1174 lines
42 KiB
Python
1174 lines
42 KiB
Python
# novel_generator.py
|
||
# -*- coding: utf-8 -*-
|
||
import os
|
||
import logging
|
||
import re
|
||
import time
|
||
import traceback
|
||
import json
|
||
from typing import List, Optional, Tuple
|
||
|
||
from langchain_chroma import Chroma
|
||
from chromadb.config import Settings
|
||
from langchain.docstore.document import Document
|
||
|
||
# nltk、sentence_transformers 及文本处理相关
|
||
import nltk
|
||
from sentence_transformers import SentenceTransformer
|
||
from sklearn.metrics.pairwise import cosine_similarity
|
||
|
||
# 工具函数
|
||
from utils import (
|
||
read_file, append_text_to_file, clear_file_content,
|
||
save_string_to_txt
|
||
)
|
||
|
||
# prompt模板
|
||
from prompt_definitions import (
|
||
core_seed_prompt,
|
||
character_dynamics_prompt,
|
||
world_building_prompt,
|
||
plot_architecture_prompt,
|
||
chapter_blueprint_prompt,
|
||
chunked_chapter_blueprint_prompt,
|
||
summary_prompt,
|
||
update_character_state_prompt,
|
||
first_chapter_draft_prompt,
|
||
next_chapter_draft_prompt,
|
||
summarize_recent_chapters_prompt,
|
||
create_character_state_prompt
|
||
)
|
||
|
||
# 章节目录解析
|
||
from chapter_directory_parser import get_chapter_info_from_blueprint
|
||
|
||
from llm_adapters import create_llm_adapter
|
||
from embedding_adapters import create_embedding_adapter
|
||
|
||
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
|
||
|
||
# ============ 通用的重试封装 ============
|
||
|
||
def call_with_retry(func, max_retries=3, sleep_time=2, fallback_return=None, **kwargs):
|
||
"""
|
||
通用的重试机制封装。
|
||
:param func: 要执行的函数
|
||
:param max_retries: 最大重试次数
|
||
:param sleep_time: 重试前的等待秒数
|
||
:param fallback_return: 如果多次重试仍失败时的返回值
|
||
:param kwargs: 传给func的命名参数
|
||
:return: func的结果，若失败则返回 fallback_return
|
||
"""
|
||
for attempt in range(1, max_retries + 1):
|
||
try:
|
||
return func(**kwargs)
|
||
except Exception as e:
|
||
logging.warning(f"[call_with_retry] Attempt {attempt} failed with error: {e}")
|
||
traceback.print_exc()
|
||
if attempt < max_retries:
|
||
time.sleep(sleep_time)
|
||
else:
|
||
logging.error("Max retries reached, returning fallback_return.")
|
||
return fallback_return
|
||
|
||
|
||
# ============ 工具函数 ============
|
||
|
||
def remove_think_tags(text: str) -> str:
|
||
"""移除 <think>...</think> 包裹的内容"""
|
||
return re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL)
|
||
|
||
def debug_log(prompt: str, response_content: str):
|
||
logging.info(
|
||
f"\n[######################################### Prompt #########################################]\n{prompt}\n"
|
||
)
|
||
logging.info(
|
||
f"\n[######################################### Response #########################################]\n{response_content}\n"
|
||
)
|
||
|
||
def invoke_with_cleaning(llm_adapter, prompt: str) -> str:
|
||
"""
|
||
对 LLM 的调用增加了重试封装，
|
||
如果多次失败，则返回空字符串以继续流程，而不是中断。
|
||
"""
|
||
def _invoke(prompt):
|
||
return llm_adapter.invoke(prompt)
|
||
|
||
response = call_with_retry(func=_invoke, max_retries=3, fallback_return="", prompt=prompt)
|
||
if not response:
|
||
logging.warning("No response from model after retry. Return empty.")
|
||
return ""
|
||
cleaned_text = remove_think_tags(response)
|
||
debug_log(prompt, cleaned_text)
|
||
return cleaned_text.strip()
|
||
|
||
|
||
# ============ 获取 vectorstore 路径 ============
|
||
|
||
def get_vectorstore_dir(filepath: str) -> str:
|
||
return os.path.join(filepath, "vectorstore")
|
||
|
||
|
||
# ============ 清空向量库 ============
|
||
|
||
def clear_vector_store(filepath: str) -> bool:
|
||
import shutil
|
||
store_dir = get_vectorstore_dir(filepath)
|
||
if not os.path.exists(store_dir):
|
||
logging.info("No vector store found to clear.")
|
||
return False
|
||
try:
|
||
shutil.rmtree(store_dir)
|
||
logging.info(f"Vector store directory '{store_dir}' removed.")
