initial: novel-app snapshot

services/chapter_writer.py

@@ -0,0 +1,194 @@
+"""Chapter generation service: context assembly + streaming chapter writing.
+
+Pure functions — no Flask imports.  Context assembly follows the 拆书→写书
+pipeline: recent chapters full-text, earlier chapters as short summaries,
+relevant characters, open foreshadowing, novel meta.  Total prompt size is
+bounded by ``config.CTX_BUDGET_MAX``.
+"""
+
+import json
+import logging
+
+import config
+from repositories import chapter_repo, character_repo, foreshadowing_repo
+from services import llm
+
+logger = logging.getLogger(__name__)
+
+SYSTEM_PROMPT = (
+    "你是一位专业网文写手, 擅长按大纲续写章节, 文风贴合用户指定风格。"
+    "只输出章节正文, 不要输出标题、解释或任何元信息。"
+)
+
+_EARLIER_SUMMARY_CHARS = 200  # per earlier-chapter heuristic summary
+_PREFIX_TAIL_CHARS = 500      # how much of a resume-prefix to echo back
+
+
+def _truncate(text, limit):
+    if not isinstance(text, str):
+        return text
+    if len(text) <= limit:
+        return text
+    return text[:limit] + "..."
+
+
+def build_chapter_context(db, novel, chapter, ctx_prev=None):
+    """Assemble the context dict for writing ``chapter`` of ``novel``.
+
+    Budget enforcement: drop earlier summaries first, then truncate prev
+    chapter contents, to stay under ``config.CTX_BUDGET_MAX``.
+    """
+    novel_id = novel["id"]
+    prev_limit = ctx_prev or config.CTX_PREV_CHAPTERS
+
+    prev = chapter_repo.get_previous_chapters(
+        db, novel_id, chapter.get("volume") or 1,
+        chapter.get("chapter_number") or 1, limit=prev_limit,
+    )
+    prev_chapters = [
+        {"title": c.get("title"), "volume": c.get("volume"),
+         "chapter_number": c.get("chapter_number"), "content": c.get("content") or ""}
+        for c in reversed(prev)  # chronological order
+    ]
+
+    prev_ids = {c["id"] for c in prev}
+    done = chapter_repo.list_done_chapters(
+        db, novel_id, limit=config.CTX_SUMMARY_CHAPTERS
+    )
+    earlier_summaries = [
+        {"title": c.get("title"), "volume": c.get("volume"),
+         "chapter_number": c.get("chapter_number"),
+         "summary": _truncate(c.get("content") or "", _EARLIER_SUMMARY_CHARS)}
+        for c in done if c["id"] not in prev_ids
+    ]
+
+    outline = chapter.get("outline") or ""
+    characters = character_repo.find_characters_in_text(db, novel_id, outline)
+    if not characters:
+        characters = character_repo.list_characters(db, novel_id)
+
+    context = {
+        "novel_meta": {
+            "title": novel.get("title"),
+            "genre": novel.get("genre"),
+            "style": novel.get("style"),
+        },
+        "prev_chapters": prev_chapters,
+        "earlier_summaries": earlier_summaries,
+        "characters": characters,
+        "open_foreshadowing": foreshadowing_repo.list_open_foreshadowing(db, novel_id),
+    }
+
+    serialized = json.dumps(context, ensure_ascii=False, default=str)
+    # Budget pass 1: drop earlier summaries.
+    if len(serialized) > config.CTX_BUDGET_MAX and context["earlier_summaries"]:
+        logger.info("chapter context over budget, dropping earlier summaries")
+        context["earlier_summaries"] = []
+        serialized = json.dumps(context, ensure_ascii=False, default=str)
+    # Budget pass 2: truncate prev chapter bodies proportionally.
