Release Notes

Release Notes: BullSequana AI v1.2.1

Release date: August 6, 2026

Version: 1.2.1

Executive summary

This is a patch release on top of 1.2.0. API keys now work as standard Authorization: Bearer tokens through the platform gateway, so OpenAI-compatible tools and SDKs connect without workarounds, and the updated CoreAI API backend delivers the Anthropic Messages API and streaming model downloads. LiteLLM gains a 600-second request budget for long generations, Redis-backed response and authentication caching, and high-availability hardening. Kubeflow default pipelines and MLflow notebook credentials are fixed for all profile namespaces. The CoreAI API backend is updated to 1.2.9 and the portal to 1.2.3.

Fixes & improvements

• API keys as Bearer tokens through the gateway — requests carrying a platform API key as Authorization: Bearer sk-bsq-... are now routed past gateway-level JWT validation and authenticated by the CoreAI API itself. Previously the gateway rejected these requests with Jwt is not in the form of Header.Payload.Signature before they reached the backend, which is why 1.2.0 recommended the X-Api-Key header for clients behind the gateway. That workaround is no longer required: OpenAI-compatible SDKs and tools that hardcode the Authorization: Bearer header now work natively. X-Api-Key remains fully supported. Invalid or malformed keys are still rejected by the backend, and JWT validation at the gateway is unchanged for all other bearer tokens. See Use local models via API.

• API keys accepted in the interactive API docs — the Swagger (/docs) and Scalar (/scalar) UIs now authenticate correctly with platform API keys, and the X-Api-Key security scheme resolves on WebSocket routes in the generated API documentation.

• Anthropic Messages API compatibility — the CoreAI API now serves an Anthropic-compatible endpoint at /anthropic. Tools that speak the Anthropic wire format, such as Claude Code and the Anthropic SDK, can use platform-hosted models with platform authentication. No Anthropic account is required. The endpoint supports streaming and non-streaming turns, model and agent configuration resolution, token counting, and a per-user model catalogue. See Claude Code.

• Streaming model downloads — model installer downloads from Hugging Face and S3-compatible storage now stream directly to the destination instead of buffering the full artifact in memory. This reduces memory pressure and improves reliability for large model imports.

• LiteLLM reliability and performance — LiteLLM is upgraded to 1.89.4 (mirrored in the private registry) with several operational fixes:

  • Gateway route and proxy request timeouts are raised to 600 seconds, so long generations and streaming responses are no longer cut off mid-request by the previous short gateway timeout.
  • A Redis instance now backs a response cache and a shared authentication cache, keeping virtual-key state consistent across workers and replicas.
  • Worker count is pinned per pod instead of derived from node cores, preventing CPU oversubscription on large nodes.
  • Replicas spread across nodes, health probes are defined for startup, liveness, and readiness, and failed upstream calls retry up to three times.

• Kubeflow default pipeline and MLflow notebook credentials — expanded Kyverno policies now generate the MLflow notebook credentials (secret and PodDefault), registry pull-secret PodDefault, pipeline S3 credentials, per-profile MLflow provisioning job, and scoped RBAC in every profile namespace — both existing namespaces and newly created profiles. This fixes the default Kubeflow pipeline failing to run and notebooks failing with a missing MLflow secret.

• Reranker model mode — reranker models are now recognized as a dedicated rerank model mode across the backend model catalogue and the portal, fixing reranker model selection in Global Configuration.

• Self-contained backend image builds — the backend container image now builds without internet access and multi-architecture image builds are fixed, supporting offline and air-gapped delivery pipelines.

• Security patches — the backend 1.2.9 image patches CVE-2026-69244, CVE-2026-69247, CVE-2026-69249, and GHSA-3f7w-8rr8-f37f; the portal 1.2.3 image patches Next.js and PostCSS dependency vulnerabilities.

• CoreAI API backend 1.2.9 — the release moves the backend from 1.2.5 to 1.2.9, which carries the API-key authentication paths, the Anthropic Messages API, streaming model downloads, the reranker model mode, and the build and security fixes above.

• CoreAI Web Portal 1.2.3 — carries the reranker model mode fix and the dependency security patches above.

