[RFC] Rethink course language: from single enum to LLM-inferred language directive

[cosarah]

Problem / 问题

The current course language system uses a single language field (zh-CN or en-US) that forces all generated content into one language. This is limiting for many real-world scenarios:

Scenario	User expectation	Current behavior
Chinese user learning English	English slides with Chinese explanations when needed	All English OR all Chinese
Japanese user uploads English paper	Japanese teaching, English source material	Cannot express "source ≠ teaching language"
Bilingual user	Technical terms in English, explanations in Chinese	No mixing support
Non-zh/en user	Korean teaching with English technical terms	Only zh-CN or en-US available

当前课程语言系统用单一的 language 字段（zh-CN 或 en-US）强制所有生成内容使用同一种语言。这在很多实际场景中存在局限：

场景	用户期望	当前行为
中文用户学英语课	课件英文，解释可夹中文	全英文或全中文二选一
日语用户上传英文论文	用日语讲解英文内容	无法表达"源语言≠教学语言"
双语用户	术语英文，讲解中文	不支持混合
非中英母语用户	韩语上课，技术术语保留英文	只能选中/英

---

Proposed design / 设计构想

Core idea

Replace the single enum with a natural-language directive inferred by the LLM from user input and uploaded documents.

用自然语言指令替代枚举值，由 LLM 从用户输入和上传文档中推断。

Language dimensions / 语言维度

Language requirements are multi-dimensional:

1. User's native language — inferred from the writing language of course requirements
2. Source material language — detected from uploaded PDF
3. Teaching language — what language the classroom should use
4. Terminology strategy — keep original / translate / bilingual
5. User proficiency — inferred from phrasing (e.g. "learn X from scratch" vs "advanced study of X")

语言需求是多维的：

1. 用户母语 — 从课程要求的书写语言推断
2. 源材料语言 — 从上传文档检测
3. 教学语言 — 课堂使用什么语言
4. 术语处理 — 保留原文 / 翻译 / 双语对照
5. 用户水平 — 从措辞推断

Examples / 示例

User input	Inferred directive
"从零学 Python"	Native: Chinese. Teaching: Chinese. Keep code terms in English.
"Explain quantum computing"	Native: likely English. Teaching: English.
"用中文讲解这篇论文" + English PDF	Native: Chinese. Source: English. Teaching: Chinese with English terms preserved.
"英語でPythonを教えてください"	Native: Japanese. Teaching: English.
"帮我准备一堂英语课，初中水平"	Native: Chinese. Teaching: mixed Chinese/English with Chinese scaffolding.

Data flow / 数据流

User input + PDF
    │
    ▼
LLM inference (new step before outline generation)
  Input: course requirements, PDF language, user profile
  Output: languageDirective (natural language, 3-5 sentences)
    │
    ▼
Stored in stage.languageDirective
    │
    ├──→ Outline generation prompt
    ├──→ Page content generation prompt
    ├──→ Agent system prompt (replaces current Language (CRITICAL) section)
    ├──→ Quiz grading prompt
    └──→ PBL prompt

Key design decisions / 关键决策

1. User can review and edit the directive / 用户可查看和编辑

LLM inference isn't 100% accurate. Show the inferred directive before generation and let the user edit it (similar to "Advanced Settings").

LLM 推断不是 100% 准确。在生成前展示推断结果，用户可编辑确认。

2. Backward compatibility / 向后兼容

stage.language is already stored in user's local DB. Add languageDirective as a new field; old classrooms fall back to language.

stage.language 已存储在用户本地数据库。languageDirective 作为新字段，旧课堂继续用 language。

3. Manual selection as shortcut / 手动选择作为快捷方式

Keep manual language picker for quick creation. When used, generate a simple directive directly without LLM inference.

保留手动语言选择作为快捷方式。选择后直接生成简单指令，跳过 LLM 推断。

4. Eliminate hardcoded language branches / 消除硬编码语言分支

Currently there are 15+ === 'zh-CN' checks across the codebase for prompt text. With a natural language directive injected into prompts, most of these can be removed — the LLM decides based on the directive.

当前代码中有 15+ 处 === 'zh-CN' 硬编码判断。自然语言指令注入 prompt 后，大部分可以删除 — 让 LLM 根据指令自行决定。

---

Trade-offs / 权衡

Aspect	Current	Proposed
Data structure	language: 'zh-CN'	languageDirective: string
Source	Manual selection	LLM-inferred + user-editable
Injection	Hardcoded ternary in code	Unified prompt paragraph
Extensibility	Each new language requires code changes	No language enumeration needed
Cost	None	One extra LLM call (or merged into outline step)
Token overhead	~10 tokens	~50-100 tokens per prompt (negligible vs full system prompt)

---

Suggested issue breakdown / 建议的 Issue 拆分

See comment below for individual issues.

---

Feedback welcome! This is a significant architectural change, so I'd like to hear thoughts on the approach before implementation.

欢迎反馈！这是一个重要的架构变更，希望在实现之前听到大家的想法。

[nkmohit]

Love this proposal @cosarah! this would make multilingual content generation much more flexible, since we would no longer need to add new languages through separate code changes. Users could either specify the desired language(s) directly in the prompt, or let the LLM infer it from context.

On timing, language inference before outline generation makes more sense, so the outline itself already follows the intended language pattern. That seems more stable. Merging it into the outline step saves an LLM call, but might risk overloading that step probably worth keeping separate initially.

One thought on the review surface: if outline review ever lands, exposing languageDirective there alongside the outline would let users validate and edit both in one place rather than adding a separate confirmation step. That feels lower-friction than a dedicated language review screen.