313 lines
10 KiB
Python
313 lines
10 KiB
Python
|
"""
|
||
|
|
成本追踪系统 - 监控Token消耗和API调用成本
|
||
|
|
参考Claude Code的cost-tracker.ts实现
|
||
|
|
"""
|
||
|
|
|
||
|
|
import json
|
||
|
|
import time
|
||
|
|
from dataclasses import dataclass, field
|
||
|
|
from datetime import datetime
|
||
|
|
from pathlib import Path
|
||
|
|
from typing import Dict, List, Optional
|
||
|
|
from threading import Lock
|
||
|
|
|
||
|
|
from gfa_core.project_paths import PROJECT_ROOT
|
||
|
|
|
||
|
|
|
||
|
|
@dataclass
|
||
|
|
class CostEntry:
|
||
|
|
"""单次调用成本记录"""
|
||
|
|
timestamp: str
|
||
|
|
agent_name: str
|
||
|
|
provider: str
|
||
|
|
model: str
|
||
|
|
input_tokens: int
|
||
|
|
output_tokens: int
|
||
|
|
total_tokens: int
|
||
|
|
estimated_cost: float # 美元
|
||
|
|
duration_ms: int
|
||
|
|
query_preview: str # 查询前50字符
|
||
|
|
|
||
|
|
|
||
|
|
@dataclass
|
||
|
|
class AgentCostStats:
|
||
|
|
"""Agent成本统计"""
|
||
|
|
agent_name: str
|
||
|
|
call_count: int = 0
|
||
|
|
total_input_tokens: int = 0
|
||
|
|
total_output_tokens: int = 0
|
||
|
|
total_tokens: int = 0
|
||
|
|
total_cost: float = 0.0
|
||
|
|
avg_cost_per_call: float = 0.0
|
||
|
|
last_call: Optional[str] = None
|
||
|
|
|
||
|
|
|
||
|
|
class CostTracker:
|
||
|
|
"""
|
||
|
|
Token成本追踪器
|
||
|
|
|
||
|
|
功能:
|
||
|
|
1. 记录每次LLM调用的token消耗
|
||
|
|
2. 按Agent统计成本
|
||
|
|
3. 按时间段统计成本
|
||
|
|
4. 生成成本报告
|
||
|
|
5. 提供优化建议
|
||
|
|
"""
|
||
|
|
|
||
|
|
# Token价格 (美元/1K tokens) - 阿里云百炼
|
||
|
|
TOKEN_PRICES = {
|
||
|
|
"glm-4-flash": {"input": 0.0001, "output": 0.0001},
|
||
|
|
"moonshot-v1-8k": {"input": 0.00015, "output": 0.00015},
|
||
|
|
"abab6.5s-chat": {"input": 0.0001, "output": 0.0001},
|
||
|
|
"deepseek-chat": {"input": 0.0001, "output": 0.0002},
|
||
|
|
"qwen3:0.6b": {"input": 0.0, "output": 0.0}, # 本地免费
|
||
|
|
"deterministic-estimate": {"input": 0.0, "output": 0.0}, # 本地规则/无LLM交互估算
|
||
|
|
}
|
||
|
|
|
||
|
|
def __init__(self, storage_dir: Optional[Path] = None):
|
||
|
|
self.storage_dir = storage_dir or PROJECT_ROOT / "cost_logs"
|
||
|
|
self.storage_dir.mkdir(parents=True, exist_ok=True)
|
||
|
|
|
||
|
|
self._entries: List[CostEntry] = []
|
||
|
|
self._lock = Lock()
|
||
|
|
self._daily_file = self.storage_dir / f"cost_{datetime.now().strftime('%Y-%m-%d')}.jsonl"
|
||
|
|
|
||
|
|
# 加载今日已有记录
|
||
|
|
self._load_today_entries()
|
||
|
|
|
||
|
|
def _load_today_entries(self):
|
||
|
|
"""加载今日成本记录"""
|
||
|
|
if self._daily_file.exists():
|
||
|
|
with open(self._daily_file, 'r', encoding='utf-8') as f:
|
||
|
|
for line in f:
|
||
|
|
line = line.strip()
|
||
|
|
if line:
|
||
|
|
try:
|
||
|
|
data = json.loads(line)
|
||
|
|
self._entries.append(CostEntry(**data))
|
||
|
|
except:
|
||
|
|
pass
|
||
|
|
|
||
|
|
def _estimate_cost(self, provider: str, model: str, input_tokens: int, output_tokens: int) -> float:
|
||
|
|
"""估算成本"""
|
||
|
|
prices = self.TOKEN_PRICES.get(model, {"input": 0.0001, "output": 0.0001})
|
||
|
|
|
||
|
|
input_cost = (input_tokens / 1000) * prices["input"]
|
||
|
|
output_cost = (output_tokens / 1000) * prices["output"]
|
||
|
|
|
||
|
|
return round(input_cost + output_cost, 6)
|
||
|
|
|
||
|
|
def track(
|
||
|
|
self,
|
||
|
|
agent_name: str,
|
||
|
|
provider: str,
|
||
|
|
model: str,
|
||
|
|
