Files
SillyTavern_replica/backend/services/prompt_assembler.py

227 lines
8.4 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""
提示词组装器 (Prompt Assembler)
负责根据 SillyTavern 规范将角色卡、世界书、聊天历史等组件
拼装成最终的 LLM 消息列表。
"""
import re
from typing import List, Dict, Optional
from langchain_core.messages import SystemMessage, HumanMessage, AIMessage, BaseMessage
from models.internal import CharacterCard, ChatMessage, WorldInfoEntry
class PromptConfig:
"""提示词组装配置"""
def __init__(
self,
an_position: str = "after_history", # "before_history" or "after_history"
an_depth: int = 4,
post_history_instructions: Optional[str] = None
):
self.an_position = an_position
self.an_depth = an_depth
self.post_history_instructions = post_history_instructions
class PromptAssembler:
"""
轻量级提示词组装核心
不依赖复杂的框架,只负责纯粹的文本拼接和位置插入。
"""
# SillyTavern 的位置枚举映射
POS_WI_BEFORE = 0
POS_WI_AFTER = 1
POS_EXAMPLES_BEFORE = 2
POS_EXAMPLES_AFTER = 3
POS_AN_TOP = 4
POS_AN_BOTTOM = 5
POS_DEPTH = 6
POS_OUTLET = 7
def assemble(
self,
character: CharacterCard,
chat_history: List[ChatMessage],
user_input: str,
active_entries: List[WorldInfoEntry],
config: PromptConfig = PromptConfig()
) -> List[BaseMessage]:
"""
执行完整的提示词组装流程
Returns:
List[BaseMessage]: 准备好发送给 LLM 的消息列表
"""
# 1. 按位置分组世界书条目
grouped_entries = self._group_entries_by_position(active_entries)
# 2. 组装 Story String (包含 Pos 0-3)
story_string = self._build_story_string(character, grouped_entries)
# 3. 组装 Author's Note (包含 Pos 4-5)
authors_note_content = self._build_authors_note(grouped_entries, config.an_depth)
# 4. 处理 Chat History 并注入 Depth 条目 (Pos 6)
processed_history = self._inject_depth_entries(chat_history, grouped_entries.get(self.POS_DEPTH, []))
# 5. 准备 Outlet 替换字典 (Pos 7)
outlet_map = {entry.uid: entry.content for entry in grouped_entries.get(self.POS_OUTLET, [])}
# 6. 最终封装为 Messages
return self._wrap_to_messages(
story_string,
authors_note_content,
processed_history,
user_input,
outlet_map,
config
)
def _group_entries_by_position(self, entries: List[WorldInfoEntry]) -> Dict[int, List[WorldInfoEntry]]:
"""将激活的条目按 position 分组"""
grouped = {}
for entry in entries:
# 这里假设 entry.position 存储的是我们定义的 0-7 整数
pos = entry.position if isinstance(entry.position, int) else 1 # 默认为 wiAfter
if pos not in grouped:
grouped[pos] = []
grouped[pos].append(entry)
# 对每个组内的条目按 order 排序
for pos in grouped:
grouped[pos].sort(key=lambda x: x.order)
return grouped
def _build_story_string(self, character: CharacterCard, grouped: Dict) -> str:
"""组装故事字符串 (Story String)"""
parts = []
# Pos 0: wiBefore
for entry in grouped.get(self.POS_WI_BEFORE, []):
parts.append(entry.content)
# 角色核心信息
parts.append(f"[Character('{character.name}')]\n{character.description}\n")
parts.append(f"Personality: {character.personality}\n")
parts.append(f"Scenario: {character.scenario}\n")
# Pos 1: wiAfter
for entry in grouped.get(self.POS_WI_AFTER, []):
parts.append(entry.content)
# Pos 2: Examples Before
for entry in grouped.get(self.POS_EXAMPLES_BEFORE, []):
parts.append(entry.content)
# 示例对话
if character.mes_example:
parts.append(f"<START>\n{character.mes_example}")
# Pos 3: Examples After
for entry in grouped.get(self.POS_EXAMPLES_AFTER, []):
parts.append(entry.content)
return "\n".join(parts)
def _build_authors_note(self, grouped: Dict, depth: int) -> str:
"""组装作者笔记 (Author's Note)"""
parts = []
# Pos 4: AN Top
for entry in grouped.get(self.POS_AN_TOP, []):
parts.append(str(entry.content) if entry.content else "")
# AN 核心内容 (这里简化为一个占位,实际应从角色卡或设置获取)
parts.append(f"[Author's note at depth {depth}]")
# Pos 5: AN Bottom
for entry in grouped.get(self.POS_AN_BOTTOM, []):
parts.append(str(entry.content) if entry.content else "")
return "\n".join(parts)
def _inject_depth_entries(self, history: List[ChatMessage], depth_entries: List[WorldInfoEntry]) -> List[Dict]:
"""
在聊天历史的指定深度插入条目 (Pos 6)
返回一个包含 role 和 content 的字典列表,方便后续转换
✅ 过滤已被总结的消息is_summarized=True 且 mes=""
"""
# 先将历史转换为中间格式,过滤掉空消息(已被总结)
msg_list = []
for msg in history:
# ✅ 跳过已被总结的空消息
if msg.is_summarized and (msg.mes == "" or msg.mes.strip() == ""):
continue
msg_list.append({"role": "user" if msg.is_user else "assistant", "content": msg.mes})
# 按 depth 分组插入
# d0 通常指最新用户输入之前,即列表末尾
for entry in depth_entries:
depth = entry.depth if entry.depth is not None else 0
# 计算插入索引 (从后往前数)
insert_index = max(0, len(msg_list) - depth)
# 确定角色
role_map = {"system": "system", "user": "user", "assistant": "assistant"}
role = role_map.get(str(entry.position).split('_')[-1] if '_' in str(entry.position) else "system", "system")
msg_list.insert(insert_index, {"role": "system", "content": entry.content})
return msg_list
def _replace_outlets(self, text: str, outlet_map: Dict[str, str]) -> str:
"""执行 Outlet 宏替换 (Pos 7)"""
def replacer(match):
uid = match.group(1)
return outlet_map.get(uid, "")
# 匹配 {{outlet::UID}}
return re.sub(r"\{\{outlet::([^}]+)\}\}", replacer, text)
def _wrap_to_messages(
self,
story_string: str,
an_content: str,
history: List[Dict],
user_input: str,
outlet_map: Dict[str, str],
config: PromptConfig
) -> List[BaseMessage]:
"""将组装好的文本块封装为 LangChain Messages"""
messages = []
# 1. System Message (Story String + Outlet 替换)
final_story = self._replace_outlets(story_string, outlet_map)
if final_story:
messages.append(SystemMessage(content=final_story))
# 2. Author's Note (根据配置位置插入)
if an_content and config.an_position == "before_history":
messages.append(SystemMessage(content=self._replace_outlets(an_content, outlet_map)))
# 3. Chat History
for msg_data in history:
if msg_data["role"] == "user":
messages.append(HumanMessage(content=msg_data["content"]))
elif msg_data["role"] == "assistant":
messages.append(AIMessage(content=msg_data["content"]))
else:
messages.append(SystemMessage(content=msg_data["content"]))
# 4. Author's Note (如果在 History 之后)
if an_content and config.an_position == "after_history":
messages.append(SystemMessage(content=self._replace_outlets(an_content, outlet_map)))
# 5. Post-History Instructions & User Input
final_input = user_input
if config.post_history_instructions:
final_input = f"{config.post_history_instructions}\n\n{user_input}"
messages.append(HumanMessage(content=final_input))
return messages