416 lines
18 KiB
Python
416 lines
18 KiB
Python
import json
|
||
import os
|
||
import time
|
||
import datetime
|
||
import uuid
|
||
from pathlib import Path
|
||
from typing import Any, Dict, Iterable, List, Optional, Sequence, Union
|
||
|
||
import requests
|
||
from qwen_agent.agents import Assistant
|
||
from qwen_agent.llm.schema import ContentItem, Message
|
||
|
||
import agent_runtime # noqa: F401
|
||
from agent_runtime import readonly_tools # noqa: F401
|
||
|
||
# 强制注册 write_file 工具，防止 build_agent 时报错
|
||
try:
|
||
import agent_runtime.write_tools
|
||
except ImportError:
|
||
pass
|
||
|
||
DEFAULT_SYSTEM_PROMPT = (
|
||
'你是 Qwen3.5，本地部署的多模态中文助手。\n'
|
||
'默认中文回答。\n'
|
||
'当用户只是打招呼或闲聊时，自然回应即可，不要主动枚举全部工具。\n'
|
||
'你的目标是先使用可用工具获得可验证信息，再给出结论。\n'
|
||
'规则:\n'
|
||
'1. 对最新信息先用 web_search，再按需用 web_fetch 或 web_extractor 抓取正文。\n'
|
||
'2. 对人名、作品名、小众概念等不确定知识先 web_search，若结果歧义则改写关键词再检索一次。\n'
|
||
'3. 核心规则：你已具备 filesystem 的读写能力。你可以读取文件，如果用户有需求，你也可以调用 write_file 工具进行写入。\n'
|
||
'4. 图片问题先看整图，细节再用 image_zoom_in_tool，使用相对坐标。\n'
|
||
'5. 工具失败时必须明确说明原因，不得伪造结果。\n'
|
||
'6. 联网任务要控制上下文预算，优先少量高质量来源。\n'
|
||
'7. 严禁在未获授权的情况下使用 filesystem 工具查看助手自身的源代码或运行目录。\n'
|
||
'8. 联网任务要控制上下文预算，优先少量高质量来源，避免搬运大段无关正文。\n'
|
||
'9. 长期记忆（主动意识）：你拥有 manage_memory 工具。当你从对话中识别出以下内容时，必须【主动】调用 add 操作：\n - 用户的明确偏好（如：喜欢 MD 格式、不喜欢繁琐说明）。\n - 重要的个人事实（如：职业、项目代号、系统配置路径）。\n - 约定的工作习惯（如：每段代码都要加注释）。\n 执行后，在回复中自然地告知用户“我已记下此习惯/信息”。当用户问“你了解我什么”或要求修改时，配合 list 和 delete 操作。\n'
|
||
)
|
||
|
||
DEFAULT_FUNCTION_LIST = [
|
||
'web_search',
|
||
'web_fetch',
|
||
'web_extractor',
|
||
'image_search',
|
||
'image_zoom_in_tool',
|
||
'filesystem',
|
||
'manage_memory',
|
||
]
|
||
|
||
TIMINGS_EMIT_INTERVAL_SEC = 0.8
|
||
MAX_FALLBACK_PART_TEXT_CHARS = 512
|
||
|
||
# ==========================================
|
||
# 🧠 记忆系统：全兼容自愈逻辑
|
||
# ==========================================
|
||
def get_injected_memory() -> str:
|
||
"""从环境变量指定的路径动态加载记忆，支持多种 JSON 格式"""
|
||
raw_path = os.getenv('MEMORY_FILE_PATH', './.tmp/super_agent_data/memory.json')
|
||
# 彻底剥离可能存在的换行符或引号
|
||
clean_path = raw_path.strip().strip('"').strip("'")
|
||
memory_path = Path(clean_path).resolve()
|
||
|
||
if not memory_path.exists():
|
||
return ""
|
||
try:
|
||
with open(memory_path, 'r', encoding='utf-8') as f:
|
||
data = json.load(f)
|
||
|
||
# 兼容：直接是列表 [item1, item2] 或 字典 {"items": [item1, ...]}
|
||
if isinstance(data, dict) and "items" in data:
|
||
raw_list = data["items"]
|
||
elif isinstance(data, list):
|
||
raw_list = data
|
||
else:
|
||
return ""
|
||
|
||
processed = []
|
||
for item in raw_list:
|
||
# 兼容项：直接是字符串 或 带 content 键的字典
|
||
val = item.get("content") if isinstance(item, dict) else item
|
||
if val: processed.append(str(val))
|
||
|
||
if not processed: return ""
|
||
memory_str = "\n".join([f"- {m}" for m in processed])
|
||
return f"\n【长期记忆库（已自动加载）】:\n{memory_str}\n"
|
||
except Exception as e:
|
||
print(f"[Memory] 记忆注入提示: 暂未读取到有效记忆或格式不匹配 ({e})")
|
||
return ""
|
||
|
||
def fetch_model_id(model_server: str, timeout_sec: int) -> str:
|
||
try:
|
||
response = requests.get(f'{model_server}/models', timeout=2)
|
||
response.raise_for_status()
|
||
return response.json()['data'][0]['id']
|
||
except Exception:
|
||
return "local-model"
|
||
|
||
# ==========================================
|
