Skip to main content
Version: 3.4.0-rc.1

Changelog

[3.4.0] - 2026-10-02​

Release focused on RAG settings in the portal, temporarily disabling API keys, group visibility for tenant members, and removal of the legacy manager.

Added​

  • RAG
    • Ingestion and search settings in the portal, on creation and on the Settings tab (renamed from Model settings): require a group for retrieval, delete files that have no group, use PDF OCR by default, and use smart temporal search by default. Request values still override the OCR and temporal search defaults
    • Tokenizers for platform-hosted embedding models are delivered to RAG automatically, so token counts and chunking match the embedding model. External models keep their existing tokenizer handling
  • API Keys
    • Personal keys and service-account keys can be disabled and enabled again from the portal without replacing secrets or grants
    • The key ID is shown under the key name in the API keys table and on the detail page
  • Chat
    • Before opening an OAuth provider, Chat shows which service an agent or tool needs and waits for Connect. Previously it opened the authorization page without explanation
  • Tenants
    • Tenant members can open the Groups page and see the groups they belong to and who is in them
  • Portal
    • Usage charts show the date and shade alternate days, so 24-hour and 7-day windows are readable
    • The models pages explain each identifier in a tooltip, and API key grants show and link the API model ID instead of the Kubernetes object name

Changed​

  • RAG
    • Removed the RAG /v1/chat/completions endpoint. Retrieve context with /context and generate answers separately through the Model Gateway
    • The generation model is optional for retrieval and ordinary ingestion. Contextual indexing still requires one
    • RAG settings are stored on the deployment and applied by restarting the RAG pod. /reload-config is removed. An upgrade migration copies existing settings over
  • Platform
    • The legacy manager and realtime services are removed. The new manager provides every capability they did, though not through the same APIs
    • The data-connector service is removed. Custom scrapers can still upload files to RAG over /internal/files
  • Model Gateway
    • /v1/audio/transcriptions accepts requests up to 128 MiB, and Chat accepts audio files up to 127 MiB. Other inference requests stay capped at 32 MiB
    • A tool call with invalid arguments returns the validation error to the model so it can retry, instead of ending the Responses turn
  • MCP
    • External MCP servers that use the 2026-07-28 MCP specification can be registered and invoked. Servers using the older handshake keep working
  • Tenants
    • Any tenant member can share resources they administer with other tenant members. Previously the share picker only worked for tenant admins, because only they could list the tenant's members

Fixed​

  • Model Gateway
    • Preset agents report their base model's input modalities on /v1/models, so Chat allows image uploads to agents built on vision models
  • Chat
    • Typing no longer re-renders every message, so long conversations stay responsive
    • The jump-to-bottom button hides when closing a reasoning block shortens the transcript
  • Portal
    • Dismissing a form dialog with unsaved input asks for confirmation instead of discarding it
    • Previous on the RAG files list loads the preceding page instead of reloading the current files
    • A newly created group shows up as soon as creation succeeds. If creation fails, the form keeps the entered name
  • API Keys
    • Displayed key prefixes match the cm_api_ secret format. Stored prefixes are migrated without changing secrets
    • Metadata updates on revoked keys are rejected, matching the portal and the documented lifecycle
  • Keycloak
    • Trusted CA names up to 30 characters work. Longer names previously left Keycloak running without the CA and reported nothing. The chart now fails the render instead

[3.3.0] - 2026-09-17​

Release focused on preset-agent output guardrails, one permission model for every resource type, and portal-v2 as the only portal.

Added​

  • Preset Agents
    • Output guardrails: up to four ordered guardrail models check a final answer against a written policy before it is stored or returned. The first block replaces the answer with the configured refusal, and only the refusal is stored or replayed in later turns
    • Guarded preset agents support streaming protocol but response is buffered and returned as a single result after the guardrail has accepted it
    • A description field, editable in place from the detail page
  • Models
    • A guardrail model type, so policy judges stay out of chat model discovery while remaining callable. The first supported judge protocol is GPT-OSS-Safeguard
    • A platform-level default transcription model, so Chat no longer fails when a user is authorized for more than one
    • Some deployment presets were added and updated
    • Presets declare their input modalities, so image-capable models are discoverable through the input_modalities filter on /v1/models
    • Each deployment reports its last observed context limit as runtime_info.max_model_len, without requiring an explicit max-model-len flag. The value is retained through outages, restarts and scale-to-zero
      • The value is available through the API only. Chat visibility is added in a later release
  • Chat
    • Tenant admins can edit the assistant instructions used by direct-model Chat from a Chat Settings page in the portal, with Reset to Default restoring the built-in baseline. Preset-agent prompts are unaffected
    • Image uploads are refused up front when the active model declares no image input, instead of failing server-side with no explanation. Documents and audio keep working on text-only models, since both are reduced to text in the browser
  • Observability
    • Audit and platform logs can be exported to a Splunk HTTP Event Collector
  • Portal
    • Status and access filters on the models lists, alongside the existing source filter
    • Members can be added to a group directly from their member detail page
    • MCP server API keys report whether a credential is stored, shown as Configured without exposing any part of the key
  • Documentation
    • Guides for per-user OAuth against external MCP servers, including SharePoint setup and troubleshooting

