last update

app.py

@@ -64,7 +64,7 @@ API_URL = os.getenv("API_URL") or os.getenv("API_UR", "")
 API_KEY = os.getenv("API_KEY", "")
 USE_MOCK_MODELS = os.getenv("USE_MOCK_MODELS", "0" if (API_URL and API_KEY) else "1") == "1"
 USE_MOCK_TTS = os.getenv("USE_MOCK_TTS", "0") == "1"
-CHAT_MODEL_ID = os.getenv("QWEN_VL_MODEL_ID", "qwen-vl-max")
+CHAT_MODEL_ID = os.getenv("QWEN_VL_MODEL_ID", "gpt-4.1")
 TTS_MODEL_ID = os.getenv("QWEN_TTS_MODEL_ID", "qwen-tts")
 TTS_SPEAKER = os.getenv("QWEN_TTS_SPEAKER", "longxiaochun_v2")
 TTS_FORMAT = os.getenv("QWEN_TTS_FORMAT", "wav")
@@ -94,58 +94,59 @@ QWEN_VL_MCQ_MAX_NEW_TOKENS = int(os.getenv("QWEN_VL_MCQ_MAX_NEW_TOKENS", "1800")
 
 
 DEFAULT_LECTURE_PROMPT_TEMPLATE = """
-你是一名课程助教。请阅读用户上传的论文内容，并输出一段中文讲解，要求：
-1. 先说明论文要解决的问题和背景；
-2. 再解释核心方法（按步骤/模块）；
-3. 再总结实验结果或亮点；
-4. 最后给出局限性与适用场景；
-5. 语言清晰，适合课堂讲解（约 400-700 字）。
-
-论文内容（可能是节选）：
+You are a teaching assistant. Read the uploaded paper content and produce a clear lecture-style explanation in English:
+1. Explain the problem and background first;
+2. Explain the core method step by step / module by module;
+3. Summarize key experiments and highlights;
+4. End with limitations and suitable use cases;
+5. Keep it classroom-friendly (about 400-700 words).
+
+Paper content (may be excerpted):
 {document}
 """.strip()
 
 
 DEFAULT_MCQ_PROMPT_TEMPLATE = """
-请基于下面论文内容，生成 5 道中文单选题（MCQ），用于课堂测验。
-严格输出 JSON（不要 markdown 代码块），格式如下：
+Based on the paper content below, generate 5 English single-choice MCQs for a classroom quiz.
+Output strict JSON only (no markdown code block), in this format:
 {{
   "questions": [
     {{
       "question": "...",
-      "options": ["A选项", "B选项", "C选项", "D选项"],
+      "options": ["Option A", "Option B", "Option C", "Option D"],
       "answer": "A",
       "explanation": "..."
     }}
   ]
 }}
 
-要求：
-1. 共 5 题；
-2. 每题 4 个选项；
-3. answer 必须是 A/B/C/D；
-4. 解析要说明为什么正确，以及常见误区；
-5. 题目应覆盖问题背景、方法、实验/结果、局限性。
+Requirements:
+1. Exactly 5 questions;
+2. 4 options per question;
+3. `answer` must be one of A/B/C/D;
+4. Explanation should tell why it is correct and common mistakes;
+5. Cover background, method, experiments/results, and limitations.
 
-论文内容（可能是节选）：
+Paper content (may be excerpted):
 {document}
 """.strip()
 
 
 DEFAULT_MCQ_RETRY_PROMPT_TEMPLATE = """
-基于以下论文内容生成 5 道中文单选题。只输出合法 JSON，不要任何解释，不要 markdown。
+Generate 5 English single-choice MCQs from the following paper content.
+Output valid JSON only. No explanation outside JSON, no markdown.
 
-限制：
-1. 必须是紧凑 JSON（单行也可以）；
-2. 共 5 题；
-3. 每题 question、options(4项)、answer(A/B/C/D)、explanation；
-4. 解析简短（1-2句），避免过长；
-5. 如果不确定，仍按论文内容出题，不要输出额外文字。
+Constraints:
+1. Compact JSON (single line is fine);
+2. Exactly 5 questions;
+3. Each question includes `question`, `options` (4 items), `answer` (A/B/C/D), `explanation`;
+4. Keep explanations short (1-2 sentences);
+5. If uncertain, still generate based on the paper content only.
 
-输出格式：
+Output format:
 {{"questions":[{{"question":"...","options":["...","...","...","..."],"answer":"A","explanation":"..."}}]}}
 
-论文内容：
+Paper content:
 {document}
 """.strip()
 
@@ -189,6 +190,10 @@ def load_character_configs() -> Dict[str, Dict[str, Any]]:
             if cid in configs:
                 cid = d.name
             avatar_rel = str(meta.get("avatar", "avatar.jpg"))
+            lecture_prompt_path = d / "lecture_prompt.txt"
+            mcq_prompt_path = d / "mcq_prompt.txt"
+            mcq_retry_prompt_path = d / "mcq_retry_prompt.txt"
+            feedback_prompt_path = d / "feedback.txt"
             config: Dict[str, Any] = {
                 "id": cid,
                 "display_name": str(meta.get("display_name", d.name)),
@@ -198,17 +203,25 @@ def load_character_configs() -> Dict[str, Dict[str, Any]]:
                 "chat_mode": str(meta.get("chat_mode", "paper mode")),
                 "avatar_path": str((d / avatar_rel).resolve()),
                 "lecture_prompt_template": _read_text_if_exists(
-                    d / str(meta.get("lecture_prompt_file", "lecture_prompt.txt")),
+                    lecture_prompt_path,
                     DEFAULT_LECTURE_PROMPT_TEMPLATE,
                 ),
                 "mcq_prompt_template": _read_text_if_exists(
-                    d / str(meta.get("mcq_prompt_file", "mcq_prompt.txt")),
+                    mcq_prompt_path,
                     DEFAULT_MCQ_PROMPT_TEMPLATE,
                 ),
                 "mcq_retry_prompt_template": _read_text_if_exists(
-                    d / str(meta.get("mcq_retry_prompt_file", "mcq_retry_prompt.txt")),
+                    mcq_retry_prompt_path,
                     DEFAULT_MCQ_RETRY_PROMPT_TEMPLATE,
                 ),
+                "feedback_prompt_template": _read_text_if_exists(
+                    feedback_prompt_path,
+                    "",
+                ),
+                "lecture_prompt_path": str(lecture_prompt_path.resolve()),
+                "mcq_prompt_path": str(mcq_prompt_path.resolve()),
+                "mcq_retry_prompt_path": str(mcq_retry_prompt_path.resolve()),
+                "feedback_prompt_path": str(feedback_prompt_path.resolve()),
             }
             configs[cid] = config
 
@@ -225,6 +238,7 @@ def load_character_configs() -> Dict[str, Dict[str, Any]]:
             "lecture_prompt_template": DEFAULT_LECTURE_PROMPT_TEMPLATE,
             "mcq_prompt_template": DEFAULT_MCQ_PROMPT_TEMPLATE,
             "mcq_retry_prompt_template": DEFAULT_MCQ_RETRY_PROMPT_TEMPLATE,
+            "feedback_prompt_template": "",
         }
     return configs
 