||
# 🚥 路由系统：指令拦截
|
||
# ==========================================
|
||
def extract_routing_command(messages: list) -> Optional[str]:
|
||
"""从后往前扫描整段对话历史，寻找最近的路由指令"""
|
||
if not messages:
|
||
return None
|
||
|
||
# 1. 优先检查最后一条消息（即时指令优先级最高）
|
||
last_msg = messages[-1]
|
||
if last_msg.get('role') == 'user':
|
||
content = last_msg.get('content', '')
|
||
if isinstance(content, str):
|
||
text = content.strip()
|
||
if text.startswith('/cloud'):
|
||
last_msg['content'] = text[len('/cloud'):].strip()
|
||
return 'FORCE_CLOUD'
|
||
elif text.startswith('/local'):
|
||
last_msg['content'] = text[len('/local'):].strip()
|
||
return 'FORCE_LOCAL'
|
||
|
||
# 2. 如果最后一条没指令，则从倒数第二条开始往回找历史指令
|
||
# 这样可以实现“记忆切换状态”
|
||
for i in range(len(messages) - 2, -1, -1):
|
||
msg = messages[i]
|
||
if msg.get('role') == 'user':
|
||
content = msg.get('content', '')
|
||
if isinstance(content, str):
|
||
if content.strip().startswith('/cloud'):
|
||
return 'FORCE_CLOUD'
|
||
if content.strip().startswith('/local'):
|
||
return 'FORCE_LOCAL'
|
||
|
||
return None
|
||
|
||
def build_agent(
|
||
model_server: str,
|
||
timeout_sec: int,
|
||
generate_cfg: Dict[str, Any],
|
||
model_id: Optional[str] = None,
|
||
system_prompt: str = DEFAULT_SYSTEM_PROMPT,
|
||
) -> Assistant:
|
||
|
||
# 获取默认模式
|
||
env_mode = os.getenv('MODEL_MODE', 'local').strip().lower()
|
||
|
||
# 决策逻辑：指令优先 > 环境变量
|
||
if model_id == 'FORCE_CLOUD':
|
||
is_cloud = True
|
||
elif model_id == 'FORCE_LOCAL':
|
||
is_cloud = False
|
||
else:
|
||
is_cloud = (env_mode == 'cloud')
|
||
|
||
if is_cloud:
|
||
target_model = os.getenv('CLOUD_MODEL_ID', 'gpt-4o').strip()
|
||
target_server = os.getenv('CLOUD_BASE_URL', '').strip()
|
||
target_key = os.getenv('CLOUD_API_KEY', '').strip()
|
||
display_name = f"Cloud-Agent ({target_model})"
|
||
else:
|
||
target_model = fetch_model_id(model_server, timeout_sec)
|
||
target_server = model_server
|
||
target_key = os.getenv('OPENAI_API_KEY', 'EMPTY')
|
||
display_name = f"Local-Agent ({target_model})"
|
||
|
||
# 日志输出（不带表情符号，防止 Windows CMD 崩溃）
|
||
print(f"\n[Route] Current Brain: {display_name}", flush=True)
|
||
|
||
llm_cfg = {
|
||
'model': target_model,
|
||
'model_server': target_server,
|
||
'api_key': target_key,
|
||
'model_type': 'qwenvl_oai',
|
||
'generate_cfg': generate_cfg,
|
||
}
|
||
|
||
# 组装功能列表
|
||
actual_function_list = list(DEFAULT_FUNCTION_LIST)
|
||
if os.getenv('ENABLE_FILE_WRITE', 'False').lower() == 'true':
|
||
actual_function_list.append('write_file')
|
||
|
||
# 组装系统提示词
|
||
persistent_memory = get_injected_memory()
|
||
now = datetime.datetime.now()
|
||
weekdays = ["一", "二", "三", "四", "五", "六", "日"]
|
||
dynamic_context = f"【系统实时状态】\n当前时间：{now.strftime('%Y年%m月%d日 %H:%M:%S')}，星期{weekdays[now.weekday()]}。\n"
|
||
actual_system_prompt = dynamic_context + persistent_memory + system_prompt
|
||
|
||
return Assistant(
|
||
name=display_name,
|
||
llm=llm_cfg,
|
||
function_list=actual_function_list,
|
||
system_message=actual_system_prompt,
|
||
)
|
||
|
||
# ==========================================
|
||
# 🔄 基础转换工具函数
|
||
# ==========================================
|
||
def _extract_image_uri(part: Dict[str, Any]) -> Optional[str]:
|
||
keys = ('image_url', 'image', 'url', 'input_image', 'image_uri')