|
||
return True
|
||
except Exception as e:
|
||
logging.error(f"无法删除向量库文件夹，请关闭程序后手动删除 {store_dir}。\n {str(e)}")
|
||
traceback.print_exc()
|
||
return False
|
||
|
||
|
||
# ============ 根据 embedding 接口创建/加载 Chroma ============
|
||
|
||
def init_vector_store(
|
||
embedding_adapter,
|
||
texts: List[str],
|
||
filepath: str
|
||
) -> Optional[Chroma]:
|
||
"""
|
||
在 filepath 下创建/加载一个 Chroma 向量库并插入 texts。
|
||
如果Embedding失败，则返回 None，不中断任务。
|
||
"""
|
||
from langchain.embeddings.base import Embeddings as LCEmbeddings
|
||
|
||
store_dir = get_vectorstore_dir(filepath)
|
||
os.makedirs(store_dir, exist_ok=True)
|
||
|
||
documents = [Document(page_content=str(t)) for t in texts]
|
||
|
||
try:
|
||
class LCEmbeddingWrapper(LCEmbeddings):
|
||
def embed_documents(self, texts: List[str]) -> List[List[float]]:
|
||
return call_with_retry(
|
||
func=embedding_adapter.embed_documents,
|
||
max_retries=3,
|
||
fallback_return=[],
|
||
texts=texts
|
||
)
|
||
|
||
def embed_query(self, query: str) -> List[float]:
|
||
res = call_with_retry(
|
||
func=embedding_adapter.embed_query,
|
||
max_retries=3,
|
||
fallback_return=[],
|
||
query=query
|
||
)
|
||
return res
|
||
|
||
chroma_embedding = LCEmbeddingWrapper()
|
||
|
||
vectorstore = Chroma.from_documents(
|
||
documents,
|
||
embedding=chroma_embedding,
|
||
persist_directory=store_dir,
|
||
client_settings=Settings(anonymized_telemetry=False),
|
||
collection_name="novel_collection"
|
||
)
|
||
return vectorstore
|
||
except Exception as e:
|
||
logging.warning(f"Init vector store failed: {e}")
|
||
traceback.print_exc()
|
||
return None
|
||
|
||
def load_vector_store(
|
||
embedding_adapter,
|
||
filepath: str
|
||
) -> Optional[Chroma]:
|
||
"""
|
||
读取已存在的 Chroma 向量库。若不存在则返回 None。
|
||
如果加载失败（embedding 或IO问题），则返回 None。
|
||
"""
|
||
store_dir = get_vectorstore_dir(filepath)
|
||
if not os.path.exists(store_dir):
|
||
logging.info("Vector store not found. Will return None.")
|
||
return None
|
||
|
||
from langchain.embeddings.base import Embeddings as LCEmbeddings
|
||
|
||
try:
|
||
class LCEmbeddingWrapper(LCEmbeddings):
|
||
def embed_documents(self, texts: List[str]) -> List[List[float]]:
|
||
return call_with_retry(
|
||
func=embedding_adapter.embed_documents,
|
||
max_retries=3,
|
||
fallback_return=[],
|
||
texts=texts
|
||
)
|
||
|
||
def embed_query(self, query: str) -> List[float]:
|
||
res = call_with_retry(
|
||
func=embedding_adapter.embed_query,
|
||
max_retries=3,
|
||
fallback_return=[],
|
||
query=query
|
||
)
|
||
return res
|
||
|
||
chroma_embedding = LCEmbeddingWrapper()
|
||
|
||
return Chroma(
|
||
persist_directory=store_dir,
|
||
embedding_function=chroma_embedding,
|
||
client_settings=Settings(anonymized_telemetry=False),
|
||
collection_name="novel_collection"
|
||
)
|
||
except Exception as e:
|
||
logging.warning(f"Failed to load vector store: {e}")
|
||
traceback.print_exc()
|
||
return None
|
||
|
||
|
||
# ============ 文本分段工具 ============
|
||
|
||
def split_by_length(text: str, max_length: int = 500) -> List[str]:
|
||
segments = []
|
||
start_idx = 0
|
||
while start_idx < len(text):
|
||
end_idx = min(start_idx + max_length, len(text))
|
||
segment = text[start_idx:end_idx]
|
||
segments.append(segment.strip())
|
||
start_idx = end_idx
|
||
return segments
|
||
|
||
def split_text_for_vectorstore(chapter_text: str,
|
||
max_length: int = 500,
|
||
similarity_threshold: float = 0.7) -> List[str]:
|
||
"""
|
||
对新的章节文本进行分段后，再用于存入向量库。
|
||
先句子切分 -> 语义相似度合并 -> 再按 max_length 切分。
|
||
"""
|
||
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 []
|
||
|
||
model = SentenceTransformer('paraphrase-MiniLM-L6-v2')
|
||
embeddings = model.encode(sentences)
|
||
|
||
merged_paragraphs = []
|
||
current_sentences = [sentences[0]]
|
||
current_embedding = embeddings[0]
|
||
|
||
for i in range(1, len(sentences)):
|
||
sim = cosine_similarity([current_embedding], [embeddings[i]])[0][0]
|
||
if sim >= similarity_threshold:
|
||
current_sentences.append(sentences[i])
|
||
current_embedding = (current_embedding + embeddings[i]) / 2.0
|
||
else:
|
||
merged_paragraphs.append(" ".join(current_sentences))
|
||
current_sentences = [sentences[i]]
|
||
current_embedding = embeddings[i]
|
||
|
||
if current_sentences:
|
||
merged_paragraphs.append(" ".join(current_sentences))
|
||
|
||
final_segments = []
|
||
for para in merged_paragraphs:
|
||
if len(para) > max_length:
|
||
sub_segments = split_by_length(para, max_length=max_length)
|
||
final_segments.extend(sub_segments)
|
||
else:
|
||
final_segments.append(para)
|
||
|
||
return final_segments
|
||
|
||
|
||
# ============ 更新向量库 ============
|
||
|
||
def update_vector_store(
|
||
embedding_adapter,
|
||
new_chapter: str,
|
||
filepath: str
|
||
):
|
||
"""
|
||
将最新章节文本插入到向量库中。
|
||
若库不存在则初始化；若初始化/更新失败，则跳过。
|
||
"""
|
||
splitted_texts = split_text_for_vectorstore(new_chapter)
|
||
if not splitted_texts:
|
||
logging.warning("No valid text to insert into vector store. Skipping.")