@@ -234,6 +248,11 @@ DEFAULT_CHARACTER_ID = next(iter(CHARACTER_CONFIGS.keys()))
 
 
 def get_character_config(character_id: Optional[str]) -> Dict[str, Any]:
+    global CHARACTER_CONFIGS, DEFAULT_CHARACTER_ID
+    # Reload from disk so prompt/template edits (e.g. mcq_prompt.txt) take effect immediately.
+    CHARACTER_CONFIGS = load_character_configs()
+    if DEFAULT_CHARACTER_ID not in CHARACTER_CONFIGS:
+        DEFAULT_CHARACTER_ID = next(iter(CHARACTER_CONFIGS.keys()))
     if character_id and character_id in CHARACTER_CONFIGS:
         return CHARACTER_CONFIGS[character_id]
     return CHARACTER_CONFIGS[DEFAULT_CHARACTER_ID]
@@ -698,34 +717,34 @@ class QwenPipelineEngine:
     def _mock_generate_mcqs(self, lecture_text: str) -> List[MCQItem]:
         base_questions = [
             MCQItem(
-                question="这篇论文最主要想解决的问题通常属于下列哪一类？",
-                options=["现有方法存在性能或效率瓶颈", "如何设计数据库索引", "如何搭建前端页面", "如何压缩视频文件"],
+                question="What type of core problem does this paper most likely address?",
+                options=["Performance or efficiency bottlenecks in existing methods", "How to design database indexes", "How to build a frontend page", "How to compress video files"],
                 answer="A",
-                explanation="课程论文阅读类任务通常围绕已有方法不足展开，作者提出新方法来提升性能、效率或鲁棒性。",
+                explanation="Paper-reading tasks usually focus on limitations of prior methods, then propose improvements in performance, efficiency, or robustness.",
             ),
             MCQItem(
-                question="在讲解论文方法时，最合理的组织方式是什么？",
-                options=["按模块或步骤解释输入到输出流程", "只列出参考文献", "只展示实验表格不解释方法", "只讲结论不讲背景"],
+                question="What is the best way to explain a paper's method?",
+                options=["Explain the pipeline from input to output by modules or steps", "Only list references", "Only show experiment tables without method details", "Only present conclusions without background"],
                 answer="A",
-                explanation="课堂讲解需要结构化地说明方法流程，这样听众才能理解论文如何从问题走到解法。",
+                explanation="A structured, step-by-step explanation helps learners understand how the paper moves from problem to solution.",
             ),
             MCQItem(
-                question="生成选择题时，为什么需要同时给出答案和解析？",
-                options=["便于交互反馈与纠错教学", "只是为了让 JSON 更长", "因为 Gradio 要求必须有解析", "为了减少题目数量"],
+                question="Why provide both answers and explanations in MCQs?",
+                options=["To enable feedback and error correction", "Only to make JSON longer", "Because Gradio requires explanations", "To reduce the number of questions"],
                 answer="A",
-                explanation="答案和解析是教学闭环的一部分，错误时给出解析能帮助用户理解常见误区。",
+                explanation="Answer + explanation completes the teaching loop and helps users learn from mistakes.",
             ),
             MCQItem(
-                question="如果论文很长，一次性输入模型的风险是什么？",
-                options=["上下文超长导致成本高、信息丢失或失败", "模型会自动变得更准确", "TTS 音频会变短", "PDF 文件会损坏"],
+                question="What is the risk of feeding a very long paper in one shot?",
+                options=["Context overflow can increase cost and cause information loss or failure", "The model automatically becomes more accurate", "TTS audio becomes shorter", "The PDF file gets corrupted"],
                 answer="A",
-                explanation="长文档通常需要分块总结再汇总，避免超出上下文窗口并降低生成质量波动。",
+                explanation="Long documents usually need chunking and summarization to avoid context-window issues and quality degradation.",
             ),
             MCQItem(
-                question="在这个 Demo 流程中，Qwen TTS 的作用是什么？",
-                options=["把讲解与错题解析转成语音输出", "把 PDF 转成图片", "训练 Qwen3-VL-8B", "生成新的选择题答案"],
+                question="In this demo pipeline, what is Qwen TTS used for?",
+                options=["Convert lecture text and explanations into audio", "Convert PDF to images", "Train Qwen3-VL-8B", "Generate new MCQ answers"],
                 answer="A",
-                explanation="TTS 用于将文本讲解/解析语音化，提高交互演示效果和可访问性。",
+                explanation="TTS turns text explanations into speech, improving interactivity and accessibility.",
             ),
         ]
         return base_questions
@@ -1007,7 +1026,7 @@ class QwenPipelineEngine:
         )
         try:
             mcqs = parse_mcq_json(raw_mcq_json)
-        except json.JSONDecodeError:
+        except (json.JSONDecodeError, ValueError):
             retry_prompt = render_prompt_template(str(mcq_retry_template), pdf_excerpt)
             retry_raw = self._real_generate_text_from_pdf(
                 pdf_path,
@@ -1029,6 +1048,27 @@ class QwenPipelineEngine:
 
 
 def parse_mcq_json(raw: str) -> List[MCQItem]:
+    def _normalize_answer_label(answer_raw: Any, options: List[str]) -> str:
+        s = str(answer_raw or "").strip()
+        if not s:
+            return ""
+        up = s.upper()
+        if up in {"A", "B", "C", "D"}:
+            return up
+        m = re.search(r"\b([ABCD])\b", up)
+        if m:
+            return m.group(1)
+        if up.startswith("OPTION "):
+            tail = up.replace("OPTION ", "", 1).strip()
+            if tail in {"A", "B", "C", "D"}:
+                return tail
+        normalized_answer_text = normalize_option_text(s).strip().lower()
+        if normalized_answer_text:
+            for i, opt in enumerate(options[:4]):
+                if normalized_answer_text == normalize_option_text(opt).strip().lower():
+                    return ["A", "B", "C", "D"][i]
+        return ""
+
     cleaned = strip_code_fence(raw)
     try:
         payload = json.loads(cleaned)
@@ -1039,13 +1079,39 @@ def parse_mcq_json(raw: str) -> List[MCQItem]:
             payload = json.loads(cleaned[start:end + 1])
         else:
             raise
-    questions = payload.get("questions", [])
+    if isinstance(payload, list):
+        questions = payload
+    else:
+        questions = payload.get("questions", []) or payload.get("items", []) or payload.get("data", [])
     parsed: List[MCQItem] = []
     for item in questions[:5]:
+        if not isinstance(item, dict):
+            continue
         q = str(item.get("question", "")).strip()
-        options = [normalize_option_text(x) for x in item.get("options", [])][:4]
-        answer = str(item.get("answer", "")).strip().upper()
-        explanation = str(item.get("explanation", "")).strip()
+        options_raw = item.get("options", [])
+        if not isinstance(options_raw, list):
+            options_raw = item.get("choices", []) if isinstance(item.get("choices", []), list) else []
+        options = [normalize_option_text(x) for x in options_raw][:4]
+        explanation = str(
+            item.get("explanation", "")
+            or item.get("rationale", "")
+            or item.get("reason", "")
+        ).strip()
+        answer = _normalize_answer_label(
+            item.get("answer", "")
+            or item.get("correct_answer", "")
+            or item.get("correctOption", "")
+            or item.get("correct", ""),
+            options,
+        )
+        if not answer:
+            idx_value = item.get("answer_index", item.get("correct_index", None))
+            try:
+                idx = int(idx_value)
+                if 0 <= idx < 4:
+                    answer = ["A", "B", "C", "D"][idx]
+            except Exception:
+                pass
         if len(options) != 4:
             continue
         if answer not in {"A", "B", "C", "D"}:
@@ -1191,35 +1257,60 @@ def _pick_variant(items: List[str], seed: int) -> str:
     return items[seed % len(items)]
 
