Losita/mai-bot
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat：优化重复思考，表情现在一次选所有

Browse Source
This commit is contained in:
SengokuCola
2026-04-04 01:48:52 +08:00
parent f807295653
commit 97fb4cb36f
2 changed files with 39 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

89
src/maisaka/builtin_tool/send_emoji.py
View File

@@ -3,7 +3,6 @@
from datetime import datetime
from io import BytesIO
from random import sample
from secrets import token_hex
from typing import Any, Dict, Optional
import asyncio
@@ -32,8 +31,7 @@ logger = get_logger("maisaka_builtin_send_emoji")
_EMOJI_SUB_AGENT_CONTEXT_LIMIT = 12
_EMOJI_SUB_AGENT_MAX_TOKENS = 240
_EMOJI_CANDIDATE_GROUP_COUNT = 3
_EMOJI_CANDIDATES_PER_GROUP = 5
_EMOJI_CANDIDATE_COUNT = 25
_EMOJI_CANDIDATE_TILE_SIZE = 256
_EMOJI_SUCCESS_MESSAGE = "表情包发送成功"
@@ -41,8 +39,8 @@ _EMOJI_SUCCESS_MESSAGE = "表情包发送成功"
class EmojiSelectionResult(BaseModel):
"""表情包子代理的结构化选择结果。"""
emoji_id: str = PydanticField(default="", description="选中的候选消息 ID。")
emoji_index: int = PydanticField(default=1, description="该候选消息中第几张图片，从 1 开始计数。")
emoji_index: int = PydanticField(default=1, description="选中的表情包序号，从 1 开始计数。")
reason: str = PydanticField(default="", description="选择这张表情包的简短理由。")
def get_tool_spec() -> ToolSpec:
@@ -136,44 +134,46 @@ def _build_labeled_tile(image_bytes: bytes, index: int, tile_size: int) -> PILIm
def _merge_emoji_tiles(image_bytes_list: list[bytes]) -> bytes:
"""将三张候选表情图拼接成一张横向图片。"""
"""将候选表情图拼接成一张 5x5 网格图片。"""
tile_size = _EMOJI_CANDIDATE_TILE_SIZE
gap = 12
grid_columns = 5
grid_rows = 5
tiles = [
_build_labeled_tile(image_bytes=image_bytes, index=index, tile_size=tile_size)
for index, image_bytes in enumerate(image_bytes_list, start=1)
]
canvas_width = tile_size * len(tiles) + gap * max(len(tiles) - 1, 0)
canvas = PILImage.new("RGBA", (canvas_width, tile_size), color=(255, 255, 255, 255))
canvas_width = tile_size * grid_columns + gap * (grid_columns - 1)
canvas_height = tile_size * grid_rows + gap * (grid_rows - 1)
canvas = PILImage.new("RGBA", (canvas_width, canvas_height), color=(255, 255, 255, 255))
current_x = 0
for tile in tiles:
canvas.paste(tile, (current_x, 0), tile)
current_x += tile_size + gap
for index, tile in enumerate(tiles):
row = index // grid_columns
column = index % grid_columns
offset_x = column * (tile_size + gap)
offset_y = row * (tile_size + gap)
canvas.paste(tile, (offset_x, offset_y), tile)
output = BytesIO()
canvas.convert("RGB").save(output, format="PNG")
return output.getvalue()
async def _build_emoji_candidate_message(
emojis: list[MaiEmoji],
candidate_id: str,
) -> SessionBackedMessage:
"""构建供子代理挑选的拼图候选消息。"""
async def _build_emoji_candidate_message(emojis: list[MaiEmoji]) -> SessionBackedMessage:
"""构建供子代理挑选的 5x5 拼图候选消息。"""
image_bytes_list = await asyncio.gather(*[_load_emoji_bytes(emoji) for emoji in emojis])
merged_image_bytes = await asyncio.to_thread(_merge_emoji_tiles, list(image_bytes_list))
raw_message = MessageSequence(
[
TextComponent(f"ID: {candidate_id}"),
TextComponent("请从这张 5x5 拼图中选择一个序号。"),
ImageComponent(binary_hash="", binary_data=merged_image_bytes),
]
)
return SessionBackedMessage(
raw_message=raw_message,
visible_text=f"ID: {candidate_id}",
visible_text="[表情包拼图候选]",
timestamp=datetime.now(),
source_kind="emoji_candidate",
)
@@ -185,6 +185,7 @@ async def _select_emoji_with_sub_agent(
reasoning: str,
context_texts: list[str],
sample_size: int,
selection_metadata: Optional[Dict[str, str]] = None,
) -> tuple[MaiEmoji | None, str]:
"""通过临时子代理从候选表情包中选出一个结果。"""
@@ -194,33 +195,18 @@ async def _select_emoji_with_sub_agent(
if not available_emojis:
return None, ""
total_candidate_count = min(
len(available_emojis),
_EMOJI_CANDIDATE_GROUP_COUNT * _EMOJI_CANDIDATES_PER_GROUP,
)
total_candidate_count = min(len(available_emojis), _EMOJI_CANDIDATE_COUNT)
sampled_emojis = sample(available_emojis, total_candidate_count)
candidate_map: dict[str, list[MaiEmoji]] = {}
candidate_messages: list[LLMContextMessage] = []
