refactor(api): 重构 LLM 统计数据收集逻辑
将原有的方式重构为直接从数据库中查询和聚合 LLM 使用记录。这提高了数据的持久性和准确性，并消除了对后台统计任务的依赖。 主要变更： - 移除旧的 `StatisticOutputTask` 和基于 Redis 的统计变量。 - 新增 `_collect_stats_in_period` 函数，用于在指定时间段内从 `LLMUsage` 表中动态收集和计算统计数据。 - 统计时，将数据库中存储的实际模型标识符（model_identifier）映射回配置文件中的模型名称，确保成本计算和数据显示的一致性。 - 扩展了 `period_type` 查询参数，增加了如 "last_hour", "last_24_hours", "last_7_days" 等多个预设时间范围，提升了 API 的易用性。
This commit is contained in:
minecraft1024a
@@ -1,48 +1,152 @@
|
||||
from collections import defaultdict
|
||||
from datetime import datetime, timedelta
|
||||
from typing import Any, Literal
|
||||
|
||||
from fastapi import APIRouter, HTTPException, Query
|
||||
|
||||
from src.chat.utils.statistic import (
|
||||
COST_BY_MODEL,
|
||||
COST_BY_MODULE,
|
||||
COST_BY_TYPE,
|
||||
COST_BY_USER,
|
||||
IN_TOK_BY_MODEL,
|
||||
IN_TOK_BY_MODULE,
|
||||
IN_TOK_BY_TYPE,
|
||||
IN_TOK_BY_USER,
|
||||
OUT_TOK_BY_MODEL,
|
||||
OUT_TOK_BY_MODULE,
|
||||
OUT_TOK_BY_TYPE,
|
||||
OUT_TOK_BY_USER,
|
||||
REQ_CNT_BY_MODEL,
|
||||
REQ_CNT_BY_MODULE,
|
||||
REQ_CNT_BY_TYPE,
|
||||
REQ_CNT_BY_USER,
|
||||
TOTAL_COST,
|
||||
TOTAL_REQ_CNT,
|
||||
TOTAL_TOK_BY_MODEL,
|
||||
TOTAL_TOK_BY_MODULE,
|
||||
TOTAL_TOK_BY_TYPE,
|
||||
TOTAL_TOK_BY_USER,
|
||||
StatisticOutputTask,
|
||||
)
|
||||
from src.common.database.sqlalchemy_database_api import db_get
|
||||
from src.common.database.sqlalchemy_models import LLMUsage
|
||||
from src.common.logger import get_logger
|
||||
from src.config.config import model_config
|
||||
|
||||
logger = get_logger("LLM统计API")
|
||||
|
||||
router = APIRouter()
|
||||
|
||||
# 定义统计数据的键,以减少魔法字符串
|
||||
TOTAL_REQ_CNT = "total_requests"
|
||||
TOTAL_COST = "total_cost"
|
||||
REQ_CNT_BY_TYPE = "requests_by_type"
|
||||
REQ_CNT_BY_USER = "requests_by_user"
|
||||
REQ_CNT_BY_MODEL = "requests_by_model"
|
||||
REQ_CNT_BY_MODULE = "requests_by_module"
|
||||
IN_TOK_BY_TYPE = "in_tokens_by_type"
|
||||
IN_TOK_BY_USER = "in_tokens_by_user"
|
||||
IN_TOK_BY_MODEL = "in_tokens_by_model"
|
||||
IN_TOK_BY_MODULE = "in_tokens_by_module"
|
||||
OUT_TOK_BY_TYPE = "out_tokens_by_type"
|
||||
OUT_TOK_BY_USER = "out_tokens_by_user"
|
||||
OUT_TOK_BY_MODEL = "out_tokens_by_model"
|
||||
OUT_TOK_BY_MODULE = "out_tokens_by_module"
|
||||
TOTAL_TOK_BY_TYPE = "tokens_by_type"
|
||||
TOTAL_TOK_BY_USER = "tokens_by_user"
|
||||
TOTAL_TOK_BY_MODEL = "tokens_by_model"
|
||||
TOTAL_TOK_BY_MODULE = "tokens_by_module"
|
||||
COST_BY_TYPE = "costs_by_type"
|
||||
COST_BY_USER = "costs_by_user"
|
||||
COST_BY_MODEL = "costs_by_model"
|
||||
COST_BY_MODULE = "costs_by_module"
|
||||
|
||||
|
||||
async def _collect_stats_in_period(start_time: datetime, end_time: datetime) -> dict[str, Any]:
|
||||
"""在指定时间段内收集LLM使用统计信息"""
|
||||
records = await db_get(
|
||||
model_class=LLMUsage,
|
||||
filters={"timestamp": {"$gte": start_time, "$lt": end_time}},
|
||||
)
|
||||
if not records:
|
||||
return {}
|
||||
|
||||
# 创建一个从 model_identifier 到 name 的映射
|
||||
model_identifier_to_name_map = {model.model_identifier: model.name for model in model_config.models}
|
||||
|
||||
stats: dict[str, Any] = {
|
||||
TOTAL_REQ_CNT: 0,
|
||||
TOTAL_COST: 0.0,
|
||||
REQ_CNT_BY_TYPE: defaultdict(int),
|
||||
REQ_CNT_BY_USER: defaultdict(int),