Changed​

  • Resource Permissions
    • Access control to resources is unified for each resource type (model deployments, external model connections, preset agents, mcp servers, rag endpoints, model aliases)
      • A platform-scoped resources can be shared with all tenants or chosen tenants as a read-only resource. Platform-scoped resources can only be managed from platform scope
      • Any tenant-scoped resources can by default be accessed by the creator and tenant admins. Resources can be shared to the entire tenant, or to specific users or groups within the tenant
  • Portal
    • The legacy V1 portal is removed; portal-v2 is the only portal and serves the portal subdomain directly. The portalv2 and oldportal subdomains no longer exist, and the enablePortalV2Frontend feature flag is gone
    • Unified the look and feel of multiple portal pages
    • Dates follow the viewer's own locale; technical timestamps such as log rows and chart buckets stay ISO 8601. Previously every viewer saw a pinned en-GB convention
  • Chat
    • Chat sends one chat_mode option that carries the fixed Chat formatting guidance, the tenant's instructions and the current date
      • Preset agents use their own configured instructions and never inherit the tenant prompt. API callers can opt in explicitly.
  • Platform
    • The tools subdomain and namespace are removed. Nothing has used them since the 3.0.0 release
  • RAG
    • CSV files are parsed into one labelled block per row instead of a markdown pipe table, so every retrieved chunk carries its column names rather than only the chunk that happens to hold the header; parsed-but-empty uploads now return 400 instead of failing later as a 500. Chunks stored before this change need re-ingestion to adopt the new format

Fixed​

  • Model Gateway
    • Chat Completions streams no longer drop when the backend pauses between events or holds its connection open after [DONE]
  • Models
    • Deleting a model deployment or external model no longer takes its aliases with it; an alias survives its target and reports the target as unresolved. References left by the previous controller are stripped on upgrade
  • Transcription
    • Whisper recovers from repetitive decoding windows: failed windows retry across the temperature fallback sequence, and earlier output no longer conditions later windows.
  • RAG
    • A retrieval config asking for reranking with no reranker model configured now checks for the model first and runs plain retrieval once, instead of embedding and searching twice before discovering the model is missing
  • Chat
    • Deleting a conversation while a fork is in flight no longer drops the view into a fork of the conversation that was just deleted, or duplicates its row in the sidebar. A failed delete now rolls back instead of leaving the view on an empty chat
    • Row menus in the conversation list popup are no longer clipped out of reach near the bottom of a small viewport
  • Portal
    • The default theme's accent now meets WCAG 2.2 AA contrast
    • A user with no tenant memberships or platform-level access rights can log out from the no-tenant page, which previously rendered outside the app shell and offered no way out of the session
    • Stopped model deployments are shown disabled in the RAG create form rather than being selectable and failing at query time, and a stale or unavailable optional default no longer blocks creation
    • An invalid or past API key expiry date reports what is wrong on the field itself, instead of a generic unexpected-error message
    • Renaming an API key no longer clears its description

[3.2.0] - 2026-09-02​

Release focused on a new portal version and on RAG data source file lifecycle.

Added​

  • Portal
    • A new portal version: the portal and portalv2 subdomains both serve it, and the legacy portal stays reachable on oldportal until the next release
    • Refreshed look and feel, a long list of fixes to problems the legacy portal had, and a clearer separation between tenant scope and platform scope
  • RAG Data Sources
    • Allow bulk removal of the synced files after disconnecting a datasource
    • Tenant-wide view of every data source across every RAG in one place
  • RAG
    • Exact file totals for the active filter set on GET /files via include_total, so the portal reports real file counts instead of stopping at 16

Changed​

  • Platform
    • Feature flags retired — the model gateway, operator-managed model deployments, OME, the chat frontend, chat tool selection and the passport auth path are now always on
  • RAG Data Sources
    • Support for legacy data sources is removed from the platfrom
  • RAG
    • Every model is resolved through the model gateway: MODEL_GATEWAY_URL is now required, and the reranker's direct TEI/Infinity endpoint discovery is gone
    • REDCap is no longer a supported data source provider; S3 and Azure remain

Fixed​

  • Authentication
    • Clean installs could return 401 from every api-services API indefinitely, because istiod resolved the passport key set while authn was still starting and never rebuilt it; the key set now travels in the RequestAuthentication
  • Model Gateway
    • Conversation history retains 128 items instead of 50, so long chats lose less context before token-aware limiting arrives
  • RAG
    • First startup no longer fails when PgBouncer closes the idle outer migration transaction
    • The vision model picker offers only OCR-typed models, matching what document extraction accepts
  • OpenShift
    • Chat and the new portal have reachable Routes in the tls-reencrypt chart

[3.1.0] - 2026-08-25​

Release focused on preset-agent creation, extensible RAG data sources, and platform-wide distributed tracing.

