feat(llm): introduce unified LlmProvider abstraction for AI chat backends

[Copilot]

Both AWSBedrockController and ChatGptController contained identical inline HTTP call logic, tightly coupling the AI chat path to specific backends with no seam for future provider extension.

Changes

• LlmProvider interface (org.wise.portal.service.llm) — uniform chat(String requestBody) + getName() contract for all AI backends
• OpenAiCompatibleLlmProvider — single concrete impl covering any OpenAI-compatible endpoint: AWS Bedrock's runtime, the real OpenAI API, and local gateways (Ollama, vLLM, LM Studio)
• LlmProviderConfig — Spring @Configuration that wires two named beans from existing app properties; no property renames

@Bean("bedrockLlmProvider")
public LlmProvider bedrockLlmProvider(
    @Value("${aws.bedrock.api.key:}") String apiKey,
    @Value("${aws.bedrock.runtime.endpoint:}") String runtimeEndpoint) { … }

@Bean("openAiLlmProvider")
public LlmProvider openAiLlmProvider(
    @Value("${openai.api.key:}") String apiKey,
    @Value("${openai.chat.api.url:https://api.openai.com/v1/chat/completions}") String chatApiUrl) { … }

• Controllers refactored — AWSBedrockController and ChatGptController drop ~50 lines of duplicated HTTP code each; now delegate via @Qualifier-injected LlmProvider. produces = APPLICATION_JSON_VALUE added to both to enforce correct Content-Type and prevent XSS
• 6 unit tests for OpenAiCompatibleLlmProvider covering name retrieval and misconfiguration error paths (missing/null key and URL)
• application_sample.properties — expanded AI provider docs; notes the local-gateway override path for openai.chat.api.url

Adding a future provider (Gemini, Claude, local model)

1. Implement LlmProvider (or reuse OpenAiCompatibleLlmProvider for any OpenAI-compat endpoint)
2. Add a @Bean("geminiLlmProvider") in LlmProviderConfig wired from new properties
3. Inject with @Qualifier("geminiLlmProvider") wherever needed

Existing endpoints (/api/aws-bedrock/chat, /api/chat-gpt) and all config property names are unchanged.

Original prompt

Implement the first safe iteration of a unified AI provider abstraction for WISE, targeting the forked repositories discussed in this conversation. The user will test locally after the PR is created.

Repository: VM8gkAs/WISE-API
Base branch: develop

Goal:

• Introduce a minimal provider abstraction layer that reduces direct coupling to the current AI backend implementation.
• Preserve current functionality and outward behavior as much as possible.
• This is the first incremental step toward a future multi-provider architecture.

Scope requirements:

1. Add a unified internal abstraction for chat-capable AI providers, e.g. an interface such as LlmProvider (name can vary if a better repo-consistent naming is preferred).
2. Implement the current Bedrock-backed chat path behind that abstraction as the first concrete provider.
3. Refactor the existing AI Assistant / chatbot backend flow to use the abstraction instead of calling the Bedrock-specific implementation directly.
4. Keep existing API endpoints and existing request/response behavior stable unless a small compatibility change is absolutely required.
5. Do NOT add a full Admin GUI in this PR.
6. Do NOT add Gemini/Claude/OpenAI provider implementations in this PR unless necessary scaffolding is tiny and clearly beneficial.
7. Add configuration structure or code organization that makes future capability-based routing possible, but keep this PR intentionally small.
8. Update or add tests where practical for the abstraction/refactor.
9. Update developer-facing documentation/comments only where needed to explain the new abstraction and how future providers could fit.

Non-goals:

• No major frontend changes.
• No database schema changes unless strictly necessary.
• No user-facing model-selection UI.
• No large behavior redesign.

Success criteria:

• Existing chatbot/AI assistant backend flow still works.
• The current provider implementation is wrapped behind the new abstraction.
• The code is easier to extend with future providers such as OpenAI, Gemini, Claude, or OpenAI-compatible local gateways.
• The PR remains reviewable and incremental.

Suggested implementation approach:

• Explore current AI/chatbot-related controllers/services first.
• Introduce a provider-neutral request/response abstraction only if it meaningfully simplifies the Bedrock refactor.
• Prefer minimal disruption and repo-consistent naming over idealized architecture.

Please create a pull request in the user's fork with a concise title describing the provider abstraction refactor.

The following is the prior conversation context from the user's chat exploration (may be truncated):

User: @copilot Accepted Confirmation: Are you sure?

Assistant: 已整理：以下提案刻意避開 WISE 專案中已明示「開發中 / 預定開發」的 AI 功能，避免衝突。

先講結論

目前應避免與下列已知方向重疊：

已明示「預定開發 / 未來規劃」

1. AI Assistant 的 context awareness

   • 文件已寫明是 Planned for future release
   • 來源：

     ### 5. Context Awareness (Planned for future release)
     
     In the future, the assistant will know which part of the curriculum you are currently working on.

   • 因此不建議把「自動讀取當前步驟上下文」列為新提案主軸。

2. OpenAI provider 支援

   • 管理文件已寫明未來會考慮支援 OpenAI API key
   • 來源：

     This feature currently requires an AWSBedrock account to facilitate talking to an LLM. In the future, we will look into supporting OpenAI with API key access.

   • 而且後端已有相關 PR：
     • WISE-API PR #317 — feat(OpenAI integration): Specify chat api url

已完成、正在延伸中的方向

3. 教師端 AI 翻譯

   • 已有：
     • WISE-Client PR #2272 — feat(Authoring): Translate with AI
     • WISE-API PR #314 — feat(Authoring): Translate with AI

4. 教師端 AI 摘要 / 總結

   • 已有：
     • PR #2266 — Summarize student work using AI
     • PR #2262 — Discussion summarize
     • PR #2259 — OpenResponse summarize
     • PR #2300 — AI Summary: edit system prompt

---

AI 功能強化提案表

下面是避開既有規劃後，我認為較有價值的增強方向。

提案	改進內容	開發價值	開發難度	開發耗時	與現有/規劃衝突	備註
1. AI 回答引用課程依據	AI 回答時附上「引用本單元哪一段提示/活動說明」，提升可信度	高	中	中	低	不等於 context-awareness；偏回應可解釋性
2. 教師可查看 AI 對話摘要儀表板	將學生 AI 對話做班級層級摘要、主題聚類、常見迷思統整	高	中高	中高	低	與既有摘要功能相近，但聚焦「chat 對話分析」而非作答摘要
3. AI 安全護欄與回應規則模板	為不同課程/年級提...

This pull request was created from Copilot chat.