|
||||
REQ_CNT_BY_MODEL: defaultdict(int),
|
||||
REQ_CNT_BY_MODULE: defaultdict(int),
|
||||
IN_TOK_BY_TYPE: defaultdict(int),
|
||||
IN_TOK_BY_USER: defaultdict(int),
|
||||
IN_TOK_BY_MODEL: defaultdict(int),
|
||||
IN_TOK_BY_MODULE: defaultdict(int),
|
||||
OUT_TOK_BY_TYPE: defaultdict(int),
|
||||
OUT_TOK_BY_USER: defaultdict(int),
|
||||
OUT_TOK_BY_MODEL: defaultdict(int),
|
||||
OUT_TOK_BY_MODULE: defaultdict(int),
|
||||
TOTAL_TOK_BY_TYPE: defaultdict(int),
|
||||
TOTAL_TOK_BY_USER: defaultdict(int),
|
||||
TOTAL_TOK_BY_MODEL: defaultdict(int),
|
||||
TOTAL_TOK_BY_MODULE: defaultdict(int),
|
||||
COST_BY_TYPE: defaultdict(float),
|
||||
COST_BY_USER: defaultdict(float),
|
||||
COST_BY_MODEL: defaultdict(float),
|
||||
COST_BY_MODULE: defaultdict(float),
|
||||
}
|
||||
|
||||
for record in records:
|
||||
if not isinstance(record, dict):
|
||||
continue
|
||||
|
||||
stats[TOTAL_REQ_CNT] += 1
|
||||
|
||||
request_type = record.get("request_type") or "unknown"
|
||||
user_id = record.get("user_id") or "unknown"
|
||||
# 从数据库获取的是真实模型名 (model_identifier)
|
||||
real_model_name = record.get("model_name") or "unknown"
|
||||
module_name = request_type.split(".")[0] if "." in request_type else request_type
|
||||
|
||||
# 尝试通过真实模型名找到配置文件中的模型名
|
||||
config_model_name = model_identifier_to_name_map.get(real_model_name, real_model_name)
|
||||
|
||||
prompt_tokens = record.get("prompt_tokens") or 0
|
||||
completion_tokens = record.get("completion_tokens") or 0
|
||||
total_tokens = prompt_tokens + completion_tokens
|
||||
|
||||
cost = 0.0
|
||||
try:
|
||||
# 使用配置文件中的模型名来获取模型信息
|
||||
model_info = model_config.get_model_info(config_model_name)
|
||||
if model_info:
|
||||
input_cost = (prompt_tokens / 1000000) * model_info.price_in
|
||||
output_cost = (completion_tokens / 1000000) * model_info.price_out
|
||||
cost = round(input_cost + output_cost, 6)
|
||||
except KeyError as e:
|
||||
logger.info(str(e))
|
||||
logger.warning(f"模型 '{config_model_name}' (真实名称: '{real_model_name}') 在配置中未找到，成本计算将使用默认值 0.0")
|
||||
|
||||
stats[TOTAL_COST] += cost
|
||||
|
||||
# 按类型统计
|
||||
stats[REQ_CNT_BY_TYPE][request_type] += 1
|
||||
stats[IN_TOK_BY_TYPE][request_type] += prompt_tokens
|
||||
stats[OUT_TOK_BY_TYPE][request_type] += completion_tokens
|
||||
stats[TOTAL_TOK_BY_TYPE][request_type] += total_tokens
|
||||
stats[COST_BY_TYPE][request_type] += cost
|
||||
|
||||
# 按用户统计
|
||||
stats[REQ_CNT_BY_USER][user_id] += 1
|
||||
stats[IN_TOK_BY_USER][user_id] += prompt_tokens
|
||||
stats[OUT_TOK_BY_USER][user_id] += completion_tokens
|
||||
stats[TOTAL_TOK_BY_USER][user_id] += total_tokens
|
||||
stats[COST_BY_USER][user_id] += cost
|
||||
|
||||
# 按模型统计 (使用配置文件中的名称)
|
||||
stats[REQ_CNT_BY_MODEL][config_model_name] += 1
|
||||
stats[IN_TOK_BY_MODEL][config_model_name] += prompt_tokens
|
||||
stats[OUT_TOK_BY_MODEL][config_model_name] += completion_tokens
|
||||
stats[TOTAL_TOK_BY_MODEL][config_model_name] += total_tokens
|
||||
stats[COST_BY_MODEL][config_model_name] += cost
|
||||
|
||||
# 按模块统计
|
||||
stats[REQ_CNT_BY_MODULE][module_name] += 1
|
||||
stats[IN_TOK_BY_MODULE][module_name] += prompt_tokens
|
||||
stats[OUT_TOK_BY_MODULE][module_name] += completion_tokens
|
||||