Added​

  • Preset Agents
    • Create and manage preset agents in the portal
    • Access rules enforced at creation and delegation time via FGA grant-time checks
  • RAG Data Sources
    • Extensible data source system for RAG deployments: S3 and Azure Blob source adapters, source-aware file operations, runtime state persistence, and reconciliation
    • Manager API for configuring RAG data sources and credentials (CRUD with existing RAG permissions)
    • Model operator projects data source configuration onto RAG pods
    • Metadata discovery and filter APIs for RAG sources
  • Distributed Tracing
    • OpenTelemetry tracing library for Python services, replacing the home-grown tracing implementation
    • RAG services report traces through the OpenTelemetry pipeline
    • Go services trace database queries (ClickHouse and Postgres), outbound calls, and record request IDs on server spans
    • Model operator traces inference on vLLM deployments and points RAG pods at the trace collector
    • Platform takes ownership of trace identity at the ingress: external traceparents are quarantined instead of controlling in-mesh trace IDs
    • Model deployments can opt out of inference tracing
  • Monitoring
    • Log and trace storage can be disabled per signal type
    • Observability data expires after a week by default
    • Prometheus scrape spans dropped at the collector to reduce noise
  • Model Gateway
    • Engine-neutral transcription API for speech-to-text backends (OpenAI-compatible, Azure supported)
  • Chat
    • Multiple concurrent active chats
    • Paste images from clipboard into chat
    • Message timestamps displayed
    • Disabled tool buttons show tooltips when a preset agent is selected
  • Infrastructure
    • Extra hosts configurable on the shared Istio ingress gateway
    • fastText language-detection model baked into the RAG image
    • vLLM serving image defaulted to v0.26.0

Fixed​

  • Platform
    • Internal CA (cm-selfsigned-ca) lifetime extended to outlive the leaf certificates it signs, fixing intermittent model-operator webhook outages
  • Model Gateway
    • Image detail preserved in tool-loop responses
    • Tool access scoped per request so historical calls cannot imply current access
  • Chat
    • Server timestamp used for loaded messages
    • Agent-to-model conversation switches now persist
  • Manager-go
    • Azure Blob data source type accepted
    • No-tenant MCP list denied instead of returning empty (preventing accidental access)
    • Data connector secrets always filtered from API responses
    • Model deployment credentials secret creation fixed
  • Legacy Portal
    • Misleading 5MB upload limit text removed
    • Native RAG file picker used instead of custom picker

[3.0.4] - 2026-08-14​

Release focused on platform-generated passport signing keys, API key usage reporting, and OAuth MCP hardening.

Added​

  • Authentication
    • Passport JWT signing key generated automatically on install (no manual key provisioning)
    • Passport JWKS published for token verification by other services
    • ECC key support for the passport JWT
  • Usage & Metering
    • API key usage details with CSV export in the portal
    • Owner-scoped API key token usage reporting
    • Tenant API key usage by model exportable as CSV
  • API Keys
    • Model alias grants — an API key can be granted access to an alias directly
  • Infrastructure
    • Extra CORS origins configurable for manager and Keycloak via corsCustomAllowedOrigins
    • Operators can serve their own TLS Secret
    • ArgoCD UI optionally published on its own subdomain
    • Installers can declare a CNPG volume that was grown manually
    • Platform installer bumped to 0.2.15

Fixed​

  • External MCP OAuth
    • Direct OAuth MCP tool calls enabled through the gateway
    • OAuth requirements exposed in MCP tool discovery
    • OAuth discovery edge cases handled; discovered resource omitted for explicit scopes
  • Model Gateway
    • User-friendly error messages returned for backend failures
  • Model Operator
    • Model startup probe window extended to survive slow model loads
  • Chat
    • Loading state shown on tool picker buttons
  • Installer
    • Installer values reliably reach their consumers
    • ArgoCD pre-install hook survives a previously failed run

[3.0.3] - 2026-08-04​

Release focused on OAuth for external MCP servers, air-gapped installation support, and content capture.

Added​

  • External MCP OAuth
    • End-to-end OAuth support for external MCP servers: configuration, secret handling, gateway auth, and in-chat connect flow
    • Portal UI for configuring external MCP server authentication
    • OAuth dynamic client registration support
  • Content Capture
    • Per-tenant, per-user gated capture of prompts, answers, and tool-call content
    • Content-capture toggle on the tenant admin page in the portal
    • Captured gen_ai events stored in a dedicated table with a tenant-scoped read endpoint
    • Tool call arguments and output recorded on trace spans
  • Air-gapped Installation
    • Install tarball assembler covering all third-party images and charts
    • OpenShift air-gap support (IDMS-based image mirroring, mirror verification)
    • HTTPS auth support for the GitOps repo in the installer
  • Model Serving
    • Pre-staged local model weights can be served without HuggingFace access, with a matching deploy-form option
    • OCR model type for RAG document processing
  • RAG
    • Per-RAG description is now a required, model-facing field
  • Chat UI
    • Accept or decline tenant invites in chat
    • Thinking duration and partial/empty responses rendered
    • Explicit confirmation required when switching models mid-conversation
    • Current date injected into responses
  • Model Gateway
    • Human-readable RAG tool names

Fixed​

  • Model Gateway
    • Bad requests rejected before committing the tool-loop stream
    • Conversation ownership enforced on the responses continuation path
    • Response stream error types preserved; empty conversation messages ignored
    • Conversation setup no longer races SSE keepalives
  • Chat
    • Assistant messages no longer get stuck streaming forever
    • RAG tool selection keyed by stable source id instead of the mutable name
    • Conversation list ordered by latest activity
  • Manager-go
    • Portal survives ClickHouse outages instead of crash-looping
    • stack-admin seeded into DEFAULT tenant admins on bootstrap
    • Legacy default-apikey policy gated on enableApiKey
  • RAG
    • Air-gapped uploads no longer fail on oversized embed inputs
  • Metering
    • ClickHouse usage insert columns named explicitly

Changed​

  • Istio ingress service type is now parameterisable

[3.0.2] - 2026-07-01​

Release focused on chat tool visibility, safer tenant deletion, and real-time tool streaming in the gateway.