 
+def _character_feedback_style_from_mcq_prompt(character_id: str) -> str:
+    cfg = get_character_config(character_id)
+    prompt_text = str(cfg.get("mcq_prompt_template", "") or "")
+    if not prompt_text.strip():
+        return ""
+
+    role_line = ""
+    tone_line = ""
+    in_tone_block = False
+    for raw in prompt_text.splitlines():
+        line = raw.strip()
+        if not line:
+            continue
+        lower = line.lower()
+        if not role_line and lower.startswith("you are "):
+            role_line = line
+            continue
+        if lower.startswith("tone:"):
+            in_tone_block = True
+            continue
+        if in_tone_block:
+            # Stop tone parsing when another section starts.
+            if line.endswith(":"):
+                in_tone_block = False
+                continue
+            tone_line = line
+            in_tone_block = False
+
+    style_parts: List[str] = []
+    if role_line:
+        style_parts.append(role_line.rstrip("."))
+    if tone_line:
+        style_parts.append(f"Tone: {tone_line}")
+    return " ".join(style_parts).strip()
+
+
 def _examiner_style_prompt(character_id: str) -> str:
-    cid = (character_id or "").lower()
-    guardrails = (
-        "Do NOT invent magical facts. "
-        "You may use light Hogwarts classroom metaphors (e.g., 'like your Potions class' / marks / exam / detention), "
-        "but do NOT mention spells, incantations, wands, named artifacts, named potions, or made-up magical theory. "
-        "Do NOT mention Harry Potter or any specific character names. "
-        "Keep remarks grounded in the student's performance on this paper/exam."
-    )
-    if "snape" in cid:
-        return (
-            "You are Professor Severus Snape. Cold, cutting, impatient with sloppy thinking; no emojis; no stage directions. "
-            "Sound like Snape: sharp disdain, controlled cruelty, short jabs, second-person address ('you'). "
-            "Avoid generic academic HR language like 'fundamental lack of precision' or 'theoretical principles'. "
-            f"{guardrails} Be brief and exam-focused."
-        )
-    if "mcgonagall" in cid or "mcg" in cid:
+    cfg = get_character_config(character_id)
+    feedback_prompt = str(cfg.get("feedback_prompt_template", "") or "").strip()
+    if feedback_prompt:
+        return feedback_prompt
+
+    character_style = _character_feedback_style_from_mcq_prompt(character_id)
+    if character_style:
         return (
-            "You are Professor Minerva McGonagall. Crisp, strict, decisive; no emojis; no stage directions. "
-            "Sound like McGonagall: brisk, no-nonsense, disciplined; firm standards with a controlled, teacherly tone; second-person address ('you'). "
-            "Be less harsh than Snape: no contempt, no insults; correct firmly and encourage disciplined improvement. "
-            "Avoid academic report phrasing (e.g., 'demonstrates', 'fundamental', 'theoretical', 'principles', 'application'). "
-            "Prefer plain classroom language: 'That will not do', 'Pay attention', 'Be precise', 'Again', 'Good—carry on'. "
-            "Be pointed, practical, and supportive when appropriate. "
-            f"{guardrails} Be brief, firm, and exam-focused."
+            f"{character_style}. "
+            "You are giving live exam feedback after each answer. "
+            "Respond in concise English, in-character, practical, and pointed. "
+            "No markdown, no emojis, no stage directions."
         )
     return (
-        "You are a strict examiner. Be brief, precise, and exam-focused. "
-        f"{guardrails} No emojis."
+        "You are an examiner giving live feedback after each answer. "
+        "Respond in concise English and focus on the student's performance. "
+        "No markdown, no emojis."
     )
 
 
@@ -1232,56 +1323,12 @@ def _llm_short_exam_remark(character_id: str, *, kind: str, context: str = "") -
     if engine.mock_mode:
         return ""
     ctx = " ".join(str(context or "").strip().split())
-    if ctx:
-        ctx = f"Context: {ctx}\n"
-    style_seed = uuid.uuid4().hex
-    cid = (character_id or "").lower()
     if kind == "correct":
-        if "snape" in cid:
-            openers = ["Correct.", "Precisely.", "Good.", "Exactly.", "That's right.", "Adequate.", "Very well."]
-        elif "mcgonagall" in cid or "mcg" in cid:
-            openers = ["Good.", "Correct.", "Quite right.", "That's right.", "That's better.", "Well done.", "Exactly."]
-        else:
-            openers = ["That's right.", "That's correct.", "Correct.", "Exactly.", "Good.", "Well done."]
-        instruction = (
-            f"{ctx}"
-            f"Style seed (do not repeat it): {style_seed}\n"
-            "Write ONE short, in-character sentence reacting to a correct answer. "
-            "Start the sentence with ONE of these openers exactly: "
-            + ", ".join([f"'{o}'" for o in openers])
-            + ". "
-            "Choose the opener to maximize variety across runs. "
-            "Max 16 words. No markdown. No emojis. "
-            "Make it sound like the character, not a generic professor."
-        )
+        instruction = f"Write ONE short English sentence for a correct answer. Context: {ctx}. Max 16 words. No markdown. No emojis."
     elif kind == "incorrect":
-        if "snape" in cid:
-            openers = ["Wrong!", "Wrong.", "No.", "Obviously not.", "Incorrect.", "Not even close."]
-        elif "mcgonagall" in cid or "mcg" in cid:
-            openers = ["No.", "Not quite.", "That will not do.", "Incorrect.", "Careful.", "Stop guessing."]
-        else:
-            openers = ["Wrong!", "Wrong.", "Not quite.", "No.", "Incorrect.", "That's wrong."]
-        instruction = (
-            f"{ctx}"
-            f"Style seed (do not repeat it): {style_seed}\n"
-            "Write ONE short, in-character sentence reacting to an incorrect answer. "
-            "Start the sentence with ONE of these openers exactly: "
-            + ", ".join([f"'{o}'" for o in openers])
-            + ". "
-            "Choose the opener to maximize variety across runs. "
-            "Do NOT mention the correct option letter. "
-            "Do NOT include the phrase 'The correct answer is'. "
-            "Max 20 words. No markdown. No emojis. "
-            "Make it sound like the character, not a generic professor."
-        )
+        instruction = f"Write ONE short English sentence for an incorrect answer without giving the option letter. Context: {ctx}. Max 20 words. No markdown. No emojis."
     else:
-        instruction = (
-            f"{ctx}"
-            f"Style seed (do not repeat it): {style_seed}\n"
-            "Write 1–2 short, in-character sentences as a final examiner remark, with ONE concrete revision instruction. "
-            "Max 28 words total. No markdown. No emojis. "
-            "Do not sound like a generic academic report."
-        )
+        instruction = f"Write 1-2 short English final remarks with one concrete revision suggestion. Context: {ctx}. Max 28 words total. No markdown. No emojis."
     text = _llm_exam_feedback(
         [
             {"role": "system", "content": _examiner_style_prompt(character_id)},
@@ -1298,20 +1345,20 @@ def exam_feedback_correct(character_id: str, *, q_index: int) -> str:
         if "snape" in cid:
             return _pick_variant(
                 [
-                    "That's right. Try not to look so astonished; it is unbecoming.",
-                    "Correct. At least you read something other than the title.",
-                    "Precisely. Keep up—this is not a guessing game.",
-                    "Good. Acceptable. Proceed before you ruin it.",
+                    "Correct. Keep going.",
+                    "Right answer. Stay focused.",
+                    "Good. Next question.",
+                    "Exactly. Keep your pace.",
                 ],
                 q_index,
             )
         if "mcgonagall" in cid or "mcg" in cid:
             return _pick_variant(
                 [
-                    "That's correct. Good—do not get complacent.",
-                    "Good. Sensible. Keep your notes straight and move on.",
-                    "Well done. Stay focused; the next will not be kinder.",
-                    "Precisely. Continue, and keep the standard.",
+                    "That's correct. Keep it up.",
+                    "Good work. Move on.",
+                    "Well done. Stay consistent.",
+                    "Precisely. Continue.",
                 ],
                 q_index,
             )
@@ -1342,10 +1389,10 @@ def exam_feedback_incorrect(
         if "snape" in cid:
             opener = _pick_variant(
                 [
-                    "Wrong! Listen carefully—if you are capable of it.",
-                    "Incorrect. Pay attention; your confidence is not evidence.",
-                    "Wrong. This is what careless reading looks like in public.",
-                    "Incorrect. I expected better discipline. Clearly, that was optimistic.",
+                    "Wrong. Read more carefully.",
+                    "Incorrect. Check the prompt details.",
+                    "Not correct. Your reading is too loose.",
+                    "Incorrect. Be more rigorous.",
                 ],
                 q_index,
             )
@@ -1353,10 +1400,10 @@ def exam_feedback_incorrect(
         if "mcgonagall" in cid or "mcg" in cid:
             opener = _pick_variant(
                 [
-                    "Incorrect. Think it through properly—do not guess.",
-                    "Not quite. Slow down and read precisely; words matter.",
-                    "Incorrect. You are guessing—stop it at once.",
-                    "Not correct. Focus on the method, not the surface wording.",
+                    "Incorrect. Think first, then answer.",
+                    "Not quite. Slow down and read precisely.",
+                    "Wrong. Stop guessing.",
+                    "Incorrect. Focus on the method itself.",
                 ],
                 q_index,
             )
@@ -1381,22 +1428,22 @@ def exam_feedback_final(character_id: str, *, score: int, total: int) -> str:
         band = _score_band(score, total)
         if "snape" in cid:
             mapping = {
-                "excellent": "Excellent. For once, you have not wasted my time.",
-                "good": "Adequate. Do not mistake adequacy for insight.",
-                "fair": "Mediocre. You have work to do—start now, not later.",
-                "poor": "Disappointing. Guesswork is not scholarship; it is laziness.",
-                "none": "No score to judge—how convenient for you.",
+                "excellent": "Excellent performance this time.",
+                "good": "Good. Keep polishing details.",
+                "fair": "Fair. More practice is needed.",
+                "poor": "Poor. Review the lecture and retry.",
+                "none": "No score available yet.",
             }
-            return mapping.get(band, "Enough.")
+            return mapping.get(band, "Quiz finished.")
         if "mcgonagall" in cid or "mcg" in cid:
             mapping = {
-                "excellent": "Excellent work. That is the standard I expect—keep it there.",
-                "good": "Good. Solid understanding—polish the details and stop rushing.",
-                "fair": "Passable, but uneven. Review the method carefully and be exact.",
-                "poor": "Not acceptable. Go back and study properly, then try again.",
-                "none": "No score to judge—start when you are ready to work seriously.",
+                "excellent": "Excellent. Keep this standard.",
+                "good": "Good understanding. Improve the details.",
+                "fair": "Passable, but not stable yet.",
+                "poor": "Not acceptable. Review and try again.",
+                "none": "No score available yet.",
             }
-            return mapping.get(band, "Well.")
+            return mapping.get(band, "Quiz finished.")
         return f"Final score: {score} / {total}."
     try:
         remark = _llm_short_exam_remark(
@@ -1414,10 +1461,10 @@ def exam_feedback_final(character_id: str, *, score: int, total: int) -> str:
 def _roleplay_explain_feedback(character_id: str) -> str:
     cid = (character_id or "").lower()
     if "snape" in cid:
-        return "Lecture is ready. If you insist, press ‘Play Lecture Audio’; then go to the exam and try not to disgrace yourself."
+        return "Lecture is ready. Select a chunk to play, then go to the exam."
     if "mcgonagall" in cid or "mcg" in cid:
-        return "Lecture is ready. Review it properly, then go to the exam when you are prepared to be examined."
-    return "Lecture is ready. Review it, then go to the exam when you are ready."
+        return "Lecture is ready. Review it carefully, then enter the exam."
+    return "Lecture is ready. Review it, then enter the exam."
 