|
||
return
|
||
|
||
store = load_vector_store(embedding_adapter, filepath)
|
||
if not store:
|
||
logging.info("Vector store does not exist or failed to load. Initializing a new one for new chapter...")
|
||
store = init_vector_store(embedding_adapter, splitted_texts, filepath)
|
||
if not store:
|
||
logging.warning("Init vector store failed, skip embedding.")
|
||
else:
|
||
logging.info("New vector store created successfully.")
|
||
return
|
||
|
||
# 如果已有store，则直接往里插入
|
||
try:
|
||
docs = [Document(page_content=str(t)) for t in splitted_texts]
|
||
store.add_documents(docs)
|
||
logging.info("Vector store updated with the new chapter splitted segments.")
|
||
except Exception as e:
|
||
logging.warning(f"Failed to update vector store: {e}")
|
||
traceback.print_exc()
|
||
|
||
|
||
# ============ 向量检索上下文 ============
|
||
|
||
def get_relevant_context_from_vector_store(
|
||
embedding_adapter,
|
||
query: str,
|
||
filepath: str,
|
||
k: int = 2
|
||
) -> str:
|
||
"""
|
||
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:
|
||
logging.info("No vector store found or load failed. Returning empty context.")
|
||
return ""
|
||
|
||
try:
|
||
docs = store.similarity_search(query, k=k)
|
||
if not docs:
|
||
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}")
|
||
traceback.print_exc()
|
||
return ""
|
||
|
||
|
||
# ============ 从目录中获取最近 n 章文本 ============
|
||
|
||
def get_last_n_chapters_text(chapters_dir: str, current_chapter_num: int, n: int = 3) -> List[str]:
|
||
texts = []
|
||
start_chap = max(1, current_chapter_num - n)
|
||
for c in range(start_chap, current_chapter_num):
|
||
chap_file = os.path.join(chapters_dir, f"chapter_{c}.txt")
|
||
if os.path.exists(chap_file):
|
||
text = read_file(chap_file).strip()
|
||
texts.append(text)
|
||
else:
|
||
texts.append("")
|
||
return texts
|
||
|
||
|
||
# ============ 提炼(短期摘要, 下一章关键字) ============
|
||
|
||
def summarize_recent_chapters(
|
||
interface_format: str,
|
||
api_key: str,
|
||
base_url: str,
|
||
model_name: str,
|
||
temperature: float,
|
||
max_tokens: int,
|
||
chapters_text_list: List[str],
|
||
timeout: int = 600
|
||
) -> Tuple[str, str]:
|
||
"""
|
||
生成 (short_summary, next_chapter_keywords)
|
||
如果解析失败，则返回 (合并文本, "")
|
||
"""
|
||
combined_text = "\n".join(chapters_text_list).strip()
|
||
if not combined_text:
|
||
return ("", "")
|
||
|
||
llm_adapter = create_llm_adapter(
|
||
interface_format=interface_format,
|
||
base_url=base_url,
|
||
model_name=model_name,
|
||
api_key=api_key,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
timeout=timeout
|
||
)
|
||
|
||
prompt = summarize_recent_chapters_prompt.format(combined_text=combined_text)
|
||
response_text = invoke_with_cleaning(llm_adapter, prompt)
|
||
|
||
short_summary = ""
|
||
next_chapter_keywords = ""
|
||
|
||
for line in response_text.splitlines():
|
||
line = line.strip()
|
||
if line.startswith("短期摘要:"):
|
||
short_summary = line.replace("短期摘要:", "").strip()
|
||
elif line.startswith("下一章关键字:"):
|
||
next_chapter_keywords = line.replace("下一章关键字:", "").strip()
|
||
|
||
if not short_summary and not next_chapter_keywords:
|
||
short_summary = response_text
|
||
|
||
return (short_summary, next_chapter_keywords)
|
||
|
||
|
||
# ============ 持久化：情节架构（partial_architecture.json） ============
|
||
|
||
def load_partial_architecture_data(filepath: str) -> dict:
|
||
"""
|
||
从 filepath 下的 partial_architecture.json 读取已有的阶段性数据。
|
||
如果文件不存在或无法解析，返回空 dict。
|
||
"""
|
||
partial_file = os.path.join(filepath, "partial_architecture.json")
|
||
if not os.path.exists(partial_file):
|
||
return {}
|
||
|
||
try:
|
||
with open(partial_file, "r", encoding="utf-8") as f:
|
||
data = json.load(f)
|
||
return data
|
||
except Exception as e:
|
||
logging.warning(f"Failed to load partial_architecture.json: {e}")
|
||
return {}
|
||
|
||
def save_partial_architecture_data(filepath: str, data: dict):
|
||
"""
|
||
将阶段性数据写入 partial_architecture.json。
|
||
"""
|
||
partial_file = os.path.join(filepath, "partial_architecture.json")
|
||
try:
|
||
with open(partial_file, "w", encoding="utf-8") as f:
|
||
json.dump(data, f, ensure_ascii=False, indent=2)
|
||
except Exception as e:
|
||
logging.warning(f"Failed to save partial_architecture.json: {e}")
|
||
|
||
|
||
# ============ 1) 生成总体架构 ============
|
||
|
||
def Novel_architecture_generate(
|
||
interface_format: str,
|
||
api_key: str,