• Speech-to-text GPU model corrected — the default GPU Whisper model is now deepdml/faster-whisper-large-v3-turbo-ct2, a CTranslate2 build compatible with FasterWhisper. The previous default referenced a model artifact the GPU speech server could not load.

• Platform CLI: optional temp-file retention — setting CLEANUP_TEMPORARY_FILES=false keeps the rendered Helm values files and the cloned GitOps repository directory after a run for debugging. Cleanup remains the default. See Platform CLI.

Platform deployment changes

• New litellm-redis StatefulSet — the litellm namespace gains a Redis StatefulSet with a persistent volume and a generated connection secret. Argo CD is configured to ignore server-side-apply drift on its volumeClaimTemplates.
• New API-key HTTPRoute — the CoreAI API gateway configuration adds a dedicated HTTPRoute matching Authorization: Bearer sk-bsq-* requests, covered by a CORS-only security policy instead of the JWT policy.
• No manual migration steps are required. Deployments pick up the changes through the standard render and GitOps sync flow.

API updates

No breaking changes in this release.

Authentication — API keys in the Authorization: Bearer header (behavior fix)

API keys are now accepted through the gateway in both header styles:

Authorization: Bearer sk-bsq-v1-<key>
X-Api-Key: sk-bsq-v1-<key>

In 1.2.0, only X-Api-Key worked for clients calling through the platform gateway. As of 1.2.1 both forms are equivalent end to end; when both headers are present, X-Api-Key still takes precedence. WebSocket connections continue to use the token query parameter.

Anthropic Messages API (new)

A new Anthropic-compatible API surface is available at /anthropic.

Authentication uses the same X-Api-Key or Authorization: Bearer headers as the main API. Model ids must use the claude-model-<name> or claude-agent-<uuid> prefix for Claude Code compatibility. Direct SDK and curl callers may also use bare model names.

The models endpoint is scoped per user: agent configurations the caller cannot access are not listed, and requesting one returns 404 rather than 403.

See Claude Code for configuration and usage.

Documentation updates

Release Notes: BullSequana AI v1.2.0

Release date: July 21, 2026

Version: 1.2.0

Executive summary

This release adds file and GPT sharing, a redesigned Service Desk configuration, admin branding customization, new settings pages for resource profiles and API keys, and significant improvements to the model deployment and setup workflows.

It also delivers a substantial platform overhaul: Python-native manifest generation replaces Terraform/OpenTofu, Argo CD adopts an app-of-apps topology, networking moves to Kubernetes Gateway API, object storage moves to Rook-Ceph, and deployment security and validation are strengthened.

What's new ?

• Custom GPT sharing — share any Custom GPT with users or Keycloak groups from the GPT card. Assign Viewer (use only), Editor (use and edit), or Owner (full control) roles. Only the GPT owner or an org admin can open the share panel. See CoreAI Portal Guide — Sharing a GPT.

• File sharing — share files and folders with users or groups from the Files area. A My Files / Shared toggle switches between owned and shared content. Subfolder URLs are bookmarkable and persist across renames. Shared files with RAG Available status appear in the chat RAG picker. See CoreAI Portal Guide — Sharing files.

• Resource Profiles settings — new settings page listing all infrastructure resource profiles available for model deployment. Operators can enable or disable individual profiles to control what appears in Easy Setup and Advanced Setup. Requires can_manage_models.

• API Keys settings — new settings page for creating personal API keys for programmatic access. The full key secret is shown only once at creation time and is scoped to the authenticated user. See CoreAI Portal Guide — API Keys.

• Admin Branding — new settings page (requires can_edit_config) for customizing portal appearance. Includes a color token editor with WCAG 2.1 contrast badges for light and dark mode, a border radius slider, logo upload, CSS/tweakcn JSON import, and a live preview panel with Theme, Chat, Files table, and GPT Cards tabs. Applied themes sync instantly to all open browser tabs. See CoreAI Portal Guide — Branding.

• Python-native platform deployment — platform configuration and manifest generation now use Python 3.12, TOML, and Jinja2. The CLI renders Kubernetes resources, resolves environment overrides, seals secrets, generates Argo CD applications, and pushes the output to the GitOps repository.

• Argo CD app-of-apps architecture — applications are organized into ordered common, coreai, and proai parent applications. Automated pruning, self-healing, retries, server-side apply, and component-specific drift handling are enabled.