 
 def _roleplay_loading_text(character_id: str, *, phase: str) -> str:
@@ -1428,13 +1475,13 @@ def _roleplay_loading_text(character_id: str, *, phase: str) -> str:
         if "snape" in cid:
             return f"Professor {name} is scrutinizing your paper…"
         if "mcgonagall" in cid or "mcg" in cid:
-            return f"Professor {name} is reviewing your paper with unforgiving precision…"
+            return f"Professor {name} is reviewing your paper with strict precision…"
         return f"Professor {name} is reviewing your paper…"
     if "snape" in cid:
-        return f"Professor {name} is preparing something unpleasantly rigorous…"
+        return f"Professor {name} is preparing a rigorous exam…"
     if "mcgonagall" in cid or "mcg" in cid:
-        return f"Professor {name} is preparing a properly challenging set of questions…"
-    return f"Professor {name} is preparing your materials…"
+        return f"Professor {name} is preparing challenging questions…"
+    return f"Professor {name} is preparing your exam materials…"
 
 
 def build_loading_html(text: str) -> str:
@@ -1497,7 +1544,13 @@ def reset_ui_from_state(
     submit_interactive = quiz_ready and not state.get("completed", False)
     radio_interactive = submit_interactive
     lecture_tts_ready = bool(state.get("lecture_text"))
-    selected_paragraph_value = str(state.get("selected_paragraph_idx", "")).strip() or None
+    picker_choices = paragraph_picker_choices(state.get("lecture_text", ""))
+    selected_paragraph_value = paragraph_picker_value_for_idx(
+        state.get("lecture_text", ""),
+        str(state.get("selected_paragraph_idx", "")).strip(),
+    )
+    if selected_paragraph_value is None and picker_choices:
+        selected_paragraph_value = picker_choices[0][1]
     if state.get("completed"):
         radio_interactive = False
     return (
@@ -1512,7 +1565,7 @@ def reset_ui_from_state(
         state.get("status", "Idle"),
         build_clickable_lecture_html(state.get("lecture_text", ""), str(state.get("selected_paragraph_idx", ""))),
         gr.update(
-            choices=paragraph_picker_choices(state.get("lecture_text", "")),
+            choices=picker_choices,
             value=selected_paragraph_value,
             interactive=lecture_tts_ready,
             visible=lecture_tts_ready,
@@ -1520,6 +1573,7 @@ def reset_ui_from_state(
         state.get("lecture_audio_path", None),
         gr.update(interactive=lecture_tts_ready),
         gr.update(visible=lecture_tts_ready, interactive=lecture_tts_ready),
+        gr.update(visible=lecture_tts_ready, interactive=lecture_tts_ready),
         gr.update(visible=exam_picker_visible),
         gr.update(value=build_exam_chat_html(state), visible=show_exam_page and (quiz_ready or bool(state.get("exam_chat")))),
         gr.update(choices=current_choices(state), value=None, interactive=radio_interactive),
@@ -1538,10 +1592,10 @@ def process_pdf(pdf_file: Optional[str], character_id: str, state: Dict[str, Any
         yield reset_ui_from_state(state, feedback="Upload a PDF to start.", results_visible=False, loading_visible=False)
         return
 
-    state["status"] = "正在生成中..."
+    state["status"] = "Generating..."
     yield reset_ui_from_state(
         state,
-        feedback="正在读取论文并生成讲解与题目，请稍候...",
+        feedback="Reading the paper and generating lecture/quiz content...",
         results_visible=False,
         loading_visible=True,
         loading_text=_roleplay_loading_text(state.get("character_id") or DEFAULT_CHARACTER_ID, phase="lecture"),
@@ -1572,7 +1626,7 @@ def process_pdf(pdf_file: Optional[str], character_id: str, state: Dict[str, Any
             loading_visible=False,
         )
     except Exception as exc:
-        state["status"] = "Failed during generation."
+        state["status"] = "Generation failed."
         state["lecture_text"] = f"Error: {type(exc).__name__}: {exc}"
         state["current_page"] = "explain"
         yield reset_ui_from_state(
@@ -1586,7 +1640,7 @@ def process_pdf(pdf_file: Optional[str], character_id: str, state: Dict[str, Any
 def submit_answer(choice: Optional[str], state: Dict[str, Any]):
     if not state.get("mcqs"):
         state["status"] = "No quiz loaded."
-        return reset_ui_from_state(state, feedback="Load a PDF first.")
+        return reset_ui_from_state(state, feedback="Upload a PDF and generate lecture first.")
     if state.get("completed"):
         return reset_ui_from_state(state, feedback="Quiz already completed.")
     if not choice:
@@ -1664,7 +1718,7 @@ def submit_answer(choice: Optional[str], state: Dict[str, Any]):
 
 def restart_quiz(state: Dict[str, Any]):
     if not state.get("mcqs"):
-        return reset_ui_from_state(new_session_state(), feedback="Load a PDF first.")
+        return reset_ui_from_state(new_session_state(), feedback="Upload a PDF and generate lecture first.")
     state["current_index"] = 0
     state["score"] = 0
     state["awaiting_next_after_wrong"] = False
@@ -1713,7 +1767,7 @@ def generate_exam_mcq(selected_character_id: Optional[str], state: Dict[str, Any
         yield reset_ui_from_state(state, feedback="Generate lecture first.", results_visible=False, loading_visible=False)
         return
     if not selected_character_id:
-        state["status"] = "Please select a character to generate MCQs."
+        state["status"] = "Please choose an examiner."
         yield reset_ui_from_state(state, feedback="", results_visible=True, loading_visible=False)
         return
 
