-orchestrate namespace.

Cluster platform​

ValueDefaultWhat it does
platform.openshiftfalsetrue on OpenShift: restricted pod security, no fixed user ids, OpenShift Routes for exposure, and dedicated service accounts for the few pods that need wider SCCs.
podSecurity.restrictedfalseRuns every pod under the Kubernetes "restricted" Pod Security Standard. Recommended for every new install. On an existing install, change the ownership of each data volume first. LiveKit uses the host network, which the restricted standard forbids.
podSecurity.runAsUser10001The user id every pod runs as, when restricted.
persistence.storageClassNameemptyThe StorageClass for every volume. Empty uses the cluster default. - binds to pre-created volumes. Each store can override it with <store>.storageClassName.

See the platform guides that come with the bundle for EKS, GKE, OpenShift and other clusters.

Exposure​

Nothing is reachable from outside the cluster unless you turn it on. Reaching an install that publishes nothing means kubectl port-forward.

ValueDefaultWhat it does
exposure.modenonenone, ingress, route (OpenShift), loadBalancer or nodePort.
exposure.classNameemptyingress mode: the IngressClass. Empty uses your cluster's default class.
exposure.loadBalancerIPemptyloadBalancer mode: a pre-assigned address, for MetalLB or a hardware balancer.
exposure.annotationsemptyAnnotations added to everything the exposure settings create.
exposure.tls.modecertManageringress and route modes: certManager, secret or none. See TLS.
exposure.tls.clusterIssueremptycertManager mode: the name of your ClusterIssuer. Required in that mode.
exposure.tls.secretNameemptysecret mode: the kubernetes.io/tls Secret you created. Required in that mode.

Modes​

ModeUse it whenResult
ingressYour cluster runs an ingress controller.An Ingress per component, on its hostname, with TLS.
routeOpenShift (platform.openshift: true).An OpenShift Route per component, with TLS.
loadBalancerYou have MetalLB or a hardware load balancer.An extra Service of type LoadBalancer per component, named <component>-external. Plain HTTP.
nodePortYou have neither.An extra NodePort Service per component. Plain HTTP.
noneDefault.Nothing is published.

In loadBalancer and nodePort modes, set each component's host to the address or DNS name people use to reach it. In nodePort mode, also set a fixed nodePort (30000 to 32767) for web and for api.

Components you can expose​

Each component takes enabled, host, nodePort, loadBalancerIP and annotations under exposure.components.<name>.

ComponentWhat it isNotes
webThe console.Its host becomes the console's URL and the API's allowed browser origin. Required in ingress mode.
apiThe Orchestrate API, including the Agent API, hosted chat, A2A and channel webhooks.The console calls it from the browser, so expose it whenever you expose web.
keycloakKeycloak's sign-in pages.Needed as soon as the console is exposed, because sign-in sends the browser there. In ingress mode only the sign-in paths are published, never Keycloak's admin console, and host can be the console's own host. In loadBalancer and nodePort modes the whole port is published, admin console included, so prefer ingress.
model-gatewayThe Model Gateway for your own applications and external agents.Off by default. See The Model Gateway address.
minioObject storage, for direct browser upload and download.The API already handles uploads and downloads, so most installs leave this off.
livekitLiveKit signalling, for voice sessions in a browser.See Voice (LiveKit).

Every other component can't be exposed. If you add any other name under exposure.components, the chart refuses to render and says why. This covers PostgreSQL, Redis, Qdrant, Open Policy Agent, Temporal and its web UI, the audit service, the knowledge-base search service, the License Service, every worker and the model predictors. Several of these have no authentication of their own. The SIP gateway and the site gateway have their own settings below.

The Model Gateway address​

The Model Gateway's external address is a switch, and it is off by default. Orchestrate is built to run air-gapped, so nothing outside the cluster can call the gateway until you turn it on.

Turn it on when applications or agents that run outside the cluster need to call models through Orchestrate: a team's own service, an agent on a developer's laptop, or a code agent framework running on other infrastructure. Leave it off when everything that calls models runs inside the cluster. Orchestrate's own agents and workers reach the gateway inside the cluster either way.

