Skip to main content

Release Notes

v2.4.1 — 20 August 2026​

This maintenance release improves the Oracle Backend as a Service (OBaaS) deployment, the mortgage demonstration, and client configuration workflows.

Highlights​

  • Mortgage demonstration: The mortgage corpus, prompts, and Deep Data Security walkthrough have been updated to reflect the current sample data and policy flow.
  • Client reliability: Testbed requests now remain disabled while they run, and clearing a Deep Data Security connect-as selection allows additional time for the server to rebuild its SQLcl connection store.
  • Settings import: Updating saved prompts through settings import now refreshes cached chat sessions so they use the updated prompts.
  • OBaaS deployment: When AI Optimizer is installed as an Oracle Backend as a Service subchart and shares an Autonomous Database Shared instance, it can reuse the wallet resources managed by Oracle Database Base Service.

For the complete set of changes, see v2.4.0...v2.4.1.

v2.4.0 — 10 August 2026​

This release adds Oracle Deep Data Security support, expands model and embedding workflows, and refreshes the installation, deployment, and documentation experience.

Upgrade considerations​

  • macOS x86_64 is no longer a supported installation target. macOS support is limited to Apple silicon (arm64).
  • Existing bare-metal environments should be recreated with uv sync --all-extras before starting the application with uv run python src/entrypoint.py client.
  • The model configuration API now reports runtime availability through status instead of the previous usable Boolean. Possible values are available, unreachable, not_pulled, and no_key.
  • Helm service types are limited to ClusterIP, NodePort, and LoadBalancer.

Highlights​

  • Oracle Deep Data Security: Create and manage data roles, end users, and data grants from the client. Vector Search and NL2SQL can connect as a selected end user so that database-enforced row- and column-level access applies to tool results.
  • Model configuration: Model availability now distinguishes between available, unreachable, missing-key, and not-yet-pulled states. Enabled model endpoints are rechecked in the background, allowing a provider such as Ollama to recover without restarting the AI Optimizer. Ollama discovery, model pulling, duplicate handling, and configuration persistence have also been improved.
  • New local model defaults: granite4.1:8b and mxbai-embed-large are included as disabled Ollama configurations, and the Helm deployment pulls them by default when its bundled Ollama service is enabled.
  • Split and Embed: Local files and OCI Object Storage objects can be grouped into one vector store per filename. SQL query results can also be used as a document source. Embedding-job persistence, cancellation, staging, cleanup, and behavior during CORE database reconnection have been made more robust.
  • SQLcl MCP: SQLcl runs with a configurable restrict level through AIO_SQLCL_MCP_LEVEL, which defaults to the most restrictive level, 4. Saved connections can be recreated when needed, and NL2SQL remains unavailable without preventing server startup when SQLcl is not installed.

Deployment and configuration​

  • Bare-metal installation and development now use uv with a committed lock file. Python 3.11 and 3.12 are supported.
  • The Helm chart can provision optional persistent storage for chart-managed SIDB-FREE and ADB-FREE databases. It also supports a fixed client nodePort and configurable Ollama service type and port.
  • The bundled SigNoz chart dependency has been updated.
  • Settings import continues to accept legacy database configuration fields, while model reachability and Deep Data Security session state are treated as runtime-only configuration.

Documentation​

  • The documentation framework has updated.
  • Installation, configuration, client, server, Helm, observability, TLS, infrastructure, and troubleshooting guidance has been reorganized and rewritten.
  • A new end-to-end walkthrough and an expanded racing use case cover local models, vector stores, RAG, NL2SQL, and Deep Data Security.

For the complete set of changes, see v2.3.2...v2.4.0.