Update app.py

app.py

@@ -1,35 +1,66 @@
-import os, json, html, re
+import os, json, html, re, random
 import gradio as gr
 
-# =============== DeepSeek 客户端（OpenAI 兼容） ===============
+# ====== 依赖：DeepSeek（OpenAI 兼容） ======
 _HAS_OPENAI = False
 client = None
 DEEPSEEK_KEY = os.getenv("DEEPSEEK_API_KEY")
 try:
     from openai import OpenAI
     if DEEPSEEK_KEY:
-        client = OpenAI(
-            api_key=DEEPSEEK_KEY,
-            base_url="https://api.deepseek.com"  # ⚠️ 不要带 /v1
-        )
+        client = OpenAI(api_key=DEEPSEEK_KEY, base_url="https://api.deepseek.com")
         _HAS_OPENAI = True
 except Exception:
     _HAS_OPENAI = False
-# ============================================================
 
-APP_TITLE = "孙子你好"
+# ====== 依赖：中文分词 + TFIDF 相似度 ======
+import jieba
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.metrics.pairwise import cosine_similarity
 
+APP_TITLE = "孙子你好"
+CORPUS_FILE = "sunzi_full.json"   # 全书语料（见下方文件格式示例）
+TOP_K = 5                         # 交给大模型挑选的候选条目数
+
+# ---------------- 读全书 + 建索引 ----------------
+def load_corpus():
+    if not os.path.exists(CORPUS_FILE):
+        raise RuntimeError("未找到 sunzi_full.json，请先上传全书语料。")
+    with open(CORPUS_FILE, "r", encoding="utf-8") as f:
+        data = json.load(f)
+    # 规范字段
+    rows = []
+    for i, it in enumerate(data):
+        rows.append({
+            "id": it.get("id", f"q_{i:04d}"),
+            "chapter": it.get("chapter", "").strip(),
+            "quote": it.get("quote", "").strip()
+        })
+    return rows
+
+CORPUS = load_corpus()
+
+def jieba_tokenizer(s):
+    return [w for w in jieba.lcut(s) if w.strip()]
+
+DOCS = [f'{r["chapter"]} {r["quote"]}' for r in CORPUS]
+VEC = TfidfVectorizer(tokenizer=jieba_tokenizer, ngram_range=(1,2), min_df=1)
+MATRIX = VEC.fit_transform(DOCS)
+
+def topk_candidates(user_text, k=TOP_K):
+    qv = VEC.transform([user_text])
+    sims = cosine_similarity(qv, MATRIX).ravel()
+    idxs = sims.argsort()[::-1][:k]
+    return [CORPUS[i] for i in idxs]
+
+# ------------- UI 渲染（沿用你的短模板） -------------
 CHAP_HINT = {
-    "计篇":"先算后战","作战篇":"速战速决","谋攻篇":"上兵伐谋","军形篇":"先固己势",
+    "始计篇":"先计而后战","作战篇":"速战速决","谋攻篇":"上兵伐谋","军形篇":"先固己势",
     "兵势篇":"势能制胜","虚实篇":"避实击虚","军争篇":"抢占先机","九变篇":"因变制胜",
     "行军篇":"察情定位","地形篇":"因地用兵","九地篇":"九地九策","火攻篇":"借势放大",
-    "用间篇":"情报为先","始计篇":"先计而后战"
+    "用间篇":"情报为先"
 }
 
-def render_tip(msg: str):
-    if not msg: return ""
-    return f"<div style='color:#9a8fb3;font-size:13px;margin-bottom:6px'>{html.escape(msg)}</div>"
-
 def render_plan(item, show_speech=False, tip_msg=""):
     chap = html.escape(item.get("chapter",""))
     hint = html.escape(CHAP_HINT.get(chap,""))
@@ -45,7 +76,7 @@ def render_plan(item, show_speech=False, tip_msg=""):
     </div>
     """ if (show_speech and speech) else ""
 
-    tip = render_tip(tip_msg)
+    tip = f"<div style='color:#9a8fb3;font-size:13px;margin-bottom:6px'>{html.escape(tip_msg)}</div>" if tip_msg else ""
 
     return f"""
     <div style="font-size:22px;line-height:1.7em;">
@@ -63,8 +94,8 @@ def render_speech_only(item):
     speech = html.escape(item.get("speech","")).replace("\n", "<br>")
     return f"""
     <div style="font-size:22px;line-height:1.8em;padding:20px;background:#fbf8ff;border-radius:12px;border:1px solid #e2d7f2;">
-        <div style="font-weight:600;margin-bottom:10px;">💬 话术示例：</div>
-        <div>{speech}</div>
+      <div style="font-weight:600;margin-bottom:10px;">💬 话术示例：</div>
+      <div>{speech}</div>
     </div>
     """
 
@@ -81,31 +112,39 @@ def pack_outputs(html_out, pool, idx, show_flag, show_switch, show_speech_btn, s
         (gr.update() if user_text_keep else gr.update(value=""))
     )
 