@@ -1749,7 +1803,7 @@ def generate_exam_mcq(selected_character_id: Optional[str], state: Dict[str, Any
         state["current_page"] = "exam"
         state["mcq_generating"] = False
         _ensure_current_question_in_exam_chat(state)
-        state["status"] = "Your exam is prepared."
+        state["status"] = "Exam prepared."
         yield reset_ui_from_state(
             state,
             feedback="",
@@ -1759,10 +1813,10 @@ def generate_exam_mcq(selected_character_id: Optional[str], state: Dict[str, Any
     except Exception as exc:
         state["current_page"] = "exam"
         state["mcq_generating"] = False
-        state["status"] = "Failed during MCQ generation."
+        state["status"] = "Exam generation failed."
         _append_exam_assistant_text(
             state,
-            f"Failed to generate the exam.\nError: {type(exc).__name__}: {exc}",
+            f"Failed to generate exam.\nError: {type(exc).__name__}: {exc}",
             kind="note",
         )
         yield reset_ui_from_state(
@@ -1839,18 +1893,18 @@ def play_lecture_audio(state: Dict[str, Any]):
     backend = _tts_backend_name()
     voice = tts_voice_for_character(state.get("character_id"))
     try:
-        state["status"] = f"Generating lecture audio ({backend})..."
+        state["status"] = f"Generating full lecture audio ({backend})..."
         state["lecture_audio_path"] = engine.synthesize_tts(state["lecture_text"], name_prefix="lecture", voice=voice)
-        state["status"] = "Lecture audio ready."
+        state["status"] = "Full lecture audio ready."
         return (
             state,
             state["status"],
             state["lecture_audio_path"],
-            f"Lecture audio generated via `{backend}`.",
+            f"Full lecture audio generated via `{backend}`.",
             build_clickable_lecture_html(state.get("lecture_text", ""), str(state.get("selected_paragraph_idx", ""))),
         )
     except Exception as exc:
-        state["status"] = "Lecture audio generation failed."
+        state["status"] = "Full lecture audio generation failed."
         return (
             state,
             state["status"],
@@ -1865,7 +1919,14 @@ def split_lecture_paragraphs(text: str) -> List[str]:
     if not s:
         return []
     pieces = re.split(r"\n\s*\n+", s)
-    return [p.strip() for p in pieces if p and p.strip()]
+    paragraphs = [p.strip() for p in pieces if p and p.strip()]
+    # If the model outputs a single giant paragraph, fall back to sentence-pair chunks
+    # so the selector always has usable granularity for TTS.
+    if len(paragraphs) <= 1:
+        fallback_chunks = split_text_every_two_sentences(s, max_len=420)
+        if len(fallback_chunks) > 1:
+            return [c.strip() for c in fallback_chunks if c and c.strip()]
+    return paragraphs
 
 
 def paragraph_picker_choices(lecture_text: str) -> List[tuple[str, str]]:
@@ -1879,6 +1940,29 @@ def paragraph_picker_choices(lecture_text: str) -> List[tuple[str, str]]:
     return choices
 
 
+def paragraph_picker_idx_from_value(value: Any) -> str:
+    s = str(value or "").strip()
+    if not s:
+        return ""
+    if s.isdigit():
+        return s
+    m = re.match(r"^\s*(\d+)\s*[\.、:：-]", s)
+    if not m:
+        return ""
+    return str(max(0, int(m.group(1)) - 1))
+
+
+def paragraph_picker_value_for_idx(lecture_text: str, idx: str) -> Optional[str]:
+    try:
+        i = int(str(idx or "").strip())
+    except Exception:
+        return None
+    paragraphs = split_lecture_paragraphs(lecture_text)
+    if i < 0 or i >= len(paragraphs):
+        return None
+    return str(i)
+
+
 def build_clickable_lecture_html(lecture_text: str, selected_idx: str = "") -> str:
     paragraphs = split_lecture_paragraphs(lecture_text)
     if not paragraphs:
@@ -1897,9 +1981,9 @@ def build_clickable_lecture_html(lecture_text: str, selected_idx: str = "") -> s
             else ""
         )
         parts.append(
-            f'<div class="lecture-paragraph{selected_cls}" data-idx="{i}" '
-            f'style="{selected_style}" '
-            f'onclick="window.__lectureSelectParagraph && window.__lectureSelectParagraph({i}, this, true);">{safe}</div>'
+            f'<div class="lecture-paragraph{selected_cls}" data-idx="{i}" style="{selected_style}">'
+            f'<div class="chunk-text">{safe}</div>'
+            f'</div>'
         )
     parts.append("</div>")
     return "".join(parts)
@@ -1909,12 +1993,12 @@ def play_lecture_paragraph_audio(paragraph_idx: str, state: Dict[str, Any]):
     lecture_text = state.get("lecture_text", "")
     paragraphs = split_lecture_paragraphs(str(lecture_text or ""))
     if not paragraphs:
-        state["status"] = "暂无讲解内容。"
+        state["status"] = "No lecture content available."
         return (
             state,
             state.get("status", "Idle"),
             state.get("lecture_audio_path"),
-            "请先生成讲解。",
+            "Generate lecture first.",
             build_clickable_lecture_html(state.get("lecture_text", ""), str(state.get("selected_paragraph_idx", ""))),
         )
 
@@ -1923,12 +2007,12 @@ def play_lecture_paragraph_audio(paragraph_idx: str, state: Dict[str, Any]):
     except Exception:
         idx = -1
     if idx < 0 or idx >= len(paragraphs):
-        state["status"] = "段落选择无效。"
+        state["status"] = "Invalid chunk selection."
         return (
             state,
             state.get("status", "Idle"),
             state.get("lecture_audio_path"),
-            "请重新点击要播放的段落。",
+            "Please select a valid chunk.",
             build_clickable_lecture_html(state.get("lecture_text", ""), str(state.get("selected_paragraph_idx", ""))),
         )
 
@@ -1936,24 +2020,24 @@ def play_lecture_paragraph_audio(paragraph_idx: str, state: Dict[str, Any]):
     voice = tts_voice_for_character(state.get("character_id"))
     try:
         state["selected_paragraph_idx"] = str(idx)
-        state["status"] = f"正在生成段落语音（{backend}）..."
+        state["status"] = f"Generating chunk audio ({backend})..."
         audio_path = engine.synthesize_tts(
             paragraphs[idx],
             name_prefix=f"lecture_p{idx+1}",
             voice=voice,
         )
         state["lecture_audio_path"] = audio_path
-        state["status"] = "段落语音已生成。"
+        state["status"] = "Chunk audio ready."
         char_len = len(paragraphs[idx])
         return (
             state,
             state["status"],
             audio_path,
-            f"已生成第 {idx+1}/{len(paragraphs)} 段语音（{char_len} 字符），可在下方播放。",
+            f"Generated chunk {idx+1}/{len(paragraphs)} ({char_len} chars). You can play it below.",
             build_clickable_lecture_html(state.get("lecture_text", ""), str(state.get("selected_paragraph_idx", ""))),
         )
     except Exception as exc:
-        state["status"] = "段落语音生成失败。"
+        state["status"] = "Chunk audio generation failed."
         return (
             state,
             state["status"],
@@ -1966,8 +2050,8 @@ def play_lecture_paragraph_audio(paragraph_idx: str, state: Dict[str, Any]):
 def play_explanation_audio(state: Dict[str, Any]):
     text = state.get("last_explanation_tts_text", "")
     if not text:
-        state["status"] = "No explanation available for TTS."
-        return state, state["status"], state.get("explanation_audio_path"), "Answer a question incorrectly first."
+        state["status"] = "No explanation available for audio."
+        return state, state["status"], state.get("explanation_audio_path"), "Answer a question first."
     voice = tts_voice_for_character(state.get("exam_character_id") or state.get("character_id"))
     try:
         state["status"] = "Generating explanation audio..."
@@ -1981,27 +2065,54 @@ def play_explanation_audio(state: Dict[str, Any]):
 
 def on_play_lecture_audio_click(state: Dict[str, Any]):
     state, status, audio_path, feedback, lecture_html = play_lecture_audio(state)
-    selected_paragraph_value = str(state.get("selected_paragraph_idx", "")).strip() or None
+    lecture_text = state.get("lecture_text", "")
+    picker_choices = paragraph_picker_choices(lecture_text)
+    selected_paragraph_value = paragraph_picker_value_for_idx(
+        lecture_text,
+        str(state.get("selected_paragraph_idx", "")).strip(),
+    )
+    if selected_paragraph_value is None and picker_choices:
+        selected_paragraph_value = picker_choices[0][1]
+    lecture_tts_ready = bool(lecture_text)
     return (
         state,
         status,
         audio_path,
         feedback,
         lecture_html,
-        gr.update(value=selected_paragraph_value),
+        gr.update(
+            choices=picker_choices,
+            value=selected_paragraph_value,
+            interactive=lecture_tts_ready,
+            visible=lecture_tts_ready,
+        ),
     )
 