|
||
for key in keys:
|
||
value = part.get(key)
|
||
if isinstance(value, str) and value.strip():
|
||
return value.strip()
|
||
if isinstance(value, dict):
|
||
nested = value.get('url') or value.get('image_url') or value.get('image')
|
||
if isinstance(nested, str) and nested.strip():
|
||
return nested.strip()
|
||
return None
|
||
|
||
def _build_compact_part_text(part: Dict[str, Any], part_type: Any) -> str:
|
||
part_keys = sorted(str(k) for k in part.keys())
|
||
payload = {'type': str(part_type or 'unknown'), 'keys': part_keys[:12]}
|
||
text = part.get('text')
|
||
if isinstance(text, str) and text.strip():
|
||
payload['text'] = text.strip()[:MAX_FALLBACK_PART_TEXT_CHARS]
|
||
return json.dumps(payload, ensure_ascii=False)
|
||
|
||
def extract_generate_cfg(payload: Dict[str, Any]) -> Dict[str, Any]:
|
||
cfg: Dict[str, Any] = {}
|
||
keys = ('temperature', 'top_p', 'top_k', 'presence_penalty', 'frequency_penalty')
|
||
for key in keys:
|
||
value = payload.get(key)
|
||
if value is not None:
|
||
cfg[key] = value
|
||
repeat_penalty = payload.get('repeat_penalty')
|
||
if repeat_penalty is not None:
|
||
cfg['repetition_penalty'] = repeat_penalty
|
||
extra_body = payload.get('extra_body', {})
|
||
chat_template_kwargs = extra_body.get('chat_template_kwargs', {})
|
||
chat_template_kwargs.setdefault('enable_thinking', True)
|
||
extra_body['chat_template_kwargs'] = chat_template_kwargs
|
||
extra_body.setdefault('reasoning_format', payload.get('reasoning_format', 'deepseek'))
|
||
extra_body.setdefault('reasoning_budget', -1)
|
||
cfg['extra_body'] = extra_body
|
||
max_tokens = payload.get('max_tokens')
|
||
if isinstance(max_tokens, int) and max_tokens > 0:
|
||
cfg['max_tokens'] = max_tokens
|
||
return cfg if cfg else {'temperature': 0.7, 'top_p': 0.9, 'max_tokens': 512}
|
||
|
||
def to_content_items(content: Any) -> Union[str, List[ContentItem]]:
|
||
if isinstance(content, str): return content
|
||
if not isinstance(content, list): return str(content)
|
||
items: List[ContentItem] = []
|
||
for part in content:
|
||
if not isinstance(part, dict):
|
||
items.append(ContentItem(text=str(part)))
|
||
continue
|
||
part_type = part.get('type')
|
||
if part_type in (None, 'text', 'input_text'):
|
||
text = part.get('text', '')
|
||
if text: items.append(ContentItem(text=str(text)))
|
||
continue
|
||
image_uri = _extract_image_uri(part)
|
||
if image_uri:
|
||
items.append(ContentItem(image=image_uri))
|
||
continue
|
||
items.append(ContentItem(text=_build_compact_part_text(part, part_type)))
|
||
return items if items else ''
|
||
|
||
def to_qwen_messages(openai_messages: Sequence[Dict[str, Any]]) -> List[Message]:
|
||
qwen_messages: List[Message] = []
|
||
for item in openai_messages:
|
||
role = str(item.get('role', '')).strip()
|
||
if role in {'system', 'user', 'assistant'}:
|
||
qwen_messages.append(Message(role=role, content=to_content_items(item.get('content', ''))))
|
||
return qwen_messages
|
||
|
||
def content_to_text(content: Any) -> str:
|
||
if isinstance(content, str): return content
|
||
if not isinstance(content, list): return str(content)
|
||
texts: List[str] = []
|
||
for item in content:
|
||
if isinstance(item, str): texts.append(item)
|
||