|
||
base_url: str,
|
||
llm_model: str,
|
||
topic: str,
|
||
genre: str,
|
||
number_of_chapters: int,
|
||
word_number: int,
|
||
filepath: str,
|
||
temperature: float = 0.7,
|
||
max_tokens: int = 2048,
|
||
timeout: int = 600
|
||
) -> None:
|
||
"""
|
||
依次调用:
|
||
1. core_seed_prompt
|
||
2. character_dynamics_prompt
|
||
3. world_building_prompt
|
||
4. plot_architecture_prompt
|
||
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)
|
||
|
||
# 加载已有的阶段性数据
|
||
partial_data = load_partial_architecture_data(filepath)
|
||
|
||
llm_adapter = create_llm_adapter(
|
||
interface_format=interface_format,
|
||
base_url=base_url,
|
||
model_name=llm_model,
|
||
api_key=api_key,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
timeout=timeout
|
||
)
|
||
|
||
# Step1: 核心种子
|
||
if "core_seed_result" not in partial_data:
|
||
logging.info("Step1: Generating core_seed_prompt (核心种子) ...")
|
||
prompt_core = core_seed_prompt.format(
|
||
topic=topic,
|
||
genre=genre,
|
||
number_of_chapters=number_of_chapters,
|
||
word_number=word_number
|
||
)
|
||
core_seed_result = invoke_with_cleaning(llm_adapter, prompt_core)
|
||
if not core_seed_result.strip():
|
||
# 多次重试依旧失败，则写入已完成内容后退出
|
||
logging.warning("core_seed_prompt generation failed and returned empty.")
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
return
|
||
partial_data["core_seed_result"] = core_seed_result
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
else:
|
||
logging.info("Step1 already done. Skipping...")
|
||
|
||
# Step2: 角色动力学
|
||
if "character_dynamics_result" not in partial_data:
|
||
logging.info("Step2: Generating character_dynamics_prompt ...")
|
||
prompt_character = character_dynamics_prompt.format(core_seed=partial_data["core_seed_result"].strip())
|
||
character_dynamics_result = invoke_with_cleaning(llm_adapter, prompt_character)
|
||
if not character_dynamics_result.strip():
|
||
logging.warning("character_dynamics_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_dynamics_result"] = character_dynamics_result
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
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 ...")
|
||
prompt_world = world_building_prompt.format(core_seed=partial_data["core_seed_result"].strip())
|
||
world_building_result = invoke_with_cleaning(llm_adapter, prompt_world)
|
||
if not world_building_result.strip():
|
||
logging.warning("world_building_prompt generation failed.")
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
return
|
||
partial_data["world_building_result"] = world_building_result
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
else:
|
||
logging.info("Step3 already done. Skipping...")
|
||
|
||
# Step4: 三幕式情节
|
||
if "plot_arch_result" not in partial_data:
|
||
logging.info("Step4: Generating plot_architecture_prompt ...")
|
||
prompt_plot = plot_architecture_prompt.format(
|
||
core_seed=partial_data["core_seed_result"].strip(),
|
||
character_dynamics=partial_data["character_dynamics_result"].strip(),
|
||
world_building=partial_data["world_building_result"].strip()
|
||
)
|
||
plot_arch_result = invoke_with_cleaning(llm_adapter, prompt_plot)
|
||
if not plot_arch_result.strip():
|
||
logging.warning("plot_architecture_prompt generation failed.")
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
return
|
||
partial_data["plot_arch_result"] = plot_arch_result
|
||
save_partial_architecture_data(filepath, partial_data)
|
||
else:
|
||
logging.info("Step4 already done. Skipping...")
|
||
|
||
# 如果能走到这里，说明全部步骤都完成了
|
||
core_seed_result = partial_data["core_seed_result"]
|
||
character_dynamics_result = partial_data["character_dynamics_result"]
|
||
world_building_result = partial_data["world_building_result"]
|
||
plot_arch_result = partial_data["plot_arch_result"]
|
||
|
||
final_content = (
|
||
"#=== 0) 小说设定 ===\n"
|
||
f"主题：{topic},类型：{genre},篇幅：约{number_of_chapters}章（每章{word_number}字）\n\n"
|
||
"#=== 1) 核心种子 ===\n"
|
||
f"{core_seed_result}\n\n"
|
||
"#=== 2) 角色动力学 ===\n"
|
||
f"{character_dynamics_result}\n\n"
|
||
"#=== 3) 世界观 ===\n"
|
||
f"{world_building_result}\n\n"
|
||
"#=== 4) 三幕式情节架构 ===\n"
|
||
f"{plot_arch_result}\n"
|
||
)
|
||
|
||
arch_file = os.path.join(filepath, "Novel_architecture.txt")
|
||
clear_file_content(arch_file)
|
||
save_string_to_txt(final_content, arch_file)
|
||
logging.info("Novel_architecture.txt has been generated successfully.")