-# ================= DeepSeek 生成（短模板） =================
-GEN_SYS = (
-    "你是中文军师，用《孙子兵法》给出极简、可执行的建议。"
-    "严格只输出一个 JSON 对象，字段仅限："
-    "chapter（篇名，如“虚实篇/谋攻篇/始计篇”等）、"
-    "quote（贴切原文，保留原句标点）、"
-    "insight（≤18字，人话，直白精炼）、"
-    "action（仅1条，≤40字，具体到动作/步骤/策略）、"
-    "speech（≤80字、可直接对对方说的话；礼貌、委婉、幽默有边界）。"
-    "不要输出除 JSON 以外的任何内容。"
+# ------------- DeepSeek 受限生成：仅在 Top-K 中选 -------------
+SYS_LOCKED = (
+    "你是中文军师，用《孙子兵法》解决现实问题。"
+    "系统会给你一个候选列表（含编号/篇名/原文）。你必须："
+    "1) 只在候选中选出【一条】最贴���原文（不可更换或自造原句）；"
+    "2) 输出一个 JSON 对象，字段仅限："
+    "id（所选编号）、chapter（篇名）、quote（原文原句）、"
+    "insight（≤18字，人话、直白精炼）、"
+    "action（仅1条，≤40字、具体到动作/步骤）、"
+    "speech（≤80字、可对对方直说；礼貌、委婉、幽默有边界）。"
+    "严格禁止输出除 JSON 外的任何内容。"
 )
 
-GEN_FEWSHOT = [
+FEWSHOT = [
     {
         "role":"user",
-        "content":"用户问题：老板让我临时加班但我状态很差，如何拒绝更体面？只输出JSON。"
+        "content":(
+            "【候选】\n"
+            "1) 作战篇｜夫兵久而国利者，未之有也。\n"
+            "2) 虚实篇｜避实而击虚。\n"
+            "【用户问题】老板让我临时加班但我状态很差，如何拒绝更体面？\n"
+            "只输出JSON。"
+        )
     },
     {
         "role":"assistant",
         "content":json.dumps({
+            "id":"1",
             "chapter":"作战篇",
             "quote":"夫兵久而国利者，未之有也。",
-            "insight":"别打消耗战，用质量说话。",
+            "insight":"别耗死自己，用质量说服。",
             "action":"给“今晚初稿/明早终版”两选项，让对方决定。",
-            "speech":"老板，我怕今晚赶出来质量不稳。不如我明早7:30交完整版本，保证更稳，您看可以吗？"
+            "speech":"老板，我担心今晚赶出来质量不稳。不如我明早7:30交完整版本，保证更稳，您看行吗？"
         }, ensure_ascii=False)
     }
 ]
@@ -115,64 +154,68 @@ def _extract_json(text: str) -> dict:
     js = m.group(0) if m else text
     return json.loads(js)
 
-def _fallback_item(msg_tip="使用本地兜底（可能未连接API）"):
+def _fallback_from_top1(cands, tip="使用本地兜底（可能未连接API）"):
+    if not cands:
+        return {
+            "item": {"chapter":"始计篇","quote":"先计而后战。","insight":"先摸清再出手。","action":"列两种稳妥方案并承诺时间点。","speech":"我们按更稳的节奏推进，彼此都放心。"},
+            "tip": tip
+        }
+    c = cands[0]
     return {
-        "item": {
-            "chapter": "始计篇",
-            "quote": "先计而后战。",
-            "insight": "先摸清再出手。",
-            "action": "列两种可行方案并承诺时间点，让对方选。",
-            "speech": "我这边按更稳的节奏推进更安全。给您两个方案，咱们选一个最稳的。"
-        },
-        "tip": msg_tip
+        "item": {"chapter": c["chapter"], "quote": c["quote"], "insight":"先捋清状况再出手。","action":"定目标与底线，给一个具体步骤。","speech":"我这边按更稳的节奏推进，给你一个清晰方案。"},
+        "tip": tip
     }
 
-def ai_generate_item(user_text: str, alt: bool=False) -> dict:
-    # 没 Key 或 SDK 不可用 —— 兜底
+def deepseek_choose_and_write(user_text: str, cands: list, alt: bool=False):
+    """
+    把 TF-IDF Top-K 候选喂给 DeepSeek，只允许从中选一条并生成短模板。
+    alt=True 时提示换表达，但仍限定在同一候选集。
+    """
+    listing = "\n".join([f'{i+1}) {c["chapter"]}｜{c["quote"]}' for i,c in enumerate(cands)])
+    ask = f"【候选】\n{listing}\n【用户问题】{user_text}\n只输出JSON。"
+    if alt:
+        ask += "\n请换一种表述思路（仍只能在候选内选），避免与上次措辞重复。"
+
     if not _HAS_OPENAI:
-        fb = _fallback_item("未检测到 DEEPSEEK_API_KEY，已使用本地兜底。")
+        fb = _fallback_from_top1(cands, "未检测到 DEEPSEEK_API_KEY，已使用本地兜底。")
         return fb["item"], fb["tip"]
 