for group_index in range(0, len(sampled_emojis), _EMOJI_CANDIDATES_PER_GROUP):
emoji_group = sampled_emojis[group_index : group_index + _EMOJI_CANDIDATES_PER_GROUP]
if not emoji_group:
continue
candidate_id = token_hex(4)
while candidate_id in candidate_map:
candidate_id = token_hex(4)
candidate_map[candidate_id] = emoji_group
candidate_messages.append(await _build_emoji_candidate_message(emoji_group, candidate_id))
candidate_message = await _build_emoji_candidate_message(sampled_emojis)
system_prompt = (
"你是 Maisaka 的临时表情包选择子代理。\n"
"你会收到群聊上下文，以及 3 条候选消息。每条候选消息都包含 5 张横向拼接的表情图。\n"
"每条候选消息都有一个临时 ID，图片左上角标有 1、2、3、4、5，对应这条消息中的第 1 到第 5 张图。\n"
"你的任务是根据上下文和当前语气，从候选中选出最合适的一张表情包。\n"
"你会收到群聊上下文，以及 1 条额外候选消息，其中包含一张 5x5 的表情包拼图，一共 25 个位置。\n"
"每张小图左上角都有一个较大的序号，范围是 1 到 25。\n"
"你的任务是根据上下文和当前语气，从这 25 张图里选出最合适的一张表情包。\n"
"如果提供了 requested_emotion，请优先考虑与其接近的候选；如果没有完全匹配，则选择最符合上下文语气的候选。\n"
"你必须返回一个 JSON 对象（json object），不要输出任何 JSON 之外的内容。\n"
'返回格式固定为：{"emoji_id":"候选消息ID","emoji_index":1}'
'返回格式固定为：{"emoji_index":1,"reason":"简短理由"}'
)
prompt_message = ReferenceMessage(
content=(
@@ -237,7 +223,7 @@ async def _select_emoji_with_sub_agent(
response = await tool_ctx.runtime.run_sub_agent(
context_message_limit=_EMOJI_SUB_AGENT_CONTEXT_LIMIT,
system_prompt=system_prompt,
extra_messages=[prompt_message, *candidate_messages],
extra_messages=[prompt_message, candidate_message],
max_tokens=_EMOJI_SUB_AGENT_MAX_TOKENS,
response_format=RespFormat(
format_type=RespFormatType.JSON_SCHEMA,
@@ -252,22 +238,17 @@ async def _select_emoji_with_sub_agent(
fallback_emoji = sampled_emojis[0] if sampled_emojis else None
return fallback_emoji, ""
selected_group = candidate_map.get(selection.emoji_id.strip())
if selected_group is None:
logger.warning(
f"{tool_ctx.runtime.log_prefix} 表情包子代理返回了无效 ID: {selection.emoji_id!r}，将回退到候选首项"
)
fallback_emoji = sampled_emojis[0] if sampled_emojis else None
return fallback_emoji, ""
if selection_metadata is not None:
selection_metadata["reason"] = selection.reason.strip()
emoji_index = int(selection.emoji_index)
if emoji_index < 1 or emoji_index > len(selected_group):
if emoji_index < 1 or emoji_index > len(sampled_emojis):
logger.warning(
f"{tool_ctx.runtime.log_prefix} 表情包子代理返回了无效序号: {emoji_index!r}，将回退到该组第 1 张"
f"{tool_ctx.runtime.log_prefix} 表情包子代理返回了无效序号: {emoji_index!r}，将回退到第 1 张"
)
emoji_index = 1
return selected_group[emoji_index - 1], ""
return sampled_emojis[emoji_index - 1], ""
async def handle_tool(
@@ -291,7 +272,9 @@ async def handle_tool(
"emotion": [],
"requested_emotion": emotion,
"matched_emotion": "",
"reason": "",
}
selection_metadata: Dict[str, str] = {"reason": ""}
logger.info(f"{tool_ctx.runtime.log_prefix} 触发表情包发送工具，请求情绪={emotion!r}")
@@ -307,6 +290,7 @@ async def handle_tool(
reasoning,
list(context_texts or []),
sample_size,
selection_metadata,
),
)
except Exception as exc:
@@ -320,6 +304,7 @@ async def handle_tool(
if send_result.success:
structured_result["message"] = _EMOJI_SUCCESS_MESSAGE
structured_result["reason"] = selection_metadata["reason"]
logger.info(
f"{tool_ctx.runtime.log_prefix} 表情包发送成功 "
f"描述={send_result.description!r} 情绪标签={send_result.emotions} "
@@ -332,7 +317,7 @@ async def handle_tool(
structured_result["success"] = True
return tool_ctx.build_success_result(
invocation.tool_name,
_EMOJI_SUCCESS_MESSAGE,
selection_metadata["reason"] or _EMOJI_SUCCESS_MESSAGE,
structured_content=structured_result,
)

4
src/maisaka/reasoning_engine.py
View File

@@ -134,8 +134,8 @@ class MaisakaReasoningEngine:
reasoning_content = response.content or ""
if self._should_replace_reasoning(reasoning_content):
response.content = "让我根据新情况重新思考："
response.raw_message.content = "让我根据新情况重新思考："
response.content = "我应该根据我上面思考的内容进行反思，重新思考我下一步的行动，我需要分析当前场景，对话，以及我可以使用的工具，然后先输出想法再使用工具"
response.raw_message.content = "我应该根据我上面思考的内容进行反思，重新思考我下一步的行动，我需要分析当前场景，对话，以及我可以使用的工具，然后先输出想法再使用工具"
logger.info(f"{self._runtime.log_prefix} 当前思考与上一轮过于相似，已替换为重新思考提示")
self._last_reasoning_content = reasoning_content