|
||
|
||
# 全部生成完成后，可以考虑删除 partial_architecture.json，或保留做追溯
|
||
# 这里选择删除
|
||
partial_arch_file = os.path.join(filepath, "partial_architecture.json")
|
||
if os.path.exists(partial_arch_file):
|
||
os.remove(partial_arch_file)
|
||
logging.info("partial_architecture.json removed (all steps completed).")
|
||
|
||
|
||
# ============ 计算分块大小的工具函数 ============
|
||
|
||
def compute_chunk_size(number_of_chapters: int, max_tokens: int) -> int:
|
||
"""
|
||
基于“每章约100 tokens”的粗略估算，
|
||
再结合当前max_tokens，计算分块大小：
|
||
chunk_size = (floor(max_tokens/100/10)*10) - 10
|
||
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
|
||
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:
|
||
chunk_size = number_of_chapters
|
||
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(
|
||
interface_format: str,
|
||
api_key: str,
|
||
base_url: str,
|
||
llm_model: str,
|
||
filepath: str,
|
||
number_of_chapters: int,
|
||
temperature: float = 0.7,
|
||
max_tokens: int = 4096,
|
||
timeout: int = 600
|
||
) -> None:
|
||
"""
|
||
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
|
||
生成完成后输出至 Novel_directory.txt。
|
||
"""
|
||
arch_file = os.path.join(filepath, "Novel_architecture.txt")
|
||
if not os.path.exists(arch_file):
|
||
logging.warning("Novel_architecture.txt not found. Please generate architecture first.")
|
||
return
|
||
|
||
architecture_text = read_file(arch_file).strip()
|
||
if not architecture_text:
|
||
logging.warning("Novel_architecture.txt is empty.")
|
||
return
|
||
|
||
llm_adapter = create_llm_adapter(
|
||
interface_format=interface_format,
|
||
base_url=base_url,
|
||
model_name=llm_model,
|
||
api_key=api_key,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
timeout=timeout
|
||
)
|
||
|
||
filename_dir = os.path.join(filepath, "Novel_directory.txt")
|
||
if not os.path.exists(filename_dir):
|
||
# 如果文件不存在，就先建一个空文件
|
||
open(filename_dir, "w", encoding="utf-8").close()
|
||
|
||
existing_blueprint = read_file(filename_dir).strip()
|
||
chunk_size = compute_chunk_size(number_of_chapters, max_tokens)
|
||
logging.info(f"Number of chapters = {number_of_chapters}, computed chunk_size = {chunk_size}.")
|
||
|
||
# 如果已经有部分章节蓝图生成了，则进行断点续跑
|
||
if existing_blueprint:
|
||
logging.info("Detected existing blueprint content. Will resume chunked generation from that point.")
|
||
|
||
pattern = r"第\s*(\d+)\s*章"
|
||
existing_chapter_numbers = re.findall(pattern, existing_blueprint)
|
||
existing_chapter_numbers = [int(x) for x in existing_chapter_numbers if x.isdigit()]
|
||
|
||
if existing_chapter_numbers:
|
||
max_existing_chap = max(existing_chapter_numbers)
|
||
else:
|
||
max_existing_chap = 0
|
||
|
||
logging.info(f"Existing blueprint indicates up to chapter {max_existing_chap} has been generated.")
|
||
|
||
final_blueprint = existing_blueprint
|
||
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=limited_blueprint, # 只保留最近100章
|
||
number_of_chapters=number_of_chapters,
|
||
n=current_start,
|
||
m=current_end
|
||
)
|
||
logging.info(f"Generating chapters [{current_start}..{current_end}] in a chunk...")
|
||
|
||
chunk_result = invoke_with_cleaning(llm_adapter, chunk_prompt)
|
||
if not chunk_result.strip():
|
||
logging.warning(f"Chunk generation for chapters [{current_start}..{current_end}] is empty.")
|
||
# 写入当前已经有的 final_blueprint，并结束
|
||
clear_file_content(filename_dir)
|
||
save_string_to_txt(final_blueprint.strip(), filename_dir)
|
||
return
|
||
|
||
final_blueprint += "\n\n" + chunk_result.strip()
|
||
|
||
# 实时写入
|
||
clear_file_content(filename_dir)
|
||
save_string_to_txt(final_blueprint.strip(), filename_dir)
|
||
|
||
current_start = current_end + 1
|
||
|
||
logging.info("All chapters blueprint have been generated (resumed chunked).")
|
||
return
|
||
|
||
# 如果 Novel_directory.txt 为空，则分情况：
|
||
# 1) 如果 chunk_size >= number_of_chapters，可以一次性生成
|
||
if chunk_size >= number_of_chapters:
|
||
prompt = chapter_blueprint_prompt.format(
|
||
novel_architecture=architecture_text,
|
||
number_of_chapters=number_of_chapters
|
||
)
|
||
blueprint_text = invoke_with_cleaning(llm_adapter, prompt)
|
||
if not blueprint_text.strip():
|
||
logging.warning("Chapter blueprint generation result is empty.")