Added​

  • Chat UI
    • Tool usage display: separate tool and knowledge-base pickers, with per-answer view of what was used
    • Improved in-chat tool call rendering
    • Model detail page
  • Tenant Lifecycle
    • Tenant deletion blocked while live resources exist, with a guided resource cleanup flow
    • Tenant rows soft-deleted on deletion
  • Portal
    • Models list split into tenant and platform models
    • External MCP server list split into tenant and platform scopes to match models
  • Model Gateway
    • Tool-loop turn deltas streamed to the client in real time
    • tools/list filtering by kind

Fixed​

  • Model Gateway
    • Keepalives sent when post-tool continuation is slow
    • One failed RAG dispatch no longer cancels its sibling dispatches
    • Service-account keys can own conversations
  • Chat
    • Tool display cleaned up mid-stream; cancelled chat input ids preserved
  • Portal
    • Documentation link fixed
  • Manager-go
    • Platform admin can list and delete manager services for any tenant
  • Keycloak Integration
    • Service-account tokens cached and refreshed instead of re-fetched per call

[3.0.1] - 2026-06-25​

Release focused on external MCP tool integrations and RAG-as-MCP capability.

Added​

  • External MCP Tools
    • Register and manage external MCP servers (e.g., web search, custom APIs) in the platform
    • MCP server CRUD API and Kubernetes CRD for persistence
    • Tenant managers can configure MCP tools that are available for everyone
    • Tenant admins can configure MCP tools that are available for others in the tenant
    • Portal UI for creating, configuring, and testing external MCP servers
    • Tools from external MCP servers appear alongside built-in tools in chat and API
  • RAG as MCP
    • Built-in RAG MCP server allows deployed RAGs to be used as tools via MCP protocol
    • RAG tools available through model gateway with consistent API
    • Tenant-scoped RAG tool access via OpenFGA authorization
    • Accessible in chat UI
  • Chat UI Improvements
    • Expandable tool list showing available tools and their descriptions
    • Display model names (display_name) in chat model picker and detail view
    • Accessibility improvements
    • Thinking block rendering from inline reasoning in message history
  • Tenant scoped external models
    • External models can now be configured inside one tenant by tenant admin

Fixed​

  • Model Gateway
    • SSE heartbeat handling improved to prevent connection timeouts during tool loop execution
    • Tool response heartbeats replaced with response.in_progress for clarity
  • Portal
    • Tenant slug uniqueness enforced across manager, portal, and Deno services
    • Resolved tenant metric names from new model resources
    • Chat opens to correct selected workspace
    • Model deployment display shows permission badges correctly
    • External model example texts updated
  • Manager-go
    • Admin token refreshed during tenant-split migration to prevent expiration
    • Group names included in Keycloak re-parent operations
    • User group pagination for large tenant memberships
    • Model-based RAG without connectors handled correctly
  • Keycloak Integration
    • Default user invite emails improved for deliverability
  • Operator
    • CNPG cluster no longer rewritten on every reconcile (performance improvement)
  • Authorization
    • External model tenant scoping enforced
    • Batch authorization support added for OpenFGA

Changed​

  • External MCP naming now scoped with tenant slug prefix to avoid collisions
  • Model deployment form displays actual default vLLM image instead of placeholder

[3.0.0] - 2026-05-28​

Major release. Four headline areas: a single model gateway replacing per-tenant model endpoints, a model deployment overhaul (V1) with an operator-managed catalog, an in-platform Chat UI (V1) bundled with the base install, and reworked authorization with a separate API key UI. Migration tooling and legacy import paths let 2.x clusters move over in place.