+    if len(serialized) > config.CTX_BUDGET_MAX and context["prev_chapters"]:
+        per = max(1000, config.CTX_BUDGET_MAX // (2 * len(context["prev_chapters"])))
+        for c in context["prev_chapters"]:
+            c["content"] = _truncate(c["content"], per)
+        serialized = json.dumps(context, ensure_ascii=False, default=str)
+        logger.info("chapter context over budget, truncated prev chapters to %d chars each", per)
+
+    context["total_chars"] = len(serialized)
+    if len(serialized) < config.CTX_BUDGET_MIN:
+        logger.info(
+            "chapter context small (%d chars < %d): early chapter or sparse data",
+            len(serialized), config.CTX_BUDGET_MIN,
+        )
+    return context
+
+
+def build_chapter_prompts(context, chapter, prefix=None):
+    """Assemble system/user messages for chapter generation."""
+    meta = context["novel_meta"]
+    parts = [
+        f"小说: 《{meta['title']}》 类型: {meta['genre']} 文风: {meta['style']}",
+        "",
+    ]
+
+    if context["earlier_summaries"]:
+        parts.append("【更早章节摘要】")
+        for s in context["earlier_summaries"]:
+            parts.append(
+                f"第{s['volume']}卷 第{s['chapter_number']}章 {s.get('title') or ''}: {s['summary']}"
+            )
+        parts.append("")
+
+    if context["prev_chapters"]:
+        parts.append("【最近章节正文】")
+        for c in context["prev_chapters"]:
+            parts.append(
+                f"--- 第{c['volume']}卷 第{c['chapter_number']}章 {c.get('title') or ''} ---"
+            )
+            parts.append(c["content"])
+        parts.append("")
+
+    if context["characters"]:
+        parts.append("【相关角色】")
+        for ch in context["characters"]:
+            parts.append(
+                f"{ch.get('name')} ({ch.get('role') or '角色'}, 状态: {ch.get('status')}): "
+                f"{_truncate(ch.get('description') or '', 300)}"
+            )
+        parts.append("")
+
+    if context["open_foreshadowing"]:
+        parts.append("【未回收伏笔】")
+        for f in context["open_foreshadowing"]:
+            parts.append(f"- {f.get('description')} (埋设: 第{f.get('planted_chapter') or '?'}章)")
+        parts.append("")
+
+    parts.append("【当前章节大纲】")
+    parts.append(
+        f"第{chapter.get('volume') or 1}卷 第{chapter.get('chapter_number') or 1}章 "
+        f"{chapter.get('title') or ''}"
+    )
+    parts.append(chapter.get("outline") or "(无大纲, 请自由发挥但承接前文)")
+    parts.append("")
+
+    if prefix:
+        parts.append("【续写指令】以下为本章已生成的开头, 请无缝续写, 不要重复已有内容:")
+        parts.append(prefix[-_PREFIX_TAIL_CHARS:])
+        parts.append("")
+
+    parts.append("要求: 写 3000-8000 字正文, 只输出正文, 不要标题不要解释。")
+
+    user_prompt = "\n".join(parts)
+    return [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user", "content": user_prompt},
+    ]
+
+
+def stream_chapter(db, novel, chapter, prefix=None):
+    """Generator yielding SSE frames: context → token… → done / error.
+
+    NOTE: ``db`` is only used here to build the context; callers must build
+    everything inside the request view and may pass a connection that dies
+    with the request — so we build context eagerly before streaming.
+    """
+    context = build_chapter_context(db, novel, chapter)
+    summary = (
+        f"已注入上下文: 前 {len(context['prev_chapters'])} 章正文, "
+        f"{len(context['earlier_summaries'])} 条更早摘要, "
+        f"{len(context['characters'])} 个角色, "
+        f"{len(context['open_foreshadowing'])} 条未回收伏笔, "
+        f"共约 {context['total_chars']} 字符"
+        + (f"; 续写前缀 {len(prefix)} 字符" if prefix else "")
+    )
+    yield llm._sse("context", {"summary": summary})
+
+    messages = build_chapter_prompts(context, chapter, prefix=prefix)
+    parts = []
+
+    def _on_token(t):
+        parts.append(t)
+
+    yield from llm.make_token_stream(messages, _on_token, max_tokens=8192)
+
+    full_text = (prefix or "") + "".join(parts)
+    yield llm._sse("done", {"content": full_text, "word_count": len(full_text)})