To turn it on, set the component and a hostname of your own:

exposure:
components:
model-gateway:
enabled: true
host: orchestrate-gateway.internal.example.com
  • The hostname can be internal-only. It only has to resolve and be reachable from wherever your applications run, such as your corporate network or a VPN. It doesn't need to be on the internet.
  • It uses the same exposure.mode and TLS settings as your other components. With tls.mode: secret, make sure the certificate also covers this hostname.
  • Every call still needs an API key. Personal keys and agent keys are both accepted, and requests without a valid key get 401.
  • When it is on, the console shows the address on API keys > Agent keys and puts it in the code snippets. When it is off, the console says that no external gateway address is set.
  • If the gateway is reachable at an address this chart didn't create, such as your own load balancer or API gateway, set modelGateway.publicUrl instead. The console then shows that address.

You can turn the switch on or off later with an upgrade. Turning it off removes the published address; keys stay valid for callers inside the cluster.

TLS​

exposure.tls.modeHow it works
secretYou create a kubernetes.io/tls Secret and name it in exposure.tls.secretName. One SAN or wildcard certificate can cover every hostname. Works on any network, including air-gapped ones.
certManagercert-manager issues a certificate for each hostname, through the ClusterIssuer in exposure.tls.clusterIssuer. Each component gets its own <component>-tls Secret. An ACME issuer needs internet access, so on an air-gapped network use a CA issuer.
nonePlain HTTP, and the console's URL starts with http://.

If a device in front of the cluster terminates TLS, none gives the console the wrong URL scheme. Contact us before you set it up that way.

Console​

ValueDefaultWhen to set it
web.nextAuthUrlhttp://localhost:3001Only if the console is not exposed. It is the address people use to reach the console, such as a port-forward.
web.apiUrlemptyOnly to override the API address the browser uses. Empty uses the exposed api host.
web.additionalOriginsemptyExtra browser origins the API accepts, besides the console's own URL.
web.temporalUiUrlemptyThe link the console shows for the Temporal web UI. Never point it at a published address: the Temporal web UI has no sign-in.

The console and API URLs have to agree with what the browser actually loads. If they don't, the console doesn't show an error. It shows pages whose panels are all empty.

Sign-in (Keycloak)​

ValueDefaultWhat it does
keycloak.publicUrlemptyThe Keycloak address a browser signs in at. Empty derives it from exposure.components.keycloak, or http://localhost:8180 for a port-forward. Set it only for an address this chart didn't create.
keycloak.sslRequiredexternalHTTPS for sign-in: external (except from private addresses), all or none. Applies when the realm is first created.
keycloak.realmSync.enabledtrueAfter every install and upgrade, a job brings an existing realm up to date: redirect addresses, the API's client, groups and events. Turn it off only if you manage the realm yourself.
keycloak.idpSecrets.enabledtrueKeeps identity providers' client secrets in a Kubernetes Secret that Keycloak reads as a vault. With it off, OpenID Connect providers with a client secret can't be saved from the console.
keycloak.adminConsoleUrlhttp://localhost:8180Where the console tells operators Keycloak's admin console is.

See Identity and single sign-on.

Multi-tenancy​

ValueDefaultWhat it does
tenancy.modesinglesingle: one organization per install. saas: many organizations on one install, isolated by PostgreSQL row-level security.
tenancy.operatorRoleplatform-operatorThe realm role of platform operators in SaaS mode.
tenancy.keycloakOrganizationsfalseCreate a Keycloak organization for each tenant. Needs Keycloak 26 with organizations enabled.

See Multi-tenancy.

Models and routing​

The Model Gateway routes each request to the model it names.

  • Catalog models. A model you deploy from the Models page is routed automatically once it is ready.
  • Static routes. modelGateway.selfHostedModels lists extra routes, each with a name and an apiBase, for an OpenAI-compatible server you run yourself. Empty by default.
modelGateway:
selfHostedModels:
- name: "llama3.1:8b"
apiBase: "http://llm01.internal.example.com:11434/v1"

Static routes are sent without an API key, win over a catalog model with the same name, and are treated as self-hosted. Point them only at servers on your own network.