 
 def on_play_paragraph_click(paragraph_idx: str, state: Dict[str, Any]):
-    state, status, audio_path, feedback, lecture_html = play_lecture_paragraph_audio(paragraph_idx, state)
-    selected_paragraph_value = str(state.get("selected_paragraph_idx", "")).strip() or None
+    idx_value = paragraph_picker_idx_from_value(paragraph_idx)
+    state, status, audio_path, feedback, lecture_html = play_lecture_paragraph_audio(idx_value, state)
+    lecture_text = state.get("lecture_text", "")
+    picker_choices = paragraph_picker_choices(lecture_text)
+    selected_paragraph_value = paragraph_picker_value_for_idx(
+        lecture_text,
+        str(state.get("selected_paragraph_idx", "")).strip(),
+    )
+    if selected_paragraph_value is None and picker_choices:
+        selected_paragraph_value = picker_choices[0][1]
+    lecture_tts_ready = bool(lecture_text)
     return (
         state,
         status,
         audio_path,
         feedback,
         lecture_html,
-        gr.update(value=selected_paragraph_value),
+        gr.update(
+            choices=picker_choices,
+            value=selected_paragraph_value,
+            interactive=lecture_tts_ready,
+            visible=lecture_tts_ready,
+        ),
     )
 
 
@@ -2316,9 +2427,11 @@ body {{
 #lecture-clickable .lecture-clickable,
 #lecture-clickable .lecture-clickable * {{
   pointer-events: auto !important;
+  opacity: 1 !important;
+  filter: none !important;
 }}
 #lecture-clickable .lecture-paragraph {{
-  cursor: pointer;
+  cursor: default;
   pointer-events: auto !important;
   padding: 10px 12px;
   border-radius: 14px;
@@ -2329,6 +2442,10 @@ body {{
   line-height: 1.45 !important;
   color: rgba(244,246,251,0.95) !important;
 }}
+#lecture-clickable .chunk-text {{
+  flex: 1 1 auto;
+  min-width: 0;
+}}
 #lecture-clickable .lecture-paragraph:hover {{
   background: rgba(255,255,255,0.08);
   border-color: rgba(255,255,255,0.14);
@@ -2345,6 +2462,15 @@ body {{
   box-shadow: 0 0 0 1px rgba(255,255,255,0.16) inset !important;
   color: #ffffff !important;
 }}
+#lecture-wrap [disabled],
+#lecture-wrap [aria-disabled="true"],
+#lecture-wrap .disabled,
+#lecture-wrap .pending,
+#lecture-wrap .loading,
+#lecture-wrap .generating {{
+  opacity: 1 !important;
+  filter: none !important;
+}}
 .lecture-empty {{
   padding: 10px 12px;
   color: rgba(244,246,251,0.72);
@@ -2378,6 +2504,63 @@ body {{
 #play-paragraph-btn {{
   display: none !important;
 }}
+#chunk-controls {{
+  margin-top: 8px !important;
+  align-items: start !important;
+  gap: 8px !important;
+  overflow: visible !important;
+  position: relative !important;
+  z-index: 60 !important;
+}}
+#tts-wrap,
+#paragraph-picker,
+#paragraph-picker .wrap {{
+  overflow: visible !important;
+}}
+#paragraph-picker .wrap {{
+  max-height: 320px !important;
+  overflow: auto !important;
+  border: 1px solid rgba(255,255,255,0.10) !important;
+  border-radius: 12px !important;
+  padding: 8px !important;
+}}
+#paragraph-picker label {{
+  border: 1px solid rgba(255,255,255,0.08) !important;
+  border-radius: 10px !important;
+  padding: 8px 10px !important;
+  margin-bottom: 6px !important;
+  background: rgba(255,255,255,0.03) !important;
+}}
+#paragraph-picker label:hover {{
+  background: rgba(255,255,255,0.06) !important;
+  border-color: rgba(255,255,255,0.14) !important;
+}}
+#paragraph-picker input[type="radio"]:checked + span {{
+  color: #f97316 !important;
+  font-weight: 700 !important;
+}}
+#play-selected-chunk-btn button {{
+  min-height: 42px !important;
+  height: 42px !important;
+  border-radius: 999px !important;
+  background: #f97316 !important;
+  border-color: #f97316 !important;
+  color: #ffffff !important;
+  font-weight: 700 !important;
+  font-size: 18px !important;
+  line-height: 1 !important;
+  padding: 0 14px !important;
+}}
+#play-selected-chunk-btn button:hover {{
+  background: #ea580c !important;
+  border-color: #ea580c !important;
+}}
+#play-selected-chunk-btn button[disabled] {{
+  background: #f97316 !important;
+  border-color: #f97316 !important;
+  color: #ffffff !important;
+  opacity: 0.75 !important;
+}}
 @keyframes tts_loading {{
   100% {{ background-size: 110%; }}
 }}
@@ -2563,6 +2746,9 @@ body {{
     border-color: rgba(15, 23, 42, 0.10);
     color: #0f172a !important;
   }}
+  #lecture-clickable .lecture-row {{
+    display: block;
+  }}
   #lecture-clickable .lecture-paragraph:hover {{
     background: rgba(15, 23, 42, 0.06);
     border-color: rgba(15, 23, 42, 0.16);
@@ -2952,14 +3138,16 @@ with gr.Blocks(css=CSS) as demo:
               const state = window.__lectureClickTtsGlobal || (window.__lectureClickTtsGlobal = {});
               if (state.bound) return;
               try {
-              const grRoot = (typeof window.gradioApp === "function") ? window.gradioApp() : null;
-              const rootCandidates = [
-                document,
-                grRoot && grRoot.shadowRoot ? grRoot.shadowRoot : null,
-                grRoot,
-              ].filter(Boolean);
+              const getRoots = () => {
+                const grRoot = (typeof window.gradioApp === "function") ? window.gradioApp() : null;
+                return [
+                  document,
+                  grRoot && grRoot.shadowRoot ? grRoot.shadowRoot : null,
+                  grRoot,
+                ].filter(Boolean);
+              };
               const q = (sel) => {
-                for (const r of rootCandidates) {
+                for (const r of getRoots()) {
                   const el = r.querySelector ? r.querySelector(sel) : null;
                   if (el) return el;
                 }
@@ -2987,8 +3175,8 @@ with gr.Blocks(css=CSS) as demo:
                 if (!audio) return;
                 if (audio.__ttsBound) return;
                 audio.__ttsBound = true;
-                audio.addEventListener("loadstart", () => showLoading("正在加载音频..."), true);
-                audio.addEventListener("waiting", () => showLoading("正在加载音频..."), true);
+                audio.addEventListener("loadstart", () => showLoading("Loading audio..."), true);
+                audio.addEventListener("waiting", () => showLoading("Loading audio..."), true);
                 audio.addEventListener("canplay", () => hideLoading(), true);
                 audio.addEventListener("playing", () => hideLoading(), true);
                 audio.addEventListener("error", () => hideLoading(), true);
@@ -3006,7 +3194,12 @@ with gr.Blocks(css=CSS) as demo:
                   boxShadow: "0 0 0 1px rgba(255,255,255,0.16) inset",
                   color: "#ffffff",
                 };
-                for (const r of rootCandidates) {
+                for (const r of getRoots()) {
+                  const rowNodes = r.querySelectorAll ? r.querySelectorAll("#lecture-clickable .lecture-row.is-selected, #lecture-clickable .lecture-row[data-selected='1']") : [];
+                  for (const row of rowNodes) {
+                    row.classList.remove("is-selected");
+                    row.removeAttribute("data-selected");
+                  }
                   const nodes = r.querySelectorAll ? r.querySelectorAll("#lecture-clickable .lecture-paragraph.is-selected") : [];
                   for (const node of nodes) {
                     node.classList.remove("is-selected");
@@ -3022,6 +3215,11 @@ with gr.Blocks(css=CSS) as demo:
                 if (para && para.classList) {
                   para.classList.add("is-selected");
                   para.setAttribute("data-selected", "1");
+                  const row = para.closest ? para.closest(".lecture-row") : null;
+                  if (row && row.classList) {
+                    row.classList.add("is-selected");
+                    row.setAttribute("data-selected", "1");
+                  }
                   if (para.style) {
                     para.style.setProperty("background", selectedInlineStyle.background, "important");
                     para.style.setProperty("border-color", selectedInlineStyle.borderColor, "important");
@@ -3036,7 +3234,7 @@ with gr.Blocks(css=CSS) as demo:
                   input = inputWrap && inputWrap.querySelector ? inputWrap.querySelector("textarea, input") : null;
                 }
                 if (!input) {
-                  showLoading("未找到段落选择控件，请刷新页面重试。");
+                  showLoading("Chunk selector not found. Please refresh the page.");
                   return;
                 }
                 input.value = indexText;
@@ -3050,10 +3248,10 @@ with gr.Blocks(css=CSS) as demo:
                   if (innerBtn) btn = innerBtn;
                 }
                 if (!btn) {
-                  showLoading("未找到段落播放控件，请刷新页面重试。");
+                  showLoading("Chunk play control not found. Please refresh the page.");
                   return;
                 }
-                showLoading("正在生成语音...");
+                showLoading("Generating audio...");
                 btn.click();
               };
               window.__lectureSelectParagraph = (idx, el, autoPlay = true) => {
@@ -3063,12 +3261,24 @@ with gr.Blocks(css=CSS) as demo:
               const paragraphFromEvent = (e) => {
                 const target = e ? e.target : null;
                 if (target && target.nodeType === 1 && target.closest) {
+                  const btn = target.closest(".chunk-play-btn");
+                  if (btn) {
+                    const row = btn.closest(".lecture-row");
+                    if (row && row.querySelector) {
+                      const p = row.querySelector(".lecture-paragraph");
+                      if (p) return p;
+                    }
+                  }
                   const p = target.closest(".lecture-paragraph");
                   if (p) return p;
                 }
                 const path = (e && typeof e.composedPath === "function") ? e.composedPath() : [];
                 for (const n of path) {
                   if (n && n.classList && n.classList.contains("lecture-paragraph")) return n;
+                  if (n && n.classList && n.classList.contains("lecture-row") && n.querySelector) {
+                    const p = n.querySelector(".lecture-paragraph");
+                    if (p) return p;
+                  }
                 }
                 return null;
               };
@@ -3076,26 +3286,51 @@ with gr.Blocks(css=CSS) as demo:
               const onParagraphClick = (e) => {
                 const para = paragraphFromEvent(e);
                 if (!para) return;
+                if (e && e.target && e.target.closest && e.target.closest(".chunk-play-btn")) {
+                  try { e.preventDefault(); } catch (_) {}
+                  try { e.stopPropagation(); } catch (_) {}
+                }
                 const idx = para.getAttribute("data-idx");
                 if (typeof idx !== "string" || idx.trim() === "") return;
                 selectParagraph(idx, para, true);
               };
+              const onChunkButtonClick = (e) => {
+                const btn = e && e.target && e.target.closest ? e.target.closest(".chunk-play-btn") : null;
+                if (!btn) return;
+                try { e.preventDefault(); } catch (_) {}
+                try { e.stopPropagation(); } catch (_) {}
+                const row = btn.closest ? btn.closest(".lecture-row") : null;
+                const para = row && row.querySelector ? row.querySelector(".lecture-paragraph") : null;
+                const idx = (btn.getAttribute && btn.getAttribute("data-idx")) || (para && para.getAttribute ? para.getAttribute("data-idx") : "");
+                if (!para || typeof idx !== "string" || idx.trim() === "") return;
+                selectParagraph(idx, para, true);
+              };
               const bindClickRoot = (root) => {
                 if (!root || !root.addEventListener) return;
                 if (root.__lectureClickBound) return;
                 root.__lectureClickBound = true;
                 root.addEventListener("click", onParagraphClick, true);
               };
+              const bindParagraphDomHandlers = () => {
+                for (const r of getRoots()) {
+                  if (!r || !r.querySelectorAll) continue;
+                  const btns = r.querySelectorAll("#lecture-clickable .chunk-play-btn");
+                  for (const btn of btns) {
+                    if (btn.__chunkPlayBound) continue;
+                    btn.__chunkPlayBound = true;
+                    btn.addEventListener("click", onChunkButtonClick, true);
+                  }
+                }
+              };
 