Added​

  • Model Gateway
    • Single gateway replacing per-tenant model endpoints
    • Streaming Responses tool loop with sequential dispatch and aggregated usage reporting
    • Function tool policy and allowlist enforcement
    • Reranking endpoint (/v1/rerank) so RAG no longer talks to model backends directly
    • Azure and OpenAI provider profiles for external model passthrough
  • Model identifiers
    • New public model identifier system: publicModelId + ModelAlias replacing ModelRoute
    • Enables no-downtime model swaps by pointing an alias at a new deployment
  • RAG
    • BM25 indexing pipeline (db migration + ingestion)
    • MCP protocol surface so RAGs are usable from MCP-aware tools
    • OpenFGA authorization enforced for RAG access
    • /v1/models endpoint for OpenWebUI compatibility
  • Chat UI (V1, bundled with the base install)
    • Conversation forking and message editing (delete-from-point API)
    • Retry / regenerate button on assistant messages
    • Settings modal with Keycloak logout
    • Pagination for conversation list and message history
    • Playwright e2e test suite
    • Auto-redirect to Keycloak login, public SPA Keycloak client
    • Accessibility improvements
  • Model Operator
    • Tekton-based model deployments migrated to the operator
    • Basic deployment presets (defaulted GPU memory utilization, restricted publicModelId character set)
    • ModelRoute CRD (later subsumed by the alias system)
    • Metrics tab for operator-deployed models in the portal
    • FGA tuples synced for ModelDeployment and ExternalModel, plus user and group APIs for deployments
  • Portal
    • API key management UI (creation, revocation, grants visualization)
    • API key creation shortcut directly from model and RAG detail pages
    • Model and RAG API URLs shown for copy-paste into SDKs
    • Model route management UI
    • New model deployment UX gated behind a feature flag
  • Authorization & Security
    • Python passport + FGA middleware for Python services
    • Gateway inference endpoint authorization via ext_authz
    • Passport identity carried into audit logs and traces
    • User JWT passports scoped to tenant via X-Cm-Tenant-Id
    • Keycloak / FGA bridge writes platform tenant tuples on tenant create
  • Manager-go (legacy migration)
    • Tenant CRUD service
    • Bulk legacy RAG migration (cmctl migrate-all-rags)
    • Legacy api-keys imported and legacy model sharing migrated
    • Legacy TEI model deployments migrated to vLLM
    • GPU config promoted into dedicated CRD fields
    • Tenant members can list invokable models
  • Observability
    • Per-request log record on every inbound RPC/HTTP request
    • Request-log and audit records emitted on handler panic
  • Infrastructure
    • OpenShift compatibility features and docs (SCC for otel-collector-logs, OCP conditions)
    • OpenTelemetry otlphttp exporter and image override support
    • Platform installer 0.2.7 with new docs

Changed​

  • RAG
    • All model calls now route through the model gateway
    • Vision model support routed through model-gateway (dots OCR)
    • Auto image updates unless an explicit spec.image is set
  • Portal
    • Tenant tagging via X-Cm-Tenant-Id header across portal calls
    • Removed unnecessary admin gates, removed custom tools auth
  • Infrastructure
    • CNPG bumped to 0.28.2
    • Istio init filtered out of workload logs
  • pnpm version pinned via packageManager, pg_textsearch bumped to 1.1.0 in cnpg-pg-bm25

Fixed​

  • ClickHouse backup CronJob spurious-failure fix

[2.7.1] - 2026-02-04​

Added​

  • RAG service enhancements
    • Contextual retrieval implementation using Anthropic's technique for improved retrieval accuracy
    • dots.OCR integration with VLM-based OCR pipeline for better document processing
    • Enhanced chunk metadata with inline citations and URL references in completions
  • Anthropic model support in external-llm service
  • GPU support for notebooks
  • Multi-platform (ARM64/AMD64) images for all services

Changed​

  • Portal
    • Tenant display now shown without requiring hover
    • Delete endpoint now prompts for database deletion confirmation
  • Manager startup probe timeout extended for reliability
  • Dependencies updated for mcp-agent, baserag-mcp, and postgres-mcp services

Fixed​

  • Portal external model deployment configuration
  • Manager API key filtering
  • Keycloak: email invite functionality
  • Portal empty config removal handling
  • Azure model configurations for external-llm

[2.7.0] - 2026-01-22​

Added​

  • External Secrets Operator (ESO) integration for secure credential management with Kubernetes secrets and Vault support
  • Chat UI [ALPHA]
    • File viewer sidebar in chat interface for viewing and uploading files directly to RAG
  • Automatic database provisioning with isolated databases per RAG deployment in shared CNPG cluster
  • Portal
    • Root admin user management panel for viewing all platform users and their tenants
    • Platform user invite functionality with optional tenant assignment
    • Customizable navbar logo via environment variable
  • Tenant ID tracking for usage metrics

Changed​

  • RAG reranker scores now use sigmoid normalization to always return values in 0-1 range instead of unbounded logits
  • Keycloak password reset enabled by default

Fixed​

  • ClickHouse operator deployment issues
  • Model endpoint configuration breaking when removing the last connector
  • Agent connector polling error when connectors are not yet set
  • Tenant ID not being included in usage records
  • Tenant overview not showing all columns

[2.6.10] - 2025-12-10​

Added​

  • Chat UI [ALPHA]
    • Chat client deployment infrastructure support
    • File and image upload with markdown rendering capabilities
  • Platform
    • Migration job support for services
    • Restored functionality to customize subdomains
    • Database backup cleanup automation with CNPG pgBackRest

Changed​

  • Upgraded CNPG from 1.25 to 1.27 (chart bump from 0.23 to 0.26)
  • Upgraded platform installer to v0.2.0
  • Improved service database implementation in manager-go
  • Simplified platform base templates and backup cleanup scripts

Fixed​

  • Portal configuration issues when creating models
  • Template parsing errors
  • Reduced pgBackRest resource requirements
  • Model endpoint timeout configuration for non-streaming requests
  • Various configuration improvements

[2.6.9] - 2025-11-14​

Fixed​

  • Fixed an issue where large number of concurrent tenant invitations could cause the portal client to not render the invitation to some users
  • Fix custom token expiration option
  • Fix an issue where the client wouldn't display all services in large deployments