ValueDefaultWhat it does
modelGateway.semanticCacheEmbeddingModelemptyAn embedding model the response cache uses to match similar prompts. Empty caches exact matches only.
modelGateway.visionModelsemptySelf-hosted models that accept images. A request with an image for any other self-hosted model is refused, rather than answered by a model that never saw it.
modelGateway.publicUrlemptyThe Model Gateway address shown with new agent keys, for agents outside the cluster. Empty derives it from the exposed model-gateway host, and shows no address when the gateway is not exposed. Display only. See The Model Gateway address.
modelServing.gpu.nodeSelector, modelServing.gpu.tolerationsemptyWhere GPU model deployments are scheduled. Empty tolerates the usual nvidia.com/gpu taint.
modelServing.gpu.memoryGiB24The memory of one GPU on your nodes. A deploy whose weights don't fit the chosen GPU size is refused, with the size that would fit.
modelServing.gpu.nodeInventorytrueLet the API read nodes and the namespace's ResourceQuota, so custom sizes and GPU types are checked against the cluster before anything is scheduled. Off, the scheduler decides.
modelServing.gpu.typeLabelnvidia.com/gpu.productThe node label that names each node's GPU model. The console offers only the types it finds.
modelServing.gpu.maxPerProject8The most GPUs one project's deployments may hold, counted at their maximum replicas. Platform admins aren't capped. 0 turns the cap off.
modelServing.vllm.variantempty (CUDA)cpu builds vLLM models from vLLM's CPU image, for an install without GPU nodes. See Sizes and GPUs.
modelServing.byo.proxyProjectsemptyFor your own serving containers: a map from a source registry to a pull-through proxy project in your registry, such as {"docker.io": "dockerhub-proxy"}.
gpuSharing.modenonetime-slicing or mig, to share GPUs between models. See GPU sharing.
gpuSharing.timeSlicing.replicas4Shares per physical GPU, with time-slicing.
gpuSharing.mig.strategymixedmixed (each slice its own resource) or single.
gpuSharing.devicePluginConfig.createfalseRender the NVIDIA device plugin's sharing config as a ConfigMap, for the plugin to load.
modelStore.urlImport.allowedHostsemptyHosts models may be downloaded from into the local model store: exact names, or *.example.com. Empty turns Add from a URL off. See Your own models.
modelStore.urlImport.maxBytes200 GiBThe largest file one URL import may download.
agentTimezoneemptyThe time zone agents are told the time in, when their organization sets none. Empty means Asia/Dubai.

Model and voice catalog​

ValueDefaultWhat it does
voiceCatalog.seedthe two baseline speech modelsVoices and speech models added to the catalog of a new install and built. [] adds none. Admins add more under Settings > Voice catalog.
modelCatalog.seedemptyChat and embedding models from the curated library to add and build on start-up.
modelCatalog.deployemptyCatalog models, of any kind, to deploy once built: shared, into the Default project, on this cluster.
modelCatalog.rebuildOnRuntimeChangeemptyRuntimes, such as ollama or vllm, whose built models are rebuilt automatically, one at a time, when the release changes that runtime's base image. Their running deployments move to the new image.

A model already in the catalog is never added twice.

External providers​

externalProviders.enabled (default false) is the ceiling for calls to AI providers outside your cluster. While it is false, no model call leaves the cluster, whatever the console says. With it true, admins connect providers on Models > External providers, a platform admin can switch external providers off and on under Settings, and an organization can turn them off for its own projects. Neither can go past this value. See External providers.

modelGateway.anthropicApiKey and modelGateway.openaiApiKey are used only when external providers are on. Leave them empty otherwise.

Guardrails​

modelGateway.guardrails sets the install's guardrail defaults, in the same shape as the chart's files/guardrails.example.json. A files/guardrails.json in the chart wins over it. Empty uses the built-in defaults.

modelGateway:
guardrails:
matrix:
completion: { credit_card: redact }
packs: [gulf, eu]
ignore_patterns: ["ORD-[0-9]{10}"]

Per organization, project, agent or version, use guardrail profiles in the console. See PII guardrails and Guardrail profiles.

Agent API and end-user chat​

ValueDefaultWhat it does
agentApi.webhookAllowedHostsemptyHosts an async agent run may call back with its webhook, comma-separated, with *.example.com for subdomains. Empty refuses every webhook. See Agent API.

Channels​

ValueDefaultWhat it does
channels.outboundHostsemptyHosts the API may call for Teams, Slack and WhatsApp, comma-separated. Empty means no channel can be connected and nothing is sent: the air-gapped setting.
channels.publicBaseUrlthe API's URLThe API's address as the vendors reach it. The webhook address is <this>/channels/<id>/events.
channels.secretBackendvault with code agents onWhere channel credentials go.
channels.simulatorfalseAllow simulator channels, for test installs only.