|
||
return
|
||
|
||
clear_file_content(filename_dir)
|
||
save_string_to_txt(blueprint_text, filename_dir)
|
||
logging.info("Novel_directory.txt (chapter blueprint) has been generated successfully (single-shot).")
|
||
return
|
||
|
||
# 2) 如果 chunk_size < number_of_chapters，则进行分块生成
|
||
logging.info("Will generate chapter blueprint in chunked mode from scratch.")
|
||
final_blueprint = ""
|
||
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=limited_blueprint, # 只保留最近100章
|
||
number_of_chapters=number_of_chapters,
|
||
n=current_start,
|
||
m=current_end
|
||
)
|
||
logging.info(f"Generating chapters [{current_start}..{current_end}] in a chunk...")
|
||
|
||
chunk_result = invoke_with_cleaning(llm_adapter, chunk_prompt)
|
||
if not chunk_result.strip():
|
||
logging.warning(f"Chunk generation for chapters [{current_start}..{current_end}] is empty.")
|
||
# 写入已经生成的 final_blueprint
|
||
clear_file_content(filename_dir)
|
||
save_string_to_txt(final_blueprint.strip(), filename_dir)
|
||
return
|
||
|
||
if final_blueprint.strip():
|
||
final_blueprint += "\n\n" + chunk_result.strip()
|
||
else:
|
||
final_blueprint = chunk_result.strip()
|
||
|
||
# 实时写入，以免中途崩溃造成丢失
|
||
clear_file_content(filename_dir)
|
||
save_string_to_txt(final_blueprint.strip(), filename_dir)
|
||
|
||
current_start = current_end + 1
|
||
|
||
logging.info("Novel_directory.txt (chapter blueprint) has been generated successfully (chunked).")
|
||
|
||
|
||
# ============ 3) 生成章节草稿 ============
|
||
|
||
def generate_chapter_draft(
|
||
api_key: str,
|
||
base_url: str,
|
||
model_name: str,
|
||
filepath: str,
|
||
novel_number: int,
|
||
word_number: int,
|
||
temperature: float,
|
||
user_guidance: str,
|
||
characters_involved: str,
|
||
key_items: str,
|
||
scene_location: str,
|
||
time_constraint: str,
|
||
embedding_api_key: str,
|
||
embedding_url: str,
|
||
embedding_interface_format: str,
|
||
embedding_model_name: str,
|
||
embedding_retrieval_k: int = 2,
|
||
interface_format: str = "openai",
|
||
max_tokens: int = 2048,
|
||
timeout: int = 600
|
||
) -> str:
|
||
"""
|
||
根据 novel_number 判断是否为第一章。
|
||
- 若是第一章，则使用 first_chapter_draft_prompt
|
||
- 否则使用 next_chapter_draft_prompt
|
||
最终将生成文本存入 chapters/chapter_{novel_number}.txt。
|
||
"""
|
||
arch_file = os.path.join(filepath, "Novel_architecture.txt")
|
||
novel_architecture_text = read_file(arch_file)
|
||
|
||
directory_file = os.path.join(filepath, "Novel_directory.txt")
|
||
blueprint_text = read_file(directory_file)
|
||
|
||
global_summary_file = os.path.join(filepath, "global_summary.txt")
|
||
global_summary_text = read_file(global_summary_file)
|
||
|
||
character_state_file = os.path.join(filepath, "character_state.txt")
|
||
character_state_text = read_file(character_state_file)
|
||
|
||
# 获取本章在目录中的信息
|
||
chapter_info = get_chapter_info_from_blueprint(blueprint_text, novel_number)
|
||
chapter_title = chapter_info["chapter_title"]
|
||
chapter_role = chapter_info["chapter_role"]
|
||
chapter_purpose = chapter_info["chapter_purpose"]
|
||
suspense_level = chapter_info["suspense_level"]
|
||
foreshadowing = chapter_info["foreshadowing"]
|
||
plot_twist_level = chapter_info["plot_twist_level"]
|
||
chapter_summary = chapter_info["chapter_summary"]
|
||
|
||
# 准备章节目录文件夹
|
||
chapters_dir = os.path.join(filepath, "chapters")
|
||
os.makedirs(chapters_dir, exist_ok=True)
|
||
|
||
# 判断是否为第一章
|
||
if novel_number == 1:
|
||
prompt_text = first_chapter_draft_prompt.format(
|
||
novel_number=novel_number,
|
||
word_number=word_number,
|
||
chapter_title=chapter_title,
|
||
chapter_role=chapter_role,
|
||
chapter_purpose=chapter_purpose,
|
||
suspense_level=suspense_level,
|
||
foreshadowing=foreshadowing,
|
||
plot_twist_level=plot_twist_level,
|
||
chapter_summary=chapter_summary,
|
||
|
