lizhilun/AI_NovelGenerator
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fixed #86 #82 ;optimized prompt words again

Browse Source
This commit is contained in:
YILING0013
2025-02-09 15:36:47 +08:00
parent 8d2d66bd79
commit 1ce0edceac
2 changed files with 121 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
novel_generator.py
+70 -16
View File
@@ -35,7 +35,8 @@ from prompt_definitions import (
update_character_state_prompt,
first_chapter_draft_prompt,
next_chapter_draft_prompt,
summarize_recent_chapters_prompt
summarize_recent_chapters_prompt,
create_character_state_prompt
)
# 章节目录解析
@@ -146,20 +147,20 @@ def init_vector_store(
try:
class LCEmbeddingWrapper(LCEmbeddings):
def embed_documents(self, doc_texts: List[str]) -> List[List[float]]:
def embed_documents(self, texts: List[str]) -> List[List[float]]:
return call_with_retry(
func=embedding_adapter.embed_documents,
max_retries=3,
fallback_return=[],
doc_texts=doc_texts
texts=texts
)
def embed_query(self, query_text: str) -> List[float]:
def embed_query(self, query: str) -> List[float]:
res = call_with_retry(
func=embedding_adapter.embed_query,
max_retries=3,
fallback_return=[],
query_text=query_text
query=query
)
return res
@@ -195,20 +196,20 @@ def load_vector_store(
try:
class LCEmbeddingWrapper(LCEmbeddings):
def embed_documents(self, doc_texts: List[str]) -> List[List[float]]:
def embed_documents(self, texts: List[str]) -> List[List[float]]:
return call_with_retry(
func=embedding_adapter.embed_documents,
max_retries=3,
fallback_return=[],
doc_texts=doc_texts
texts=texts
)
def embed_query(self, query_text: str) -> List[float]:
def embed_query(self, query: str) -> List[float]:
res = call_with_retry(
func=embedding_adapter.embed_query,
max_retries=3,
fallback_return=[],
query_text=query_text
query=query
)
return res
@@ -247,8 +248,9 @@ def split_text_for_vectorstore(chapter_text: str,
"""
if not chapter_text.strip():
return []
nltk.download('punkt', quiet=True)
nltk.download('punkt_tab', quiet=True)
sentences = nltk.sent_tokenize(chapter_text)
if not sentences:
return []
@@ -331,6 +333,7 @@ def get_relevant_context_from_vector_store(
"""
从向量库中检索与 query 最相关的 k 条文本，拼接后返回。
如果向量库加载/检索失败，则返回空字符串。
最终只返回最多2000字符的检索片段。
"""
store = load_vector_store(embedding_adapter, filepath)
if not store:
@@ -343,6 +346,9 @@ def get_relevant_context_from_vector_store(
logging.info(f"No relevant documents found for query '{query}'. Returning empty context.")
return ""
combined = "\n".join([d.page_content for d in docs])
# 限制长度最多2000字符
if len(combined) > 2000:
combined = combined[:2000]
return combined
except Exception as e:
logging.warning(f"Similarity search failed: {e}")
@@ -470,6 +476,10 @@ def Novel_architecture_generate(
若在中间任何一步报错且重试多次失败，则将已经生成的内容写入 partial_architecture.json 并退出；
下次调用时可从该步骤继续。
最终输出 Novel_architecture.txt
新增：
- 在完成角色动力学设定后，依据该角色体系，使用 create_character_state_prompt 生成初始角色状态表，
并存储到 character_state.txt，后续维护更新。
"""
os.makedirs(filepath, exist_ok=True)
@@ -521,6 +531,28 @@ def Novel_architecture_generate(
else:
logging.info("Step2 already done. Skipping...")
# 在完成角色动力学设定后，生成初始角色状态表
if "character_dynamics_result" in partial_data and "character_state_result" not in partial_data:
logging.info("Generating initial character state from character dynamics ...")
prompt_char_state_init = create_character_state_prompt.format(
character_dynamics=partial_data["character_dynamics_result"].strip()
)
character_state_init = invoke_with_cleaning(llm_adapter, prompt_char_state_init)
if not character_state_init.strip():
logging.warning("create_character_state_prompt generation failed.")
# 写入目前已有结果，然后退出
save_partial_architecture_data(filepath, partial_data)
return
partial_data["character_state_result"] = character_state_init
# 保存到文件
character_state_file = os.path.join(filepath, "character_state.txt")
clear_file_content(character_state_file)
save_string_to_txt(character_state_init, character_state_file)
save_partial_architecture_data(filepath, partial_data)
logging.info("Initial character state created and saved.")