input_tokens: int,
|
||
|
|
output_tokens: int,
|
||
|
|
duration_ms: int,
|
||
|
|
query: str = "",
|
||
|
|
) -> CostEntry:
|
||
|
|
"""
|
||
|
|
记录一次LLM调用成本
|
||
|
|
|
||
|
|
Args:
|
||
|
|
agent_name: Agent名称
|
||
|
|
provider: 提供商(glm/kimi/deepseek等)
|
||
|
|
model: 模型名称
|
||
|
|
input_tokens: 输入token数
|
||
|
|
output_tokens: 输出token数
|
||
|
|
duration_ms: 调用耗时(毫秒)
|
||
|
|
query: 查询内容(用于记录)
|
||
|
|
"""
|
||
|
|
entry = CostEntry(
|
||
|
|
timestamp=datetime.now().isoformat(),
|
||
|
|
agent_name=agent_name,
|
||
|
|
provider=provider,
|
||
|
|
model=model,
|
||
|
|
input_tokens=input_tokens,
|
||
|
|
output_tokens=output_tokens,
|
||
|
|
total_tokens=input_tokens + output_tokens,
|
||
|
|
estimated_cost=self._estimate_cost(provider, model, input_tokens, output_tokens),
|
||
|
|
duration_ms=duration_ms,
|
||
|
|
query_preview=query[:50] if query else "",
|
||
|
|
)
|
||
|
|
|
||
|
|
with self._lock:
|
||
|
|
self._entries.append(entry)
|
||
|
|
|
||
|
|
# 写入文件
|
||
|
|
with open(self._daily_file, 'a', encoding='utf-8') as f:
|
||
|
|
f.write(json.dumps({
|
||
|
|
"timestamp": entry.timestamp,
|
||
|
|
"agent_name": entry.agent_name,
|
||
|
|
"provider": entry.provider,
|
||
|
|
"model": entry.model,
|
||
|
|
"input_tokens": entry.input_tokens,
|
||
|
|
"output_tokens": entry.output_tokens,
|
||
|
|
"total_tokens": entry.total_tokens,
|
||
|
|
"estimated_cost": entry.estimated_cost,
|
||
|
|
"duration_ms": entry.duration_ms,
|
||
|
|
"query_preview": entry.query_preview,
|
||
|
|
}, ensure_ascii=False) + '\n')
|
||
|
|
|
||
|
|
return entry
|
||
|
|
|
||
|
|
def get_agent_stats(self, agent_name: Optional[str] = None) -> Dict[str, AgentCostStats]:
|
||
|
|
"""获取Agent成本统计"""
|
||
|
|
stats: Dict[str, AgentCostStats] = {}
|
||
|
|
|
||
|
|
for entry in self._entries:
|
||
|
|
if agent_name and entry.agent_name != agent_name:
|
||
|
|
continue
|
||
|
|
|
||
|
|
if entry.agent_name not in stats:
|
||
|
|
stats[entry.agent_name] = AgentCostStats(agent_name=entry.agent_name)
|
||
|
|
|
||
|
|
s = stats[entry.agent_name]
|
||
|
|
s.call_count += 1
|
||
|
|
s.total_input_tokens += entry.input_tokens
|
||
|
|
s.total_output_tokens += entry.output_tokens
|
||
|
|
s.total_tokens += entry.total_tokens
|
||
|
|
s.total_cost += entry.estimated_cost
|
||
|
|
s.last_call = entry.timestamp
|
||
|
|
|
||
|
|
# 计算平均值
|
||
|
|
for s in stats.values():
|
||
|
|
if s.call_count > 0:
|
||
|
|
s.avg_cost_per_call = round(s.total_cost / s.call_count, 6)
|
||
|
|
|
||
|
|
return stats
|
||
|
|
|
||
|
|
def get_summary(self) -> dict:
|
||
|
|
"""获取成本摘要"""
|
||
|
|
if not self._entries:
|
||
|
|
return {
|
||
|
|
"total_calls": 0,
|
||
|
|
"total_tokens": 0,
|
||
|
|
"total_cost": 0.0,
|
||
|
|
"agents": {},
|
||
|
|
}
|
||
|
|
|
||
|
|
total_calls = len(self._entries)
|
||
|
|
total_tokens = sum(e.total_tokens for e in self._entries)
|
||
|
|
total_cost = sum(e.estimated_cost for e in self._entries)
|
||
|
|
|
||
|
|
agent_stats = self.get_agent_stats()
|
||
|
|
|
||
|
|
return {
|
||
|
|
"total_calls": total_calls,
|
||
|
|
"total_tokens": total_tokens,
|
||
|
|
"total_cost": round(total_cost, 6),
|
||
|
|
"avg_cost_per_call": round(total_cost / total_calls, 6) if total_calls > 0 else 0,
|
||
|
|
"avg_tokens_per_call": total_tokens // total_calls if total_calls > 0 else 0,
