Release Notes — WUIC Framework v1.3.2

Date: 18 June 2026 Previous published version: 1.3.0 (11 June 2026) Backend: .NET 10 + IIS / Linux nginx Frontend: Angular 21

A consolidation release on the RAG chatbot introduced in 1.3.0: the conversational model is no longer tied to Anthropic — any OpenAI-compatible endpoint, including local runtimes such as Ollama with open models (Qwen), is now configurable and runs without an API key. Alongside this, a set of fixes to the first-run installer, the source package and metadata scaffolding that surfaced on fresh installations, plus a workspace ready for AI coding assistants.

🤖 RAG Chatbot — flexible LLM provider (including local and free)

The chatbot's conversational model is now provider-agnostic. In addition to Anthropic, OpenAI-compatible endpoints are supported, which includes local runtimes (e.g. Ollama): you can run open, free models such as Qwen on your own machine, without an API key and with no per-token cost.

• rag-llm-provider — anthropic (default) / openai / openrouter. Selects the provider's wire dialect.
• rag-llm-base-url — endpoint override; pointing it at a local server URL (e.g. http://localhost:11434/v1 for Ollama) makes the chatbot talk to the model locally.
• rag-llm-default-chat-model — model id for the chosen provider (e.g. a Qwen model on Ollama).
• llm-api-key — key for the active provider; for local runtimes that don't validate it, a placeholder value (e.g. ollama) is enough. The legacy anthropic-api-key remains valid when rag-llm-provider=anthropic (zero migration).

All keys are hot-reloaded from appsettings.json: switching provider or model requires no restart.

More accurate retrieval — result re-ranking has been refined: the chatbot cites more relevant sources on natural-language queries.

Setup notifications — on first use the .NET engine downloads the ONNX models on demand. The administrator now receives started / ready / error notifications for the download in the bell, across all four DB providers, even when initialization is triggered by a request with no logged-in user.

Automatic GPU acceleration — on a machine with an NVIDIA GPU the engine uses the GPU without installing CUDA: on first launch, besides the ONNX models, it also downloads the required CUDA 12 + cuDNN 9 runtime on demand (~1.8 GB, one time, only if a GPU is present) and wires it up itself. Without a GPU → CPU, no extra download. Manual override with rag-engine-cuda-path.

🧩 Workspace ready for AI coding assistants

Applications generated with the framework now include a set of markdown context files (project description, conventions, operating rules) at the workspace root. These files make agentic AI assistants — Continue, Cline, Cursor and similar — immediately aware of the WUIC structure and conventions, with no proprietary extension to install. Any client that reads the workspace context behaves as a "WUIC-native" assistant.

🐛 Notable bug fixes

• First-run installer — non-tutorial mode on every DB provider: installing with scaffolding of an existing database (without the tutorial sample data) has been fixed and unified across all supported providers — SQL Server, MySQL, PostgreSQL and Oracle. Resolved the failures caused by SQL dialect differences, target database/schema selection and connection handling that surfaced outside tutorial mode.

• First-run installer — SQL script path (non-BAK): when provisioning the metadata DB via the incremental SQL script (the alternative to restoring from a .bak), the parser for GO-separated batches mishandled some separators, causing schema creation to fail on fresh installations. The splitter has been fixed and script-based installs now complete cleanly.

• Source package — .NET RAG engine not found at runtime: in the source package (-src-) the WuicRagEngine.dll engine was placed at the package root, while the executable, started from bin/, looked for it next to itself — the RAG chatbot would not start ("WuicRagEngine.dll not found"). The loader now searches the rag-engine/ folder in several locations (build output, content root, working directory) and finds the engine in both deploy layouts.

• First-run — chatbot API key persistence: the LLM key entered in the first-install wizard is now written to the canonical appsettings.json actually read by the runtime. Previously, in some layouts, it could land in a copy the process never reads, leaving the chatbot without a key right after install.

• Metadata scaffolding — diagnostics and robustness: scaffolding the metadata for certain tables could fail with a generic message ("Unable to scaffold metadata table") that masked the real cause. The actual SQL error now propagates to the caller, and the case that triggered it is fixed.

• Source package — realtime notifications in dev: in the -src- package the dev-server (ng serve) proxy did not forward WebSocket connections to the backend; the notification channel (/ws) timed out and updates only appeared after a manual page reload. The proxy now forwards WebSockets too: notifications arrive in real time.

📦 Updated packages

🔧 Recommended operational updates for upgraders

1. To run the chatbot with a local, free model (e.g. Qwen via Ollama): set rag-llm-provider=openai, rag-llm-base-url to the local endpoint (e.g. http://localhost:11434/v1) and rag-llm-default-chat-model to the model id; set llm-api-key to a placeholder (e.g. ollama) if the runtime doesn't validate it. No restart: the keys are hot-reloaded.
2. To stay on Anthropic, no action is needed: anthropic-api-key keeps working with rag-llm-provider=anthropic (default).
3. The source (-src-) package is lighter: it no longer includes the redundant framework DLLs at the root, which are recreated by dotnet build from the NuGet packages. Downloading the new -src- requires no action.
4. On first chatbot use with the .NET engine, the administrator will see the ONNX model download progress in the bell. Wait for the "ready" notification before the first Ask.
5. New apps generated by the framework automatically include the AI-assistant context files at the workspace root; for existing apps they can be regenerated.