• Kubernetes Gateway API networking — Envoy Gateway and HTTPRoute resources replace APISIX and the default Ingress NGINX path. JWT authentication, CORS policies, cross-namespace ReferenceGrant resources, and Gateway API routing now protect platform services.

• Rook-Ceph storage platform — MinIO-based object storage has been replaced by Rook-Ceph RGW. Platform services now receive dedicated object-store users and policies. Ceph dashboard SSO, TLS, IAM bootstrapping, orphan cleanup, and dedicated Ceph CSI drivers are included.

• KServe model serving — KServe and its CRDs have been added to the CoreAI application stack, including Ceph RGW integration for model storage.

• Speech-to-text infrastructure — the backend deployment now includes Speaches-based Whisper inference, configurable CPU or GPU execution, separate default models for each mode, and persistent model pre-downloads.

• Kubeflow integration preview — the Kubeflow deployment was rebuilt around Kustomize and Jinja templates with Keycloak SSO, Istio Ambient, profile-level RBAC, Kyverno policies, per-profile artifact buckets, and customized dashboard, notebook, and pipelines images. The component remains under active rework.

Improvements & enhancements

• Service Desk moved to Global Configuration and redesigned — Service Desk is now a subsection of Global Configuration (the standalone sidebar entry has been removed). Operators select a Custom GPT to drive the conversation instead of a raw model name, and choose an optional retrieval folder from the Files library. The enable toggle requires a Custom GPT to be selected first. A custom icon (with crop dialog) and localizable display name and welcome text can be configured per language (English, French, German, Swedish). The end-user chat experience is locked down: no model selector, no file upload, no RAG or web search controls. See CoreAI Portal Guide — Service Desk.

• Service Desk chat session continuity — the Service Desk chat now resumes the user's most recent session automatically on open, rather than always starting a new conversation.

• Global Configuration — several new operator controls have been added to the Global Configuration settings page:

  • Reranker controls — select the reranker model and set Reranker Top N (1–50). The toggle requires a model to be selected before it can be enabled.
  • Retrieval K — new setting (1–50, default 3) controls the number of document chunks passed to the model as RAG context.
  • Embedding reindex progress — changing the embedding model or dimension triggers a reindex with live progress tracking (completed, failed, skipped, total). Uploads during reindex are routed to the target collection so search continues without interruption. Operators can cancel a running reindex at any time.
  • Reset to default — restores all system configuration settings to their defaults. A confirmation dialog is shown before applying.

  See CoreAI Portal Guide — Global Configuration.

• Advanced Setup — multi-step form — the Advanced Setup deployment flow has been restructured into a sectioned multi-step form with a stepper component, clearer section grouping (model information, deployment configuration, scaling, timeout, advanced), and a live YAML preview of the resulting kubeai.org/v1 Model resource.

• Easy Setup — loading and error states — Easy Setup now shows a skeleton loading state while presets load and an inline error state with a Retry button if loading fails, instead of a blank page.

• Model Presets — live backend CRUD — Model Presets settings now reads and writes directly through backend API endpoints instead of portal server actions. Create, edit, and delete operations are immediately reflected without page reloads.

• Model downloader — live download status — the downloader shows a live status indicator. If a download is already in progress when the downloader is opened, the existing status is restored automatically. Download errors from the Model Installer are surfaced in the UI.

• Model repository detail — inline error recovery — if the MLflow model detail page fails to load, an inline error state with a Retry button is shown without a full page reload.

• Homepage — refreshed design — the portal homepage has a new animated design with theme and locale switchers and a version number display. The login flow is handled directly from the homepage without a separate auth page.

• Permission gating — Chat, Files, Custom GPTs, and Settings now show a dedicated "permission denied" view for users lacking can_access_api instead of a broken page. The Models sidebar entry shows a tooltip when can_manage_models is not granted. Settings items requiring elevated permissions are disabled with a tooltip rather than hidden.

• Helm, Kustomize, and plain YAML support — the platform renderer now handles all three component formats with format-specific validation.

• Typed configuration overrides — environment overrides are converted to their corresponding TOML types, including booleans, numbers, lists, and objects.