     try:
-        user = (
-            f"用户问题：{user_text}\n"
-            "请从《孙子兵法》全书挑选最贴切的一句原文，并按上述字段输出。只输出 JSON。"
-        )
-        if alt:
-            user += "\n请更换角度或章节，避免与上一次重复；确保action是全新思路。"
-
-        msgs = [{"role":"system","content":GEN_SYS}] + GEN_FEWSHOT + [{"role":"user","content":user}]
-
+        msgs = [{"role":"system","content":SYS_LOCKED}] + FEWSHOT + [{"role":"user","content":ask}]
         resp = client.chat.completions.create(
             model="deepseek-chat",
             messages=msgs,
-            temperature=0.3 if not alt else 0.6,
-            max_tokens=480,
+            temperature=0.25 if not alt else 0.55,
+            max_tokens=420,
             timeout=30
         )
-        content = resp.choices[0].message.content.strip()
-        data = _extract_json(content)
-        chapter = str(data.get("chapter","始计篇")).strip()
+        data = _extract_json(resp.choices[0].message.content.strip())
+        sel_id = str(data.get("id","1")).strip()
+        # 把编号对应回候选，确保不越界
+        try:
+            idx = max(1, min(len(cands), int(re.sub(r"[^0-9]", "", sel_id) or "1"))) - 1
+        except Exception:
+            idx = 0
+        c = cands[idx]
         item = {
-            "chapter": chapter,
-            "quote": str(data.get("quote","先计而后战。")).strip(),
-            "insight": str(data.get("insight","先捋清状况，再出手。")).strip(),
-            "action": str(data.get("action","给出两种稳妥方案并承诺时间点。")).strip(),
+            "chapter": c["chapter"],
+            "quote": c["quote"],
+            "insight": str(data.get("insight","先捋清状况再出手。")).strip(),
+            "action": str(data.get("action","给出一个具体且可执行的步骤。")).strip(),
             "speech": str(data.get("speech","我们按更稳的节奏推进，彼此都放心。")).strip()
         }
         return item, ""
     except Exception as e:
-        # 任何异常都不抛出到前端，安静兜底
-        fb = _fallback_item(f"已使用兜底（{type(e).__name__}）")
+        fb = _fallback_from_top1(cands, f"已使用兜底（{type(e).__name__}）")
         return fb["item"], fb["tip"]
 
-# ----------------- 回调（沿用按钮逻辑） -----------------
+# ----------------- 回调（维持你的按钮可见逻辑） -----------------
 def on_generate(user_text, state_pool, state_idx, state_show):
     user_text = (user_text or "").strip()
     if not user_text:
-        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题。</div>",
-                            [], 0, False, False, False, False, False, True)
-    item, tip = ai_generate_item(user_text, alt=False)
+        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题。</div>", [], 0, False, False, False, False, False, True)
+    cands = topk_candidates(user_text, TOP_K)
+    item, tip = deepseek_choose_and_write(user_text, cands, alt=False)
     pool = [item]
     html_out = render_plan(item, False, tip)
     return pack_outputs(html_out, pool, 0, False, True, True, False, True, True)
@@ -180,9 +223,9 @@ def on_generate(user_text, state_pool, state_idx, state_show):
 def on_switch(state_pool, state_idx, state_show, user_text):
     user_text = (user_text or "").strip()
     if not user_text:
-        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题并点击生成兵法。</div>",
-                            [], 0, False, False, False, False, False, True)
-    item, tip = ai_generate_item(user_text, alt=True)
+        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题并点击生成兵法。</div>", [], 0, False, False, False, False, False, True)
+    cands = topk_candidates(user_text, TOP_K)
+    item, tip = deepseek_choose_and_write(user_text, cands, alt=True)
     pool = [item]
     html_out = render_plan(item, False, tip)
     return pack_outputs(html_out, pool, 0, False, True, True, False, True, True)
@@ -190,8 +233,7 @@ def on_switch(state_pool, state_idx, state_show, user_text):
 def on_show_speech(state_pool, state_idx, state_show):
     pool = state_pool or []
     if not pool:
-        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题并点击生成兵法。</div>",
-                            pool, state_idx, state_show, False, False, False, False, True)
+        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题并点击生成兵法。</div>", pool, state_idx, state_show, False, False, False, False, True)
     idx = max(0, min(state_idx, len(pool)-1))
     html_out = render_speech_only(pool[idx])
     return pack_outputs(html_out, pool, idx, True, True, True, True, True, True)
@@ -199,8 +241,7 @@ def on_show_speech(state_pool, state_idx, state_show):
 def on_back(state_pool, state_idx, state_show):
     pool = state_pool or []
     if not pool:
-        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题并点击生成兵法。</div>",
-                            pool, state_idx, state_show, False, False, False, False, True)
+        return pack_outputs("<div style='font-size:20px;'>请先输入你的问题并点击生成兵法。</div>", pool, state_idx, state_show, False, False, False, False, True)
     idx = max(0, min(state_idx, len(pool)-1))
     html_out = render_plan(pool[idx], False, "")
     return pack_outputs(html_out, pool, idx, False, True, True, False, True, True)