initial: novel-app snapshot

services/deconstruct.py

@@ -0,0 +1,137 @@
+"""Book deconstruction (拆书) service: novel architect that produces
+concept, world, characters, outline, and foreshadowing sections.
+
+Pure functions — no Flask imports.  Each section is streamed via
+:func:`llm.make_token_stream` with ``section_start`` / ``section_end``
+SSE bookends.
+"""
+
+import logging
+
+from services import llm
+
+logger = logging.getLogger(__name__)
+
+SECTIONS = ["concept", "world", "characters", "outline", "foreshadowing"]
+
+SYSTEM_PROMPT = (
+    "你是一位资深小说架构师, 精通网文结构设计与节奏把控。"
+    "请根据用户的需求, 逐步构建完整的小说框架。"
+    "输出纯文本, 不要使用 Markdown 标题标记 (# / ## 等), 用换行分段即可。"
+)
+
+# ---- per-section Chinese user prompts -----------------------------------
+
+_SECTION_PROMPTS = {
+    "concept": (
+        "请为小说《{title}》设计核心概念, 包括:\n"
+        "1. 主题 (这部小说想表达什么)\n"
+        "2. 核心冲突 (主角面临的主要矛盾)\n"
+        "3. 一句话卖点 (最能吸引读者的概括)\n"
+        "类型: {genre} | 风格: {style} | 目标字数: {target_words}"
+    ),
+    "world": (
+        "请为小说《{title}》设计世界观设定, 包括:\n"
+        "1. 时代背景\n"
+        "2. 主要地点\n"
+        "3. 力量体系 (如有)\n"
+        "4. 社会结构\n"
+        "类型: {genre} | 风格: {style}"
+    ),
+    "characters": (
+        "请为小说《{title}》设计主要角色表。\n"
+        "每个角色单独一段, 以【角色】名字 开头, 包含:\n"
+        "名字 / 定位 / 性格 / 动机 / 成长弧线。\n"
+        "类型: {genre} | 风格: {style}"
+    ),
+    "outline": (
+        "请为小说《{title}》规划三幕结构与卷章大纲。\n"
+        "格式: 第N卷 / 第N章 标题 — 一句话概要\n"
+        "预估总章数按目标字数 {target_words} 除以每章 5000 字计算。\n"
+        "类型: {genre} | 风格: {style}"
+    ),
+    "foreshadowing": (
+        "请为小说《{title}》设计伏笔计划。\n"
+        "每条伏笔单独一段, 以【伏笔】开头, 包含:\n"
+        "描述 / 埋设位置 (第几卷第几章) / 回收位置。\n"
+        "类型: {genre} | 风格: {style}"
+    ),
+}
+
+
+def _format_inputs(inputs):
+    """Return (title, genre, style, target_words, reference_novel, confirmed)."""
+    return (
+        inputs.get("title", "未命名"),
+        inputs.get("genre", "未指定"),
+        inputs.get("style", "未指定"),
+        inputs.get("target_words", "未指定"),
+        inputs.get("reference_novel"),
+        inputs.get("confirmed", {}),
+    )
+
+
+def _confirmed_context(confirmed):
+    """Render previously confirmed sections as context for the next prompt."""
+    if not confirmed:
+        return ""
+    parts = []
+    for sec in SECTIONS:
+        text = confirmed.get(sec)
+        if text:
+            parts.append(f"【已确认 - {sec}】\n{text}")
+    return "\n\n".join(parts)
+
+
+def build_section_prompt(section, inputs):
+    """Build the messages list for a given section.
+
+    ``inputs`` is a dict with keys: title, genre, style, target_words,
+    reference_novel (optional), confirmed (dict of previous section texts).
+    Returns a list of ``{"role": ..., "content": ...}`` dicts ready for
+    the LLM chat API.
+    """
+    title, genre, style, target_words, reference_novel, confirmed = _format_inputs(inputs)
+
+    template = _SECTION_PROMPTS.get(section)
+    if template is None:
+        raise ValueError(f"unknown section: {section}")
+
+    user_prompt = template.format(
+        title=title,
+        genre=genre,
+        style=style,
+        target_words=target_words,
+    )
+
+    if reference_novel:
+        user_prompt += f"\n\n参考小说: 《{reference_novel}》, 参考其节奏与结构但不抄袭。"
+
+    ctx = _confirmed_context(confirmed)
+    if ctx:
+        user_prompt = f"以下是已确认的前序设定:\n\n{ctx}\n\n---\n\n{user_prompt}"
+
+    return [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user", "content": user_prompt},
+    ]
+
+
+def stream_section(section, inputs):
+    """Generator yielding SSE frames for a section.
+
+    Emits ``section_start`` → token frames → ``section_end`` (with full
+    accumulated content).  Uses :func:`llm.make_token_stream` under the hood.
+    """
+    messages = build_section_prompt(section, inputs)
+    yield llm._sse("section_start", {"section": section})
+
+    parts = []
+
+    def _on_token(t):
+        parts.append(t)
+
+    yield from llm.make_token_stream(messages, _on_token, max_tokens=4096)
+
+    full_text = "".join(parts)
+    yield llm._sse("section_end", {"section": section, "content": full_text})