The hosts each vendor needs are listed in Channels. Behind a proxy, set HTTPS_PROXY and NO_PROXY on the API.

Code agents​

Code agents run your code, so they are off by default.

ValueDefaultWhat it does
agents.enabledfalseTurns on code agents, agent keys' A2A surface, the agent edge, and the build and runtime namespaces.
agents.networkPolicy.enforcedfalseYour statement that the cluster enforces NetworkPolicy. Without it, deploys are refused.
agents.networkPolicy.allowUnenforcedfalseDeploy anyway on a cluster that doesn't enforce it. Development only: every such deployment is marked.
agents.networkPolicy.privateEgressCidrsemptyPrivate ranges an agent's network.egress may open, such as an internal CRM. Never list the cluster's pod, service or node ranges.
agents.runtimeNamespace, agents.buildNamespacevois-orchestrate-agents, vois-orchestrate-agent-buildsWhere agents run and build. No platform Secret is mounted there.
agents.build.harborProjectvois-orchestrate-agentsThe registry project for agent images.
agents.build.pushCredentialsSecretharbor-agents-push-credentialsA docker config Secret in the build namespace that can push agent images and pull base images.
agents.build.pipIndexUrl and relatedemptyYour Python package mirror.
agents.build.packageIndex.sourceautoauto uses your mirror if set, else the offline package index if on, else public PyPI.
agents.build.egressemptyWhere build pods may connect: your registry and package index. With an enforcing engine and this empty, builds can't reach the registry.
agents.build.maxConcurrent, agents.build.timeoutMinutes2, 45Builds at once across the install, and the longest one may take.
agents.wheelhouse.enabledfalseThe offline package index from the bundle. voisctl upgrade --bundle turns it on.
agents.runtime.integrityEnforcementenforceRefuse agent images Orchestrate didn't build.
agents.runtime.previousLiveWarmReplicas1Replicas kept on the previous live version for a fast rollback.
agents.quotas.projectDefaultsemptyCode agent quotas for projects without their own.
agents.a2a.grpc.enabled, agents.a2a.grpc.portfalse, 50051Serve the A2A gRPC binding from the API. With ingress exposure and the API exposed, it is routed on the API host and cards list <api host>:443. See Agents calling agents.
agents.a2a.grpc.publicAddressemptyThe host:port cards list for gRPC, for other exposure modes. Empty derives it from the API's ingress.
agents.a2a.identity.installId, agents.a2a.identity.signingKeyemptyThis install's A2A identity. Empty generates one on first install and keeps it. Set them only to carry an identity over to a rebuilt install. See Trusted installs.

Code agents also need Vault roles for their sources, secrets and runtime keys. See Code agents.

Connectors and tool execution​

ValueDefaultWhat it does
toolExecution.sandbox.platformAddressestrueRefuse every pod, node and service address in the platform's namespaces as a connector destination. The list is refreshed every 30 seconds.
toolExecution.sandbox.blockedCidrsemptyMore ranges to refuse outright.
toolExecution.networkPolicy.enabledfalseA NetworkPolicy on tool execution: it then reaches only the platform services it needs and the destinations below.
toolExecution.networkPolicy.publicHttpstrueAllow 443 to public addresses, for SaaS connectors and OAuth providers.
toolExecution.networkPolicy.egressemptyYour databases and APIs, by cidr outside the cluster or by selector (to) inside it, with ports. On an air-gapped install list every database, REST API and MCP server.

Phone calls (SIP)​

ValueDefaultWhat it does
sipGateway.enabledfalseRuns the SIP gateway.
sipGateway.publicIp, sipGateway.mediaIpemptyThe addresses carriers reach for SIP and media. Empty means calls from inside the cluster only.
sipGateway.publicHostemptyShown in the console as where to point a trunk.
sipGateway.hostNetworkfalseMedia on the node's own interfaces, the usual way a carrier reaches the RTP ports. Needs a node with a routable address.
sipGateway.publicService.enabledfalseA LoadBalancer Service for SIP on 5060.
sipGateway.rtpPortMin, sipGateway.rtpPortMax20000 to 20199The media ports, one even port per call.
sipGateway.maxCalls, sipGateway.replicas50, 1Calls per gateway pod, and pods.
sipGateway.codecsPCMU,PCMAG.711 µ-law and A-law.
sipGateway.rtpStrictSourcetrueTake media only from the address the far end's SDP names.
sipGateway.tls.enabled, sipGateway.tls.secretNamefalse, emptySIP over TLS on port 5061 (tls.port), with a kubernetes.io/tls Secret in the release namespace.
sipGateway.tls.verifyPeer, sipGateway.tls.minVersiontrue, TLSv1.2Check the carrier's certificate on outbound TLS, and the lowest TLS version.
sipGateway.srtpoffoptional or required: encrypted media (SDES SRTP). Use it with TLS.