|
||
|
|
"agents": {
|
||
|
|
name: {
|
||
|
|
"calls": s.call_count,
|
||
|
|
"tokens": s.total_tokens,
|
||
|
|
"cost": round(s.total_cost, 6),
|
||
|
|
"avg_cost": s.avg_cost_per_call,
|
||
|
|
}
|
||
|
|
for name, s in agent_stats.items()
|
||
|
|
},
|
||
|
|
}
|
||
|
|
|
||
|
|
def generate_report(self) -> str:
|
||
|
|
"""生成成本报告"""
|
||
|
|
summary = self.get_summary()
|
||
|
|
|
||
|
|
lines = [
|
||
|
|
"=" * 60,
|
||
|
|
"BBAV20311 成本追踪报告",
|
||
|
|
"=" * 60,
|
||
|
|
f"生成时间: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}",
|
||
|
|
f"数据日期: {datetime.now().strftime('%Y-%m-%d')}",
|
||
|
|
"",
|
||
|
|
"【总体统计】",
|
||
|
|
f" 总调用次数: {summary['total_calls']}",
|
||
|
|
f" 总Token数: {summary['total_tokens']:,}",
|
||
|
|
f" 总成本: ${summary['total_cost']:.6f}",
|
||
|
|
f" 平均每次成本: ${summary['avg_cost_per_call']:.6f}",
|
||
|
|
f" 平均每次Token: {summary['avg_tokens_per_call']:,}",
|
||
|
|
"",
|
||
|
|
"【Agent统计】",
|
||
|
|
]
|
||
|
|
|
||
|
|
for agent_name, stats in summary['agents'].items():
|
||
|
|
lines.extend([
|
||
|
|
f" {agent_name}:",
|
||
|
|
f" 调用: {stats['calls']}",
|
||
|
|
f" Token: {stats['tokens']:,}",
|
||
|
|
f" 成本: ${stats['cost']:.6f}",
|
||
|
|
f" 平均: ${stats['avg_cost']:.6f}/call",
|
||
|
|
])
|
||
|
|
|
||
|
|
lines.extend([
|
||
|
|
"",
|
||
|
|
"=" * 60,
|
||
|
|
])
|
||
|
|
|
||
|
|
return '\n'.join(lines)
|
||
|
|
|
||
|
|
def get_optimization_suggestions(self) -> List[str]:
|
||
|
|
"""获取优化建议"""
|
||
|
|
suggestions = []
|
||
|
|
stats = self.get_agent_stats()
|
||
|
|
|
||
|
|
# 找出高消耗的Agent
|
||
|
|
high_cost_agents = [
|
||
|
|
(name, s) for name, s in stats.items()
|
||
|
|
if s.total_cost > 0.01 # 超过1美分
|
||
|
|
]
|
||
|
|
|
||
|
|
if high_cost_agents:
|
||
|
|
suggestions.append(f"高消耗Agent: {', '.join(name for name, _ in high_cost_agents)}")
|
||
|
|
suggestions.append("建议: 考虑缓存常用查询结果")
|
||
|
|
|
||
|
|
# 检查token效率
|
||
|
|
for name, s in stats.items():
|
||
|
|
if s.avg_cost_per_call > 0.005: # 单次超过0.5美分
|
||
|
|
suggestions.append(f"{name}单次成本较高(${s.avg_cost_per_call:.6f})，建议优化提示词长度")
|
||
|
|
|
||
|
|
if not suggestions:
|
||
|
|
suggestions.append("当前成本在合理范围内，无需特别优化")
|
||
|
|
|
||
|
|
return suggestions
|
||
|
|
|
||
|
|
|
||
|
|
# 全局成本追踪器实例
|
||
|
|
_cost_tracker: Optional[CostTracker] = None
|
||
|
|
|
||
|
|
|
||
|
|
def get_cost_tracker() -> CostTracker:
|
||
|
|
"""获取全局成本追踪器"""
|
||
|
|
global _cost_tracker
|
||
|
|
if _cost_tracker is None:
|
||
|
|
_cost_tracker = CostTracker()
|
||
|
|
return _cost_tracker
|
||
|
|
|
||
|
|
|
||
|
|
def track_cost(
|
||
|
|
agent_name: str,
|
||
|
|
provider: str,
|
||
|
|
model: str,
|
||
|
|
input_tokens: int,
|
||
|
|
output_tokens: int,
|
||
|
|
duration_ms: int,
|
||
|
|
query: str = "",
|
||
|
|
):
|
||
|
|
"""便捷函数：记录成本"""
|
||
|
|
tracker = get_cost_tracker()
|
||
|
|
return tracker.track(agent_name, provider, model, input_tokens, output_tokens, duration_ms, query)
|
||
|
|
|
||
|
|
|
||
|
|
def print_cost_report():
|
||
|
|
"""便捷函数：打印成本报告"""
|
||
|
|
tracker = get_cost_tracker()
|
||
|
|
print(tracker.generate_report())
|
||
|
|
|
||
|
|
suggestions = tracker.get_optimization_suggestions()
|
||
|
|
if suggestions:
|
||
|
|
print("\n【优化建议】")
|
||
|
|
for s in suggestions:
|
||
|
|
print(f" • {s}")
|