# Step3: 世界观
if "world_building_result" not in partial_data:
logging.info("Step3: Generating world_building_prompt ...")
@@ -595,9 +627,9 @@ def compute_chunk_size(number_of_chapters: int, max_tokens: int) -> int:
并确保 chunk_size 不会小于1或大于实际章节数。
"""
tokens_per_chapter = 100.0
ratio = max_tokens / tokens_per_chapter # 例如：8192 / 100 = 81.92
ratio_rounded_to_10 = int(ratio // 10) * 10 # => 80
chunk_size = ratio_rounded_to_10 - 10 # => 70
ratio = max_tokens / tokens_per_chapter
ratio_rounded_to_10 = int(ratio // 10) * 10
chunk_size = ratio_rounded_to_10 - 10
if chunk_size < 1:
chunk_size = 1
if chunk_size > number_of_chapters:
@@ -605,6 +637,22 @@ def compute_chunk_size(number_of_chapters: int, max_tokens: int) -> int:
return chunk_size
def limit_chapter_blueprint(blueprint_text: str, limit_chapters: int = 100) -> str:
"""
从已有章节目录中只取最近的 limit_chapters 章，以避免 prompt 超长。
"""
pattern = r"(第\s*\d+\s*章.*?)(?=第\s*\d+\s*章|$)"
chapters = re.findall(pattern, blueprint_text, flags=re.DOTALL)
if not chapters:
return blueprint_text
if len(chapters) <= limit_chapters:
return blueprint_text
selected = chapters[-limit_chapters:]
return "\n\n".join(selected).strip()
# ============ 2) 生成章节蓝图（新增分块逻辑 + 断点续跑） ============
def Chapter_blueprint_generate(
@@ -621,6 +669,7 @@ def Chapter_blueprint_generate(
"""
若 Novel_directory.txt 已存在且内容非空，则表示可能是之前的部分生成结果；
解析其中已有的章节数，从下一个章节继续分块生成；
对于已有章节目录，传入时仅保留最近100章目录，避免prompt过长。
否则：
- 若章节数 <= chunk_size，直接一次性生成
- 若章节数 > chunk_size，进行分块生成
@@ -674,10 +723,11 @@ def Chapter_blueprint_generate(
current_start = max_existing_chap + 1
while current_start <= number_of_chapters:
current_end = min(current_start + chunk_size - 1, number_of_chapters)
limited_blueprint = limit_chapter_blueprint(final_blueprint, 100)
chunk_prompt = chunked_chapter_blueprint_prompt.format(
novel_architecture=architecture_text,
chapter_list=final_blueprint, # 已有的章节列表文本
chapter_list=limited_blueprint, # 只保留最近100章
number_of_chapters=number_of_chapters,
n=current_start,
m=current_end
@@ -694,7 +744,7 @@ def Chapter_blueprint_generate(
final_blueprint += "\n\n" + chunk_result.strip()
# 实时写入，以免中途崩溃造成丢失
# 实时写入
clear_file_content(filename_dir)
save_string_to_txt(final_blueprint.strip(), filename_dir)
@@ -726,10 +776,11 @@ def Chapter_blueprint_generate(
current_start = 1
while current_start <= number_of_chapters:
current_end = min(current_start + chunk_size - 1, number_of_chapters)
limited_blueprint = limit_chapter_blueprint(final_blueprint, 100)
chunk_prompt = chunked_chapter_blueprint_prompt.format(
novel_architecture=architecture_text,
chapter_list=final_blueprint, # 已有的章节列表文本
chapter_list=limited_blueprint, # 只保留最近100章
number_of_chapters=number_of_chapters,
n=current_start,
m=current_end
@@ -969,6 +1020,7 @@ def finalize_chapter(
timeout=timeout
)
# 更新全局摘要
prompt_summary = summary_prompt.format(
chapter_text=chapter_text,
global_summary=old_global_summary
@@ -977,6 +1029,7 @@ def finalize_chapter(
if not new_global_summary.strip():
new_global_summary = old_global_summary
# 更新角色状态
prompt_char_state = update_character_state_prompt.format(
chapter_text=chapter_text,
old_state=old_character_state
@@ -1040,6 +1093,7 @@ def advanced_split_content(content: str,
similarity_threshold: float = 0.7,
max_length: int = 500) -> List[str]:
nltk.download('punkt', quiet=True)
nltk.download('punkt_tab', quiet=True)
sentences = nltk.sent_tokenize(content)
if not sentences:
return []