elif isinstance(item, dict) and item.get('text'): texts.append(str(item['text']))
|
||
elif hasattr(item, 'text'): texts.append(str(item.text))
|
||
return '\n'.join(texts).strip()
|
||
|
||
def extract_answer_and_reasoning(messages: Sequence[Message]) -> Dict[str, str]:
|
||
answer = ''
|
||
reasoning_parts = []
|
||
for message in messages:
|
||
if getattr(message, 'role', '') != 'assistant': continue
|
||
content_text = content_to_text(message.get('content', ''))
|
||
if content_text: answer = content_text
|
||
reasoning_text = content_to_text(message.get('reasoning_content', ''))
|
||
if reasoning_text: reasoning_parts.append(reasoning_text)
|
||
return {'answer': answer, 'reasoning': '\n'.join(reasoning_parts).strip()}
|
||
|
||
# ==========================================
|
||
# 📡 核心接口实现
|
||
# ==========================================
|
||
def run_chat_completion(payload: Dict[str, Any], model_server: str, timeout_sec: int) -> Dict[str, str]:
|
||
openai_messages = payload.get('messages', [])
|
||
|
||
# 拦截路由指令
|
||
route_signal = extract_routing_command(openai_messages)
|
||
requested_model = route_signal if route_signal else payload.get('model')
|
||
|
||
agent = build_agent(model_server, timeout_sec, extract_generate_cfg(payload), model_id=requested_model)
|
||
qwen_messages = to_qwen_messages(openai_messages)
|
||
|
||
final_batch = None
|
||
for batch in agent.run(qwen_messages):
|
||
final_batch = batch
|
||
|
||
if not final_batch: raise RuntimeError('未收到模型输出')
|
||
texts = extract_answer_and_reasoning(final_batch)
|
||
return {'model': agent.llm.model, 'answer': texts['answer'], 'reasoning': texts['reasoning']}
|
||
|
||
def build_sse_chunk(chat_id: str, created: int, model: str, delta: Dict[str, Any], finish_reason: Optional[str] = None, timings: Optional[Dict[str, Any]] = None) -> bytes:
|
||
chunk = {
|
||
'id': chat_id, 'object': 'chat.completion.chunk', 'created': created, 'model': model,
|
||
'choices': [{'index': 0, 'delta': delta, 'finish_reason': finish_reason}],
|
||
}
|
||
if timings: chunk['timings'] = timings
|
||
return f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n".encode('utf-8')
|
||
|
||
def text_delta(previous: str, current: str) -> str:
|
||
return current[len(previous):] if current.startswith(previous) else current
|
||
|
||
def count_text_tokens(model_server: str, timeout_sec: int, text: str) -> int:
|
||
try:
|
||
url = f"{model_server.rstrip('/')}/tokenize" if not model_server.endswith('/v1') else f"{model_server[:-3]}/tokenize"
|
||
response = requests.post(url, json={'content': text}, timeout=timeout_sec)
|
||
return len(response.json().get('tokens', []))
|
||
except: return 0
|
||
|
||
def build_live_timings(token_count: int, elapsed_sec: float) -> Dict[str, Any]:
|
||
safe_elapsed = max(elapsed_sec, 0.001)
|
||
return {'prompt_n': 0, 'prompt_ms': 0, 'predicted_n': token_count, 'predicted_ms': safe_elapsed * 1000.0, 'predicted_per_second': token_count / safe_elapsed, 'cache_n': 0}
|
||
|
||
def merge_generated_text(reasoning: str, answer: str) -> str:
|
||
return f"{reasoning}\n{answer}" if reasoning and answer else (reasoning or answer)
|
||
|
||
def stream_chat_completion(payload: Dict[str, Any], model_server: str, timeout_sec: int) -> Iterable[bytes]:
|
||
openai_messages = payload.get('messages', [])