Changed​

  • Improved ConfidentialMind platform installer and it's documentation

[2.6.8] - 2025-10-22​

Fixed​

  • Fixed an issue where UI would sometimes fail to load services after refresh
  • Deployment issue that caused deployments to report failure even if platform was operational

[2.6.7] - [SKIPPED]​

[2.6.6] - 2025-10-20​

Added​

  • Alerts for undeployed connectors in service UI
  • Information about connected services for database listing UI

Changed​

  • Improved multi-turn conversation accuracy in RAG

Fixed​

  • Removed deployment of unnecessary PVCs when creating a volume through UI

[2.6.5] - 2025-10-14​

Added​

  • Tenant-based model cache salting
    • This divides the shared model cache into tenant segments
    • Makes sure there's no cross-tenant data access through the cache

Changed​

  • More compact and readable log display format in portal
  • Improved RAG query encoding

Fixed​

  • Fixed an issue where non-vllm model deployments wouldn't work due to a vllm-specific flag
  • Various small UI fixes

[2.6.4] - 2025-10-03​

Added​

  • Tenant root admin role to create and edit tenants
  • Tenant level usage tracking
    • Can be used to integrate automated token based billing
  • Tenant specific usage limits
    • Can be used to create different tenant based business tiers
  • Portal UI for tenant root admin
  • Self-signed certificate support

Fixed​

  • Correct title for model deployment page
  • Correct model endpoint image versioning

[2.6.3] - 2025-09-18​

Added - Platform​

  • Multitenancy system with base tenant segregation
  • Tenant sub-groups for sharing deployments and structuring inter-tenant user separation
  • Tenant invite system for existing users (in-platform popup)
  • Email invites for non-existing users with temporary password
  • Ability for new users to create their own tenant
  • Tenant admin capabilities to see all services and manage sub-groups

Changed - Platform​

  • Every user must now be under at least one tenant
  • Only one tenant can be active in portal UI at a time
  • All deployments are now under a specific tenant
  • User/group visibility limited to same tenant when sharing services

Added - CM-RAG v0.7.4​

  • Enhanced tracing and usage tracking

Added - CM-Agent v0.2.13​

  • Enhanced tracing and usage tracking
  • Improved local development capabilities

Added - CM-Model-Endpoint v0.1.4​

  • Enhanced tracing and usage tracking
  • Handling for canceled requests

Changed - CM-Model-Endpoint v0.1.4​

  • Increased stream request timeout to 10 minutes to support long-context models

Fixed - CM-Model-Endpoint v0.1.4​

  • Bug in stream processing which could sometimes break client SSE parsing

[2.6.2] - 2025-08-26​

Added - Platform​

  • Default GPU memory utilization with smart middle-value selection
  • Database status indicators and loading bars in file upload modal
  • New deploy model table with enhanced visualization
  • Support for editing paused service configurations with config pasting

Changed - Platform​

  • Enhanced model configuration interface with improved GPU memory sliders
  • Better connector visual integration in advanced configuration settings
  • Improved default model listing

Fixed - Platform​

  • Endpoint title clickable area being larger than expected
  • CNPG Database backup parameters being ignored in vector database deployments

[2.6.1] - 2025-08-20​

Added - Platform​

  • ClickHouse metrics table creation on startup for better observability
  • /podhealth endpoint for real-time pod status checking
  • /usage endpoint for comprehensive token usage metrics logging
  • Robust WebSocket client management and connection tracking
  • Automatic fail-safe reconnection with gap recovery for realtime service
  • PostgreSQL connection pooling (max: 3 connections) with optimized timeouts

Changed - Platform​

  • Complete refactoring of realtime service to global database listener pattern
  • Replaced per-client listeners with centralized global listener for better scalability
  • Enhanced deployment service metadata overrides functionality
  • Improved connector connection and deployment logic
  • Enhanced graceful shutdown handling to prevent reconnection loops
  • Reduced log flooding with cleanup of verbose manager logs

[2.6.0] - 2025-08-09​

Added - Platform​

  • Token/Embedding usage counting for RAG services
  • API key filtering for Request/Token counts
  • Business logic integration capabilities for usage tracking
  • Import/Export configs functionality
  • Ability to create desired spaCy RAG configurations
  • Reusable configurations for new RAG service creation
  • Better deployment state indication for RAG service dependencies (database boot-up)
  • Temporal RAG features available through RAG config

Changed - Platform​

  • Improved model deployments with better out-of-the-box functionality
  • Enhanced model information
  • Updated Agent, RAG, and Model Endpoint versions
  • Multiple user experience enhancements
  • Improved navigation and interface design

Added - CM-RAG v0.6.2​

  • Enhanced usage metrics integration with comprehensive token and embedding tracking
  • Improved spaCy-based retrieval for temporal queries

Added - CM-Model-Endpoint v0.0.8​

  • Integrated usage metrics system for token counting and API key filtering
  • Enhanced request tracking and analytics capabilities

[2.5.0] - 2025-07-09​

Added - Platform​

  • Enhanced OpenTelemetry (OTEL) integration with ClickHouse
  • Distributed tracing support with TraceID and SpanID
  • Workload-specific log collection and processing
  • Portal features for tracing and logs
  • Automated log retention and cleanup mechanisms
  • Enhanced JWT payload processing for API key tracking
  • Improved health check filtering
  • Enhanced GPU node tolerations for log collectors
  • Better request filtering for API services