-              for (const r of rootCandidates) bindClickRoot(r);
+              for (const r of getRoots()) bindClickRoot(r);
               bindClickRoot(window);
+              bindParagraphDomHandlers();
 
               if (!state.rebindObserver) {
                 state.rebindObserver = new MutationObserver(() => {
-                  const nextRoot = (typeof window.gradioApp === "function") ? window.gradioApp() : null;
-                  for (const r of [document, nextRoot && nextRoot.shadowRoot ? nextRoot.shadowRoot : null, nextRoot]) {
-                    bindClickRoot(r);
-                  }
+                  for (const r of getRoots()) bindClickRoot(r);
+                  bindParagraphDomHandlers();
                 });
                 state.rebindObserver.observe(document.body, { childList: true, subtree: true });
               }
@@ -3119,28 +3354,30 @@ with gr.Blocks(css=CSS) as demo:
                             build_clickable_lecture_html(""),
                             elem_id="lecture-clickable",
                         )
-                    with gr.Row(elem_id="lecture-actions"):
-                        play_lecture_btn = gr.Button("Play Lecture Audio", interactive=False, scale=0)
-                    gr.Markdown("提示：可直接点击段落播放；若浏览器拦截点击，请使用下方 Chunk selector。", elem_id="paragraph-tts-tip")
+                    play_lecture_btn = gr.Button("Play Lecture Audio", interactive=False, visible=False)
+                    gr.Markdown("Tip: Select a chunk from the list below (left dot), then click the play button on the right.", elem_id="paragraph-tts-tip")
+                    lecture_feedback = gr.Markdown("")
+
+            with gr.Column(elem_id="tts-wrap"):
+                lecture_audio = gr.Audio(label="Lecture TTS", type="filepath", elem_id="lecture-audio")
+                gr.HTML(
+                    '<div id="tts-loading" aria-hidden="true" style="display:none"><div class="tts-loading-row"><div class="tts-loading-bar"></div><div class="tts-loading-text" id="tts-loading-text">Loading audio...</div></div></div>',
+                )
+                with gr.Row(elem_id="chunk-controls"):
                     paragraph_picker = gr.Radio(
                         choices=[],
                         value=None,
                         interactive=False,
                         visible=False,
-                        label="Chunks (fallback selector)",
+                        label="Select Chunk",
                         elem_id="paragraph-picker",
+                        scale=8,
                     )
-                    lecture_feedback = gr.Markdown("")
-                    with gr.Row(elem_id="exam-entry-wrap"):
-                        exam_btn = gr.Button("Go to Exam", interactive=False, variant="secondary", scale=0)
-
-            with gr.Column(elem_id="tts-wrap"):
-                lecture_audio = gr.Audio(label="Lecture TTS", type="filepath", elem_id="lecture-audio")
-                gr.HTML(
-                    '<div id="tts-loading" aria-hidden="true" style="display:none"><div class="tts-loading-row"><div class="tts-loading-bar"></div><div class="tts-loading-text" id="tts-loading-text">正在加载音频...</div></div></div>',
-                )
+                    play_selected_chunk_btn = gr.Button("▶", elem_id="play-selected-chunk-btn", visible=False, interactive=False, scale=1, min_width=52)
                 paragraph_idx = gr.Textbox(value="", label="", show_label=False, elem_id="selected-paragraph")
-                play_paragraph_btn = gr.Button("Play paragraph", elem_id="play-paragraph-btn")
+                play_paragraph_btn = gr.Button("Play Chunk", elem_id="play-paragraph-btn")
+                with gr.Row(elem_id="exam-entry-wrap"):
+                    exam_btn = gr.Button("Go to Exam", interactive=False, variant="secondary", scale=0)
 