• Automated secret sealing — rendered Kubernetes Secrets are converted to SealedSecrets offline. The CLI can bootstrap the sealing key and registry pull secret directly into the cluster.

• Keycloak integration — added reusable SSO template generation and CA bundle integration for platform components.

• Backend deployment consolidation — Model Installer functionality and storage configuration have moved into the LLM backend deployment while retaining the existing API routes. Portal-owned database and model-preset bootstrap resources were removed.

• Model installer fixes — S3 file deletion during model removal now works correctly. The MLflow artifact path is correctly included in S3 paths. Model names are sanitized to prevent invalid Kubernetes resource names.

• MLflow improvements — MLflow was upgraded to 3.13.0 with a custom OIDC image and S3 artifact proxy support.

• MLflow deployment reliability — corrected provisioner service-account permissions and moved database setup Jobs to explicit sync-wave ordering.

• MLflow artifact storage — MLflow experiment artifacts are served through the tracking server and stored in the Ceph RGW S3 backend.

• Milvus modernization — replaced the previous integration with the Python-native manifest pipeline and upgraded Milvus to v2.5.16 to address compatibility and security issues.

• Ceph storage and CSI support — added Ceph CSI drivers for RBD and CephFS. The current platform provisions a Ceph RBD block-storage class and Ceph RGW object storage; CephFS-backed workloads require additional filesystem and StorageClass configuration.

• CloudNativePG backups — scheduled database backups now target Ceph RGW and include backup rotation, dedicated credentials, and configurable database resources.

• Observability improvements — Loki retention is enabled for 30 days, and Grafana, Loki, Tempo, Alloy, and OpenTelemetry configurations were refreshed.

• Deployment ordering — sync waves were revised across the platform, unnecessary wait jobs and PreSync hooks were removed, and Argo CD drift handling was expanded.

• Certificate and DNS ordering — corrected sync-wave ordering for ExternalDNS, the self-signed CA chain, certificates, and Sealed Secrets.

• Security refresh — Argo CD, Argo Events, CloudNativePG, Grafana, KubeAI, Milvus, MLflow, OpenBao, OpenFGA, OAuth2 Proxy, Reflector, and Temporal received vulnerability-related upgrades. Argo Events moved from NATS Streaming to JetStream.

• CloudNativePG DatabaseRole CRDs — per-service PostgreSQL role provisioning now uses declarative DatabaseRole custom resources instead of imperative psql Jobs. The CNPG operator reconciles roles directly, reducing boilerplate and removing the need for superuser-credential mounting. Permission-granting Jobs remain, with backoffLimit increased from 3 to 5 for resilience.

• Platform component upgrades — upgraded Grafana to 13.0.1-security-01, OpenBao to 2.5.4, KubeAI to v0.23.2, the CloudNativePG operator to 1.30.0 (chart 0.29.0), and the PostgreSQL image to 17.10-system-trixie.

• Kubeflow dependency management — added Metacontroller as a Kubeflow Pipelines dependency, externalized image references, and standardized per-profile bucket naming.

• Container supply chain — additional platform and Kubeflow images are mirrored into the private ACR, with automated image builds and Trivy scanning.

• PR deployment validation — infrastructure pull requests now receive isolated AKS clusters, full GitOps deployment validation, Argo CD health monitoring, automatic diagnostics, and cleanup.

• Helm security scanning — a Python-based scanner renders platform charts, discovers referenced images, runs Trivy, publishes reports, and supports scheduled scanning of release branches.

Platform deployment changes

• Breaking change: Terraform/OpenTofu is no longer the supported platform deployment path. Deployments now use the Python CLI and GitOps workflow.
• Breaking change: APISIX has been removed and Ingress NGINX is no longer enabled by default. External routes must use Kubernetes Gateway API resources.
• Transition note: Ingress NGINX is deprecated and disabled by default but remains available as an opt-in component for transitional deployments and proofs of concept.
• Breaking change: MinIO Operator and MinIO Tenant have been removed. S3-compatible workloads now use Rook-Ceph RGW.
• Breaking change: The standalone Model Installer deployment has been removed. Its functionality is now part of the LLM backend, while the existing /v1/model-installer API routes remain available.
• Operational change: Generated manifests are organized under common, coreai, and proai app-of-apps directories and should be regenerated with replace mode during migration.