Changed - Platform​

  • Improved structured logging with JSON parsing
  • Enhanced log filtering and correlation capabilities
  • Better log parsing and attribute extraction
  • Optimized log storage and retrieval performance
  • Manager and Portal upgraded to v0.3.36
  • Model Endpoint upgraded to v0.0.7
  • Latest RAG and Agent service versions
  • Realtime service version improvements
  • Optimized database setup and configuration

Added - CM-Model-Endpoint v0.0.7​

  • Support for Agents and RAG systems

Added - CM-RAG v0.0.3​

  • Multimodal ingestion with vision model (can handle charts and images in pdf files)
  • Initial support for temporal retrieval
  • Structured logging

[2.4.0] - 2025-06-30​

Added - Platform​

  • Enhanced CloudNative PostgreSQL (CNPG) cluster redundancy
  • Scaled up CNPG clusters to 3 replicas for high availability

Changed - Platform​

  • Improved database backup and recovery mechanisms
  • Enhanced vector database chart with better redundancy
  • Manager and Portal upgraded to v0.3.25
  • Model Endpoint upgraded to v0.0.5
  • RAG and Agent service template improvements
  • Vector database upgraded to pgvector 0.8.0
  • Improved variable naming consistency
  • Enhanced API services Istio injection configuration
  • Better default configurations for stack-admin and realm-manager
  • Streamlined repository authentication setup
  • Enhanced container registry authentication
  • Improved namespace and service mesh configurations
  • Better integration with Azure container registry

Fixed - Platform​

  • Various deployment pipeline issues

[2.0.5] - 2025-06-12​

Added - Platform​

  • In-portal request metrics for endpoints (request counts, error rates, latency metrics)
  • P50, P90, and P99 latency metrics
  • Configurable OTEL Collector for logs
  • ClickHouse DB support

[2.0.4] - 2025-05-26​

Added - Platform​

  • Ability to pause API keys
  • Endpoint limits for permission management
  • Portal support for top-level access management
  • Ability to allow on-the-spot user creation (defaults to model viewers)

Changed - Platform​

  • Improved release pipeline

Fixed - Platform​

  • Fixes for ConfidentialMind stack-deployer

Added - CM-RAG v0.0.2​

  • Support for embedding/re-ranker load balancing
  • Open WebUI context example

Changed - CM-RAG v0.0.2​

  • Improved database query performance

[2.0.3] - 2025-05-12​

Added - Platform​

  • Enhanced AMD support with AMD-ready model templates
  • AMD benchmarks and ROCm optimization
  • AMD CM-RAG and CM-Agent tests

Fixed - Platform​

  • Installer fixes

Added - CM-Agent v0.2.5​

  • "How to get started" examples
  • Ability to run MCP servers locally & communicate over HTTP

Changed - CM-Agent v0.2.5​

  • Restructured for clear MCP pattern

Added - CM-Model-Endpoint v0.0.4​

  • Routing through ConfidentialMind service ID

Changed - CM-Model-Endpoint v0.0.4​

  • Improved model naming
  • Improved Open WebUI support

[2.0.2] - 2025-04-23​

Added - Platform​

  • New dedicated Model and Agent Endpoint UI in portal
  • More granular access control for model endpoints
  • Separation between access to models vs model deployment management

Changed - Platform​

  • Models now controllable separately from model endpoints
  • Improved model management and deployment workflow

Added - CM-Agent v0.0.1​

  • Powerful AI agent interface with OpenAI-compatible API
  • Database interaction through natural language
  • Integration with RAG systems for information retrieval
  • Access to intranets and private information sources
  • Action performance capabilities
  • Multi-step request processing through LangGraph

[CM-Model-Endpoint-0.0.2] - 2025-04-10​

Added - CM-Model-Endpoint v0.0.2​

  • OpenAI-compatible API for language model access
  • Support for text and image inputs (multimodal)
  • Structured output generation with guided_json
  • Multiple model access through unified interface
  • Streaming response capabilities

[2.0.1] - 2025-03-24​

Added - Platform​

  • Better model deployment assistant with VRAM estimation
  • Shows active free VRAM on selected GPUs
  • Deployment failure prediction with current configuration
  • Track GPU usage on a per-pod basis
  • Historical record of GPU metrics for deeper insights

Added - CM-RAG v0.0.1​

  • Optimized database structure
  • Automatic database migration during version update
  • Group enforcement settings
  • Improved filtering capabilities

Changed - CM-RAG v0.0.1​

  • Performance improvements in the internal logic

[2.0.0] - 2025-03-10​

Added - Platform​

  • Simple admin portal with RAG Endpoint and Thinking Endpoint
  • Optimal template configs for use cases (Internal knowledge assistant, Semantic search, Ask Questions From Document)
  • Data connectors to sync data to AI Systems
  • In portal API usage guide
  • Automatic deployment of database and other required services
  • Automatically connecting the right models to the endpoints
  • Advanced and basic UI toggle (old/new version of portal)
  • Active VRAM usage per node and GPU
  • Select specific node and GPU on model deployment