         with gr.Column(visible=False, elem_id="exam-picker-overlay") as exam_picker_overlay:
             with gr.Column(elem_id="exam-picker-modal"):
@@ -3175,25 +3412,7 @@ with gr.Blocks(css=CSS) as demo:
                         requestAnimationFrame(doScroll);
                         setTimeout(doScroll, 50);
                       };
-                      const disableGlobalBlockers = () => {
-                        const nodes = document.querySelectorAll('.wrap.default, .wrap.center');
-                        nodes.forEach((n) => {
-                          const rect = n.getBoundingClientRect();
-                          const nearFullScreen =
-                            rect.width >= window.innerWidth - 4 &&
-                            rect.height >= window.innerHeight - 4 &&
-                            rect.left <= 2 &&
-                            rect.top <= 2;
-                          if (!nearFullScreen) return;
-                          const cs = window.getComputedStyle(n);
-                          if (cs.position !== 'fixed') return;
-                          n.style.setProperty('display', 'none', 'important');
-                          n.style.setProperty('pointer-events', 'none', 'important');
-                          n.style.setProperty('background', 'transparent', 'important');
-                        });
-                      };
                       const ensure = () => {
-                        disableGlobalBlockers();
                         const root = document.querySelector('#exam-chat');
                         const wrap = root ? root.querySelector('.exam-chat-wrap') : null;
                         if (!root || !wrap) return;
@@ -3248,6 +3467,7 @@ with gr.Blocks(css=CSS) as demo:
         paragraph_picker,
         lecture_audio,
         play_lecture_btn,
+        play_selected_chunk_btn,
         exam_btn,
         exam_picker_overlay,
         exam_chat,
@@ -3283,7 +3503,7 @@ with gr.Blocks(css=CSS) as demo:
         outputs=[state, status_box, lecture_audio, lecture_feedback, lecture_box, paragraph_picker],
         show_progress="minimal",
     )
-    paragraph_picker.change(
+    play_selected_chunk_btn.click(
         fn=on_play_paragraph_click,
         inputs=[paragraph_picker, state],
         outputs=[state, status_box, lecture_audio, lecture_feedback, lecture_box, paragraph_picker],

characters/Mcgonagall/mcq_retry_prompt.txt

@@ -1,26 +1,17 @@
 You are Professor Minerva McGonagall.
 
-Your previous output failed to parse. Correct it by producing STRICTLY valid JSON only.
+Output STRICT JSON only for 5 MCQs, with this exact shape:
+{"questions":[{"question":"...","options":["...","...","...","..."],"answer":"A","explanation":"..."}]}
 
-Hard rules:
-- Ignore References, Bibliography, Appendix, and supplementary sections.
-- Create exactly 5 MCQs (A–D).
-- Do not invent details.
-- Options must be plain strings WITHOUT leading labels like "A.", "B)", etc.
-- Output JSON only. No extra text, no markdown, no code fences.
-- Keep explanations in character (stern, academic, pointed), without stage directions.
+Hard constraints:
+- Exactly 5 questions.
+- Exactly 4 options per question.
+- answer MUST be A/B/C/D only.
+- explanation must be non-empty and concise.
+- No markdown, no preface, no trailing text.
+- No option prefixes like "A." or "B)".
 
-JSON schema (exact):
-{
-  "questions": [
-    {
-      "question": "…",
-      "options": ["…", "…", "…", "…"],
-      "answer": "A",
-      "explanation": "…"
-    }
-  ]
-}
+Keep tone strict, fair, and academically rigorous.
 
 Paper text:
 {document}

characters/snape/feedback.txt

@@ -0,0 +1,18 @@
+You are Professor Severus Snape giving live feedback in an oral exam.
+
+Voice and style:
+- Cold, precise, cutting, academically strict.
+- Sound unmistakably like Snape: terse, disdainful of sloppy thinking, controlled sarcasm.
+- Address the student directly as "you".
+- Keep each response short and punchy.
+
+Behavior rules:
+- Always stay in character.
+- Focus only on the student's answer quality and reasoning discipline.
+- Do not invent facts beyond the paper content.
+- No emojis, no markdown, no stage directions, no roleplay brackets.
+- Do not mention Harry Potter universe lore, spells, or named artifacts.
+
+Output constraints:
+- Return plain English text only.
+- Keep comments practical and exam-focused.

characters/snape/mcq_prompt.txt

@@ -45,4 +45,4 @@ Requirements:
  - Options must NOT include leading labels like "A.", "B)", etc. The UI will add A/B/C/D.
 
 Paper text:
-{document}
+{document}

characters/snape/mcq_retry_prompt.txt

@@ -1,26 +1,17 @@
 You are Professor Severus Snape.
 
-Your previous output failed to parse. Fix it. Produce STRICTLY valid JSON only.
+Output STRICT JSON only for 5 MCQs, with this exact shape:
+{"questions":[{"question":"...","options":["...","...","...","..."],"answer":"A","explanation":"..."}]}
 
-Hard rules:
-- Ignore References, Bibliography, Appendix, and supplementary sections.
-- Create exactly 5 MCQs (A–D).
-- Do not invent details.
-- Options must be plain strings WITHOUT leading labels like "A.", "B)", etc.
-- Output JSON only. No extra text, no markdown, no code fences.
-- Keep explanations in character (cold, sharp, slightly theatrical), without stage directions.
+Hard constraints:
+- Exactly 5 questions.
+- Exactly 4 options per question.
+- answer MUST be A/B/C/D only.
+- explanation must be non-empty and concise.
+- No markdown, no preface, no trailing text.
+- No option prefixes like "A." or "B)".
 
-JSON schema (exact):
-{
-  "questions": [
-    {
-      "question": "…",
-      "options": ["…", "…", "…", "…"],
-      "answer": "A",
-      "explanation": "…"
-    }
-  ]
-}
+Keep tone strict, sharp, and academically demanding.
 
 Paper text:
 {document}

characters/test/lecture_prompt.txt

@@ -1,10 +0,0 @@
-你是一名风格冷静、严谨但清晰的课程助教（Snape 风格），请阅读用户上传的论文内容，并输出一段中文讲解，要求：
-1. 先说明论文要解决的问题和背景；
-2. 再解释核心方法（按步骤/模块）；
-3. 再总结实验结果或亮点；
-4. 最后给出局限性与适用场景；
-5. 语言精炼、逻辑清楚，适合课堂讲解（约 400-700 字）；
-6. 不要虚构论文中不存在的实验细节。
-
-论文内容（可能是节选）：
-{document}

characters/test/mcq_prompt.txt

@@ -1,23 +0,0 @@
-请基于下面论文内容，生成 5 道中文单选题（MCQ），用于课堂测验。
-严格输出 JSON（不要 markdown 代码块），格式如下：
-{
-  "questions": [
-    {
-      "question": "...",
-      "options": ["A选项", "B选项", "C选项", "D选项"],
-      "answer": "A",
-      "explanation": "..."
-    }
-  ]
-}
-
-要求：
-1. 共 5 题；
-2. 每题 4 个选项；
-3. answer 必须是 A/B/C/D；
-4. 解析要说明为什么正确，以及常见误区；
-5. 题目覆盖背景、方法、结果、局限性；
-6. 题目难度适中，适合课程课堂测验。
-
-论文内容（可能是节选）：
-{document}

characters/test/mcq_retry_prompt.txt

@@ -1,14 +0,0 @@
-基于以下论文内容生成 5 道中文单选题。只输出合法 JSON，不要任何解释，不要 markdown。
-
-限制：
-1. 必须是紧凑 JSON（单行也可以）；
-2. 共 5 题；
-3. 每题字段：question、options(4项)、answer(A/B/C/D)、explanation；
-4. explanation 保持简短（1-2句）；
-5. 不要输出任何 JSON 以外内容。
-
-输出格式：
-{"questions":[{"question":"...","options":["...","...","...","..."],"answer":"A","explanation":"..."}]}
-
-论文内容：
-{document}

characters/test/meta.json

@@ -1,12 +0,0 @@
-{
-  "id": "test",
-  "display_name": "Test",
-  "tagline": "Research paper explainer · MCQ coach",
-  "byline": "By @local-demo",
-  "chat_label": "Test",
-  "chat_mode": "paper mode",
-  "avatar": "avatar.jpg",
-  "lecture_prompt_file": "lecture_prompt.txt",
-  "mcq_prompt_file": "mcq_prompt.txt",
-  "mcq_retry_prompt_file": "mcq_retry_prompt.txt"
-}