Opening the SIP and RTP ports to your carrier is infrastructure work outside the chart. See Phone calls (SIP).

Hybrid sites​

hybrid.gateway.enabled turns on the site gateway that sites on customer premises dial out to. Expose it with a TCP load balancer (hybrid.gateway.service.type: LoadBalancer), list its public name in hybrid.gateway.hosts, and set hybrid.gateway.publicUrl. See Hybrid sites.

Voice (LiveKit)​

Voice sessions in the browser use LiveKit. The browser connects to LiveKit's signalling port through your ingress, and sends media straight to the node that runs LiveKit.

  1. Set LiveKit's public hostname:

    livekit:
    publicHost: orchestrate-voice.example.com

    With ingress exposure and TLS, browsers connect over wss://. The chart sets one-hour proxy timeouts on LiveKit's Ingress for ingress-nginx, so long sessions aren't cut off. With another ingress controller, set the equivalent through exposure.components.livekit.annotations.

  2. Open 7881/TCP and 7882/UDP from your users to the nodes that can run LiveKit.

ValueDefaultWhat it does
livekit.publicHostemptyThe hostname browsers use for LiveKit signalling. Empty publishes nothing.
livekit.nodeSelectoremptyPins LiveKit to labelled nodes, so media ports stay on known addresses.
livekit.stunServersemptySTUN servers given to browsers. Empty gives them none that work, so media uses LiveKit's own node address, which clients on your network reach directly. Set an internal STUN server when clients are behind a NAT that needs one.
livekit.publicStunfalsetrue lets LiveKit hand browsers its default public STUN servers when stunServers is empty. Only for an internet-connected install that accepts this.
livekit.externalTurnServersemptyTURN servers given to browsers, each with host, port, protocol, username and credential. The credential is sent to every browser as is.
livekit.turn.*offLiveKit's built-in TURN server. Contact us before you turn it on.

LiveKit uses a single pod on the node's host network, and it is replaced, not rolled, on a change. Voice is unavailable for everyone while it restarts. If LiveKit moves to another node, restart the voice worker too:

kubectl --context <your-context> -n vois-orchestrate rollout restart deployment/realtime-agent-worker

Governance​

ValueDefaultWhat it does
opa.bundlePolling.minDelaySeconds, opa.bundlePolling.maxDelaySeconds5, 15How often Open Policy Agent fetches the custom policies you save in the console.

Custom policies, guardrail profiles, model access and compliance settings are stored in the database, so they carry across upgrades.

Audit log​

ValueDefaultWhat it does
audit.chainKey, audit.signingKey, audit.writeKeygeneratedThe chain's HMAC key, the Ed25519 key that signs checkpoints, and the key services use to write events. Generated once into the audit-keys Secret and kept. Set them, or audit.existingSecret, to bring your own.
audit.chainPreviousKeys, audit.signingPreviousPublicKeysemptyOld keys that must still verify, after a rotation. Never drop one while events signed with it are kept.
audit.writerRole.enabledtrueThe audit service connects as a role that can only read and add events.
audit.checkpointIntervalSeconds300How often the chain head is signed, when it has moved.
audit.retention.days, audit.retention.minDays0 (forever), 365How long events are kept, and the lowest the console may set.
audit.export.sinksemptySIEM sinks managed here: syslog, webhook, Splunk HEC or S3. Shown read-only in the console.

See Audit log.

Licence​

ValueDefaultWhat it does
license.secretNameemptyA Secret holding your licence under the key license.txt.
license.storage.enabledtrueA small volume for a licence uploaded on the console's Licence page. Needs a default StorageClass.
license.graceDays14Days after expiry during which everything keeps working, with a warning.
license.revalidateSeconds600How often the License Service reads the licence again.