API updates

This section lists all user-facing API surface changes. Breaking changes are marked explicitly.

STT WebSocket endpoint (new)

A new WebSocket endpoint is available for real-time speech-to-text transcription.

WS /v1/stt

Query parameters:

Breaking change: The query parameter was renamed from language to lang. Clients using the old name receive a 1008 Policy Violation close frame with a structured error listing all supported languages.

The endpoint accepts a binary audio stream after an initial JSON handshake and returns transcript frames as JSON messages. The server sends a session_ready frame before audio is accepted. Sending an unsupported language code closes the socket immediately with a structured unsupported_language error frame.

Collection search — structured filters replace raw filter string

Breaking change: The filter field on POST /v1/collections/{name}/search has been replaced by a structured filters array. Raw Milvus filter strings are no longer accepted.

New request shape (partial):

{
  "query": "...",
  "filters": [
    { "field": "metadata_key", "operator": "==", "value": "some_value" },
    { "field": "tag", "operator": "in", "value": ["a", "b"] }
  ]
}

Supported operators: == (scalar string value) and in (list of strings). The == and in operators are mutually exclusive with their value types — mismatched pairs are rejected at the API boundary with a 422 response.

Files API — folder_path replaced by folder_id

Breaking change: All file and folder write endpoints that previously accepted a folder_path string parameter now require folder_id: UUID | null instead. The /v1/files/children and /v1/files/search endpoints also accept a scope parameter.

Affected endpoints: file upload, create folder, move file, move folder, list children, search.

New scope values:

Model download status and logs (new endpoints)

Two new endpoints expose live download progress for model installer jobs. Both require the can_manage_models permission.

GET /v1/model-installer/downloads
GET /v1/model-installer/downloads/{download_id}/logs

GET /v1/model-installer/downloads returns a list of active downloads. Completed and failed downloads are not included.

GET /v1/model-installer/downloads/{download_id}/logs returns buffered log lines for a download. The response is empty once the download reaches a terminal state.

Translations API (new)

A new endpoint exposes tenant-level translation overrides for portal UI strings.

GET  /v1/translations
POST /v1/translations/reset
PATCH /v1/translations/{language}

GET /v1/translations returns all overrides as a nested blob shaped language → namespace → key → value. Requires authentication; no specific permission required.

POST /v1/translations/reset clears all overrides and restores defaults. Requires can_edit_config.

PATCH /v1/translations/{language} merges a partial update for a single language. Requires can_edit_config.

Branding API (new)

New endpoints for admin-controlled portal branding. GET /v1/branding/theme and GET /v1/branding/logo.

GET  /v1/branding/theme
PUT  /v1/branding/theme
POST /v1/branding/theme/preview
POST /v1/branding/theme/reset
GET  /v1/branding/logo
POST /v1/branding/logo

Write endpoints require can_edit_config. Logo upload accepts SVG, PNG, or WebP up to 512 KB.

Group authorization enforcement on all API routers

All API routers now require the can_access_api organization permission. Previously, some endpoints only checked token validity. Users or service accounts in groups without can_access_api receive a 403 Forbidden response on all endpoints.

Service Desk configuration — collection_name replaced by document_ids

Breaking change: The collection_name field in ServiceDeskConfiguration (readable and writable via GET/PATCH /v1/config) has been replaced by document_ids: list[string] | null.

Retrieval configuration — new admin-configurable fields

Two new fields are available in SystemConfiguration via GET/PATCH /v1/config:

Reranker error responses

The reranker now returns structured 400 Bad Request errors (error code rerank_bad_request) instead of a generic 500 when the query or document list is empty or invalid.

Authentication — X-Api-Key header support (new)

All HTTP endpoints now accept API keys via the X-Api-Key request header in addition to the existing Authorization: Bearer header.

X-Api-Key: sk-bsq-v1-<key>

When both X-Api-Key and Authorization: Bearer are present, X-Api-Key takes precedence. This header bypasses gateway-level JWT validation, making it the correct authentication method for API key clients sitting behind a JWT-enforcing gateway. (Superseded in 1.2.1: API keys in the Authorization: Bearer header now pass the gateway too.)

WebSocket connections continue to use the token query parameter. API key format and scoping are unchanged.

Documentation updates