||
characters_involved=characters_involved,
|
||
key_items=key_items,
|
||
scene_location=scene_location,
|
||
time_constraint=time_constraint,
|
||
user_guidance=user_guidance,
|
||
|
||
novel_setting=novel_architecture_text
|
||
)
|
||
else:
|
||
# 若不是第一章，则获取最近几章文本，并做摘要与检索
|
||
recent_3_texts = get_last_n_chapters_text(chapters_dir, novel_number, n=3)
|
||
short_summary, next_chapter_keywords = summarize_recent_chapters(
|
||
interface_format=interface_format,
|
||
api_key=api_key,
|
||
base_url=base_url,
|
||
model_name=model_name,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
chapters_text_list=recent_3_texts,
|
||
timeout=timeout
|
||
)
|
||
|
||
# 从最近章节中获取最后一段作为前章结尾
|
||
previous_chapter_excerpt = ""
|
||
for text_block in reversed(recent_3_texts):
|
||
if text_block.strip():
|
||
# 取后1500字符左右
|
||
if len(text_block) > 1500:
|
||
previous_chapter_excerpt = text_block[-1500:]
|
||
else:
|
||
previous_chapter_excerpt = text_block
|
||
break
|
||
|
||
# 从向量库检索上下文
|
||
embedding_adapter = create_embedding_adapter(
|
||
embedding_interface_format,
|
||
embedding_api_key,
|
||
embedding_url,
|
||
embedding_model_name
|
||
)
|
||
retrieval_query = short_summary + " " + next_chapter_keywords
|
||
relevant_context = get_relevant_context_from_vector_store(
|
||
embedding_adapter=embedding_adapter,
|
||
query=retrieval_query,
|
||
filepath=filepath,
|
||
k=embedding_retrieval_k
|
||
)
|
||
if not relevant_context.strip():
|
||
relevant_context = "（无检索到的上下文）"
|
||
|
||
prompt_text = next_chapter_draft_prompt.format(
|
||
novel_number=novel_number,
|
||
word_number=word_number,
|
||
chapter_title=chapter_title,
|
||
chapter_role=chapter_role,
|
||
chapter_purpose=chapter_purpose,
|
||
suspense_level=suspense_level,
|
||
foreshadowing=foreshadowing,
|
||
plot_twist_level=plot_twist_level,
|
||
chapter_summary=chapter_summary,
|
||
|
||
characters_involved=characters_involved,
|
||
key_items=key_items,
|
||
scene_location=scene_location,
|
||
time_constraint=time_constraint,
|
||
user_guidance=user_guidance,
|
||
|
||
novel_setting=novel_architecture_text,
|
||
global_summary=global_summary_text,
|
||
character_state=character_state_text,
|
||
context_excerpt=relevant_context,
|
||
previous_chapter_excerpt=previous_chapter_excerpt
|
||
)
|
||
|
||
llm_adapter = create_llm_adapter(
|
||
interface_format=interface_format,
|
||
base_url=base_url,
|
||
model_name=model_name,
|
||
api_key=api_key,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
timeout=timeout
|
||
)
|
||
|
||
chapter_content = invoke_with_cleaning(llm_adapter, prompt_text)
|
||
if not chapter_content.strip():
|
||
logging.warning("Generated chapter draft is empty.")
|
||
|
||
chapter_file = os.path.join(chapters_dir, f"chapter_{novel_number}.txt")
|
||
clear_file_content(chapter_file)
|
||
save_string_to_txt(chapter_content, chapter_file)
|
||
|
||
logging.info(f"[Draft] Chapter {novel_number} generated as a draft.")
|
||
return chapter_content
|
||
|
||
|
||
# ============ 4) 定稿章节 ============
|
||
|
||
def finalize_chapter(
|
||
novel_number: int,
|
||
word_number: int,
|
||
api_key: str,
|
||
base_url: str,
|
||
model_name: str,
|
||
temperature: float,
|
||
filepath: str,
|
||
embedding_api_key: str,
|
||
embedding_url: str,
|
||
embedding_interface_format: str,
|
||
embedding_model_name: str,
|
||
interface_format: str,
|
||
max_tokens: int,
|
||
timeout: int = 600
|
||
):
|
||
"""
|
||
对指定章节做最终处理：更新全局摘要、更新角色状态、插入向量库等。
|
||
默认无需再做扩写操作，若有需要可在外部调用 enrich_chapter_text 处理后再定稿。
|
||
"""
|
||
chapters_dir = os.path.join(filepath, "chapters")
|
||
chapter_file = os.path.join(chapters_dir, f"chapter_{novel_number}.txt")
|
||
chapter_text = read_file(chapter_file).strip()
|
||
if not chapter_text:
|
||
logging.warning(f"Chapter {novel_number} is empty, cannot finalize.")