Added - CM-RAG v0.0.0​

  • Simple endpoint to push files
  • OpenAI style API chat endpoint that answers with the internal knowledge
  • Multiple improvements over "naive" RAG implementations

Added - S3 Connector v1.0.0​

  • Connect data from S3 bucket to RAG deployment
  • Automatic file sync
  • Filter to specific files

[1.1.0] - 2025-02-10​

Added - Platform​

  • Simple admin portal with RAG and ASG endpoints
  • Data connectors to sync data to AI Systems
  • Use case specific pre-configured deployments
  • Advanced and basic UI toggle (old/new version of portal)

Added - CM-RAG v0.0.1 (Beta)​

  • Simple endpoint to push files
  • OpenAI style API chat endpoint that answers with the internal knowledge
  • Multiple improvements over "naive" RAG implementations

Added - S3 Connector v1.0.0 (Beta)​

  • Connect data from S3 bucket to RAG deployment
  • Automatic file sync
  • Filter to specific files

[1.0.10] - 2025-01-22​

Added - Platform​

  • Optimized inference deployment and replication with weight cache
  • Customizable timeout for deployment pipelines

Changed - Platform​

  • Template updates and better search

[1.0.9] - 2025-01-06​

Added - Platform​

  • Base-stack helm AMD GPU inference deployments
  • AMD GPU Operator
  • rocm/vllm support for inference
  • Templates for Llama 3.1 on AMD
  • Benchmarks on AMD MI300X (8xGPU)

[1.0.8] - 2024-12-16​

Added - Platform​

  • Optimized inference deployment and replication with weight cache
  • Customizable timeout for deployment pipelines

Changed - Platform​

  • Template updates and better search

Added - rag-service v2.0.0​

  • Improved file management (group, metadata, filtering)

Changed - rag-service v2.0.0​

  • Re-chunk only files if they have changed and other performance updates
  • Improved performance for semantic search and other use cases

[1.0.7] - 2024-12-10​

Added - Platform​

  • More configurations to stack-base chart

Changed - Platform​

  • Template updates and better search

Fixed - Platform​

  • Long pipelines error bug

[1.0.6] - 2024-11-28​

Added - Platform​

  • CNPG backup feature support for vector databases
  • Improved API key management with labels
  • API key expiration date
  • New API key management UI
  • API key expiration alerts

[1.0.5] - 2024-11-12​

Added - Platform​

  • More models to select from template UI
  • Automatic JWK Creation on installation tooling

Fixed - Platform​

  • Portal UI bug-fixes

[1.0.4] - 2024-10-25​

Added - Platform​

  • Optional Longhorn CSI and Nvidia Plugin support for enhanced storage and GPU capabilities
  • Updated application cards with refreshed colors and design
  • Enhanced navigation layout and functionality
  • Alpha version of CNPG backup feature support for vector databases

Added - Python SDKs v0.1.3​

  • Query parameter options for app helper

Changed - Python SDKs v0.1.3​

  • Updated to new API key header format using authorization
  • Improved local development workflow

Fixed - Python SDKs v0.1.3​

  • Compatibility issues with RAG interface and vLLM
  • Added missing dependencies

[1.0.3] - 2024-09-25​

Added - Platform​

  • Support for persistent volume claims across workload namespaces
  • Storage class suggestions in portal UI
  • Enhanced volume mount support for web services

Changed - Platform​

  • Updated header format for API key authentication (follows OpenAI API)
  • Improved volume management and storage capabilities

[1.0.2] - 2024-08-16​

Added - Platform​

  • Support for more advanced deployment options
  • Support for API-key enabled communication for local development

Added - Python SDKs v0.1.2​

  • Handling of .env files for local development
  • New functions load_environment and get_api_parameters
  • New application and API examples for getting-started tutorial

[1.0.1] - 2024-06-17​

Added - Platform​

  • Poro 34B chat Model added to easily deployable models

Changed - Platform​

  • Improved support for common GenAI frameworks and SDKs

Fixed - llm-service v1.0.1​

  • Fixed endpoints to more closely match OpenAI's API for better third-party SDK support

[1.0.0] - 2024-06-03​

Added - Platform v1.0.0​

  • Easily deploy GenAI applications, APIs, models and vector databases
  • Connect workflows to existing data sources and external models
  • Enable connectivity between micro services
  • Configure access on per-service level

Added - Python SDK v0.1.0​

  • Communicate with LLMs and various data sources
  • Sync and configure app-specific settings
  • Streamlines management of connectors for enhanced usability

Added - llm-service v1.0.0​

  • Connect to external models in Azure, IBM cloud or AWS
  • Connect to OpenAI API
  • Simple abstraction to streamline communication

Added - chat-template v1.0.0​

  • Simple Streamlit-based template for creating applications
  • Can be connected to any internal or external LLM supported by the stack

Added - rag-service v1.0.0​

  • Enables uploading new documents and querying relevant chunks
  • Configuration for Section Length, Section Overlap, Number of Included Neighbours
  • Connects seamlessly with vector database deployed within the stack

Added - rag-interface v1.0.0​

  • Simple streamlit based UI app that connects to rag-service
  • Allows to upload files and ask questions about them