stats[TOTAL_TOK_BY_MODULE][module_name] += total_tokens
|
||||
stats[COST_BY_MODULE][module_name] += cost
|
||||
|
||||
return stats
|
||||
|
||||
|
||||
@router.get("/llm/stats")
|
||||
async def get_llm_stats(
|
||||
period_type: Literal["fixed", "daily", "custom"] = Query(
|
||||
"daily", description="查询的时间段类型: 'fixed' (固定), 'daily' (按天), 'custom' (自定义)"
|
||||
),
|
||||
days: int = Query(1, ge=1, description="当 period_type 为 'daily' 时，指定查询过去多少天的数据"),
|
||||
start_time_str: str = Query(None, description="当 period_type 为 'custom' 时，指定查询的开始时间 (ISO 8601)"),
|
||||
end_time_str: str = Query(None, description="当 period_type 为 'custom' 时，指定查询的结束时间 (ISO 8601)"),
|
||||
period_type: Literal[
|
||||
"daily", "custom", "last_hour", "last_24_hours", "last_7_days", "last_30_days"
|
||||
] = Query("daily", description="查询的时间段类型"),
|
||||
days: int = Query(1, ge=1, description="当 period_type 为 'daily' 时,指定查询过去多少天的数据"),
|
||||
start_time_str: str = Query(None, description="当 period_type 为 'custom' 时,指定查询的开始时间 (ISO 8601)"),
|
||||
end_time_str: str = Query(None, description="当 period_type 为 'custom' 时,指定查询的结束时间 (ISO 8601)"),
|
||||
group_by: Literal["model", "module", "user", "type"] = Query("model", description="按指定维度对结果进行分组"),
|
||||
):
|
||||
"""
|
||||
@@ -50,10 +154,19 @@ async def get_llm_stats(
|
||||
"""
|
||||
try:
|
||||
now = datetime.now()
|
||||
start_time, end_time = None, now
|
||||
end_time = now
|
||||
start_time = None
|
||||
|
||||
if period_type == "daily":
|
||||
start_time = now - timedelta(days=days)
|
||||
elif period_type == "last_hour":
|
||||
start_time = now - timedelta(hours=1)
|
||||
elif period_type == "last_24_hours":
|
||||
start_time = now - timedelta(days=1)
|
||||
elif period_type == "last_7_days":
|
||||
start_time = now - timedelta(days=7)
|
||||
elif period_type == "last_30_days":
|
||||
start_time = now - timedelta(days=30)
|
||||
elif period_type == "custom":
|
||||
if not start_time_str or not end_time_str:
|
||||
raise HTTPException(status_code=400, detail="自定义时间段必须提供开始和结束时间")
|
||||
@@ -61,22 +174,16 @@ async def get_llm_stats(
|
||||
start_time = datetime.fromisoformat(start_time_str)
|
||||
end_time = datetime.fromisoformat(end_time_str)
|
||||
except ValueError:
|
||||
raise HTTPException(status_code=400, detail="无效的日期时间格式，请使用ISO 8601格式")
|
||||
elif period_type == "fixed":
|
||||
# 预设的固定时间段，这里以最近一小时为例
|
||||
start_time = now - timedelta(hours=1)
|
||||
raise HTTPException(status_code=400, detail="无效的日期时间格式,请使用ISO 8601格式")
|
||||
|
||||
if start_time is None:
|
||||
raise HTTPException(status_code=400, detail="无法确定查询的起始时间")
|
||||
|
||||
# 调用统计函数
|
||||
stats_data = await StatisticOutputTask._collect_model_request_for_period([("custom", start_time)])
|
||||
period_stats = stats_data.get("custom", {})
|
||||
period_stats = await _collect_stats_in_period(start_time, end_time)
|
||||
|
||||
if not period_stats:
|
||||
return {"period": {"start": start_time.isoformat(), "end": end_time.isoformat()}, "data": {}}
|
||||
|
||||
# 根据 group_by 参数选择对应的数据
|
||||
key_mapping = {
|
||||
"model": (REQ_CNT_BY_MODEL, COST_BY_MODEL, IN_TOK_BY_MODEL, OUT_TOK_BY_MODEL, TOTAL_TOK_BY_MODEL),
|
||||
"module": (
|
||||
|
||||
Reference in New Issue
Block a user