|
||
|
||
# 拦截路由指令
|
||
route_signal = extract_routing_command(openai_messages)
|
||
requested_model = route_signal if route_signal else payload.get('model')
|
||
|
||
agent = build_agent(model_server, timeout_sec, extract_generate_cfg(payload), model_id=requested_model)
|
||
qwen_messages = to_qwen_messages(openai_messages)
|
||
|
||
now, chat_id = int(time.time()), f'chatcmpl-{uuid.uuid4().hex}'
|
||
current_model = agent.llm.model
|
||
yield build_sse_chunk(chat_id, now, current_model, {'role': 'assistant'})
|
||
|
||
previous_answer, previous_reasoning = '', ''
|
||
started_at = time.perf_counter()
|
||
last_timing_at, last_reported_tokens, last_counted_text = started_at, -1, ''
|
||
|
||
for batch in agent.run(qwen_messages):
|
||
texts = extract_answer_and_reasoning(batch)
|
||
answer, reasoning = texts['answer'], texts['reasoning']
|
||
|
||
reasoning_inc = text_delta(previous_reasoning, reasoning)
|
||
if reasoning_inc: yield build_sse_chunk(chat_id, now, current_model, {'reasoning_content': reasoning_inc})
|
||
|
||
answer_inc = text_delta(previous_answer, answer)
|
||
if answer_inc: yield build_sse_chunk(chat_id, now, current_model, {'content': answer_inc})
|
||
|
||
# Timing 逻辑
|
||
gen_text = merge_generated_text(reasoning, answer)
|
||
cur_t = time.perf_counter()
|
||
if gen_text and gen_text != last_counted_text and (cur_t - last_timing_at) >= TIMINGS_EMIT_INTERVAL_SEC:
|
||
t_count = count_text_tokens(model_server, timeout_sec, gen_text)
|
||
if t_count != last_reported_tokens:
|
||
yield build_sse_chunk(chat_id, now, current_model, {}, timings=build_live_timings(t_count, cur_t - started_at))
|
||
last_reported_tokens = t_count
|
||
last_counted_text, last_timing_at = gen_text, cur_t
|
||
|
||
previous_answer, previous_reasoning = answer, reasoning
|
||
|
||
yield build_sse_chunk(chat_id, now, current_model, {}, 'stop')
|
||
yield b'data: [DONE]\n\n'
|
||
|
||
# ==========================================
|
||
# 🎁 响应构建工具（补全缺失的函数）
|
||
# ==========================================
|
||
|
||
def build_non_stream_response(answer: str, model: str, reasoning: str = '') -> Dict[str, Any]:
|
||
"""构建 OpenAI 兼容的非流式响应 JSON"""
|
||
now = int(time.time())
|
||
message = {'role': 'assistant', 'content': answer}
|
||
if reasoning:
|
||
message['reasoning_content'] = reasoning
|
||
|
||
return {
|
||
'id': f'chatcmpl-{uuid.uuid4().hex}',
|
||
'object': 'chat.completion',
|
||
'created': now,
|
||
'model': model,
|
||
'choices': [{
|
||
'index': 0,
|
||
'message': message,
|
||
'finish_reason': 'stop',
|
||
}],
|
||
'usage': {'prompt_tokens': 0, 'completion_tokens': 0, 'total_tokens': 0},
|
||
}
|
||
|
||
def sse_lines(answer: str, model: str, reasoning: str = '') -> Iterable[bytes]:
|
||
"""将完整答案转换为 SSE 流格式（用于特殊兼容性场景）"""
|
||
now = int(time.time())
|
||
chat_id = f'chatcmpl-{uuid.uuid4().hex}'
|
||
|
||
# 1. 角色初始化块
|
||
yield build_sse_chunk(chat_id, now, model, {'role': 'assistant'})
|
||
|
||
# 2. 思维链块（如果有）
|
||
if reasoning:
|
||
yield build_sse_chunk(chat_id, now, model, {'reasoning_content': reasoning})
|
||
|
||
# 3. 正文块
|
||
yield build_sse_chunk(chat_id, now, model, {'content': answer})
|
||
|
||
# 4. 结束块
|
||
yield build_sse_chunk(chat_id, now, model, {}, 'stop')
|
||
yield b'data: [DONE]\n\n' |