|
||
return
|
||
|
||
# 进行摘要、角色状态更新
|
||
global_summary_file = os.path.join(filepath, "global_summary.txt")
|
||
old_global_summary = read_file(global_summary_file)
|
||
|
||
character_state_file = os.path.join(filepath, "character_state.txt")
|
||
old_character_state = read_file(character_state_file)
|
||
|
||
llm_adapter = create_llm_adapter(
|
||
interface_format=interface_format,
|
||
base_url=base_url,
|
||
model_name=model_name,
|
||
api_key=api_key,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
timeout=timeout
|
||
)
|
||
|
||
# 更新全局摘要
|
||
prompt_summary = summary_prompt.format(
|
||
chapter_text=chapter_text,
|
||
global_summary=old_global_summary
|
||
)
|
||
new_global_summary = invoke_with_cleaning(llm_adapter, prompt_summary)
|
||
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
|
||
)
|
||
new_char_state = invoke_with_cleaning(llm_adapter, prompt_char_state)
|
||
if not new_char_state.strip():
|
||
new_char_state = old_character_state
|
||
|
||
clear_file_content(global_summary_file)
|
||
save_string_to_txt(new_global_summary, global_summary_file)
|
||
|
||
clear_file_content(character_state_file)
|
||
save_string_to_txt(new_char_state, character_state_file)
|
||
|
||
# 更新向量库
|
||
embedding_adapter = create_embedding_adapter(
|
||
embedding_interface_format,
|
||
embedding_api_key,
|
||
embedding_url,
|
||
embedding_model_name
|
||
)
|
||
update_vector_store(embedding_adapter, chapter_text, filepath)
|
||
|
||
logging.info(f"Chapter {novel_number} has been finalized.")
|
||
|
||
|
||
def enrich_chapter_text(
|
||
chapter_text: str,
|
||
word_number: int,
|
||
api_key: str,
|
||
base_url: str,
|
||
model_name: str,
|
||
temperature: float,
|
||
interface_format: str,
|
||
max_tokens: int,
|
||
timeout: int=600
|
||
) -> str:
|
||
"""
|
||
对章节文本进行扩写，使其更接近 word_number 字数，保持剧情连贯。
|
||
"""
|
||
llm_adapter = create_llm_adapter(
|
||
interface_format=interface_format,
|
||
base_url=base_url,
|
||
model_name=model_name,
|
||
api_key=api_key,
|
||
temperature=temperature,
|
||
max_tokens=max_tokens,
|
||
timeout=timeout
|
||
)
|
||
prompt = f"""以下章节文本较短，请在保持剧情连贯的前提下进行扩写，使其更充实，接近 {word_number} 字左右：
|
||
原内容：
|
||
{chapter_text}
|
||
"""
|
||
enriched_text = invoke_with_cleaning(llm_adapter, prompt)
|
||
return enriched_text if enriched_text else chapter_text
|
||
|
||
|
||
# ============ 导入知识文件到向量库 ============
|
||
|
||
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 []
|
||
|
||
model = SentenceTransformer('paraphrase-MiniLM-L6-v2')
|
||
embeddings = model.encode(sentences)
|
||
|
||
merged_paragraphs = []
|
||
current_sentences = [sentences[0]]
|
||
current_embedding = embeddings[0]
|
||
|
||
for i in range(1, len(sentences)):
|
||
sim = cosine_similarity([current_embedding], [embeddings[i]])[0][0]
|
||
if sim >= similarity_threshold:
|
||
current_sentences.append(sentences[i])
|
||
current_embedding = (current_embedding + embeddings[i]) / 2.0
|
||
else:
|
||
merged_paragraphs.append(" ".join(current_sentences))
|
||
current_sentences = [sentences[i]]
|
||
current_embedding = embeddings[i]
|
||
|
||
if current_sentences:
|
||
merged_paragraphs.append(" ".join(current_sentences))
|
||
|
||
final_segments = []
|
||
for para in merged_paragraphs:
|
||
if len(para) > max_length:
|
||
sub_segments = split_by_length(para, max_length=max_length)
|
||
final_segments.extend(sub_segments)
|
||
else:
|
||
final_segments.append(para)
|
||
|
||
return final_segments
|
||
|
||
def import_knowledge_file(
|
||
embedding_api_key: str,
|
||
embedding_url: str,
|
||
embedding_interface_format: str,
|
||
embedding_model_name: str,
|
||
file_path: str,
|
||
filepath: str
|
||
):
|
||
logging.info(f"开始导入知识库文件: {file_path}, 接口格式: {embedding_interface_format}, 模型: {embedding_model_name}")
|
||
if not os.path.exists(file_path):
|
||
logging.warning(f"知识库文件不存在: {file_path}")
|
||
return
|
||
|
||
content = read_file(file_path)
|
||
if not content.strip():
|
||
logging.warning("知识库文件内容为空。")
|
||
return
|
||
|
||
paragraphs = advanced_split_content(content)
|
||
|
||
embedding_adapter = create_embedding_adapter(
|
||
interface_format=embedding_interface_format,
|
||
api_key=embedding_api_key,
|
||
base_url=embedding_url if embedding_url else "http://localhost:11434/api",
|
||
model_name=embedding_model_name
|
||
)
|
||
|
||
store = load_vector_store(embedding_adapter, filepath)
|
||
if not store:
|
||
logging.info("Vector store does not exist or load failed. Initializing a new one for knowledge import...")
|
||
store = init_vector_store(embedding_adapter, paragraphs, filepath)
|
||
if store:
|
||
logging.info("知识库文件已成功导入至向量库(新初始化)。")
|
||
else:
|
||
logging.warning("知识库导入失败，跳过。")
|
||
else:
|
||
try:
|
||
docs = [Document(page_content=str(p)) for p in paragraphs]
|
||
store.add_documents(docs)
|
||
logging.info("知识库文件已成功导入至向量库(追加模式)。")
|
||
except Exception as e:
|
||
logging.warning(f"知识库导入失败: {e}")
|
||
traceback.print_exc()
|