There is no public key setting: the License Service trusts only the VoisX keys built into it. See Licence.

High availability and backups​

ValueDefaultWhat it does
availability.replicasemptyRun stateless services more than once, for example { api: 2, web: 2, modelGateway: 2 }. Allowed: api, web, modelGateway, toolExecution, ragIngestion, audit, opa, orchestrationWorker, realtimeAgentWorker, eval.
availability.podDisruptionBudgetstrueA PodDisruptionBudget for each service with two or more replicas.
postgresHA.modeoffPostgreSQL on CloudNativePG. A fresh install uses cnpg with postgresHA.bootstrap: initdb. An existing install moves over in two gated steps.
postgresHA.instances3A primary and streaming replicas.
postgresHA.synchronous.enabledfalseSynchronous replication, for no loss of acknowledged writes on failover. See Synchronous replication.
postgresHA.synchronous.method, .numberany, 1Quorum (any) or priority (first), and how many standbys each commit waits for, 1 to instances - 1.
postgresHA.synchronous.dataDurabilityrequiredrequired blocks writes while too few standbys are up. preferred never blocks, and suspends the guarantee meanwhile.
postgresHA.synchronous.allowWriteStallsfalseNeeded for required with no spare standby, where routine restarts would stop all writes. The chart refuses that combination without it.
postgresHA.synchronous.failoverQuorumfalseWait rather than promote a replica that may lack a synchronous commit. Matters from 3 instances.
backups.s3.*emptyOne S3-compatible target for every backup.
backups.postgres, .qdrant, .minio, .redis, .licensePostgreSQL on, the rest offWhat is backed up, when, and how many copies are kept.

See High availability and backups.

Observability​

ValueDefaultWhat it does
observability.metrics.enabledfalsePrometheus metrics from the API and the Model Gateway, on a separate port (9464), with scrape annotations.
observability.metrics.serviceMonitor.enabledfalseA prometheus-operator ServiceMonitor per service. Needs the operator's CRDs.
observability.tracing.otlpEndpointemptyAn OTLP/HTTP collector for traces. Empty sends none anywhere.

Temporal​

ValueDefaultWhat it does
temporal.payloadEncryption.enabledtrueEvery payload the orchestration engine stores is encrypted with AES-256-GCM by the services themselves, so its database and web UI hold only ciphertext. The key is generated once and kept, or comes from temporal.payloadEncryption.existingSecret. Losing every key makes running workflows unreadable.
temporal.namespaceRetention24hHow long a closed run's history is kept. Applies when the namespace is first created.

Projects​

projects.purgeAfterHours (default 168) is how long a deleted project can be restored before it is purged. See Delete a project.

Storage​

ValueDefault
postgres.storageSize5Gi
qdrant.storageSize5Gi
minio.storageSize10Gi
redis.storageSize1Gi
postgresHA.storage.size10Gi per instance
license.storage.size1Gi

Every volume is ReadWriteOnce.

Set the sizes before the first install. Kubernetes doesn't let a chart change the volume size of an existing StatefulSet, so changing these values later makes the upgrade fail. To grow a volume later, resize its PersistentVolumeClaim directly, if your StorageClass allows expansion. The claims are named data-postgres-0, data-qdrant-0, data-minio-0 and data-redis-0.

Resources​

The chart sizes components through four tiers. You can change the requests and limits of each tier.

TierDefault requestsDefault limitsUsed by
resources.small100m CPU, 128Mi500m CPU, 256MiLicense Service, audit, eval worker, Open Policy Agent, LiveKit, Redis
resources.medium250m CPU, 256Mi1 CPU, 512MiAPI, Model Gateway, console, orchestration worker, tool execution, deployment controller, knowledge-base ingestion, agent edge, Qdrant, MinIO
resources.jvm500m CPU, 512Mi1 CPU, 1GiKeycloak
resources.ml_inference500m CPU, 1Gi2 CPU, 2GiRealtime voice worker

Models you deploy are sized separately, by the resource profile you choose when you deploy them. Code agents are sized by their runtime.resources profile. The SIP gateway has its own sipGateway.edgeResources and sipGateway.bridgeResources.

Other values​

ValueDefaultNotes
deploymentModestandaloneLeave as standalone.
Was this page helpful?