guke
5c6840dd71
- comparison_record 加 llm_cost_yuan(元/float)+ llm_price_snapshot(JSON)两列 - _backfill_llm_calls 回填时按 app_config 当时单价逐模型算成本、冻结成本+快照到记录 - app_config 新增 llm_token_price 配置(per_model + default 兜底,运营在系统配置页可改) - services/llm_cost.py:compute_llm_cost 纯函数(按 model 分桶、error/无 usage 跳过、 脏价格当 unpriced 不抛异常以免连累 token 回填)+ get_llm_prices reader - admin schema 暴露成本:列表项带 llm_cost_yuan,详情另带价格快照 - tests/test_llm_cost.py(10 测试);scripts/seed_mock_llm_cost.py(mock seeder) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: guke <guke@autohome.com.cn> Reviewed-on: #133
54 lines
2.7 KiB
Python
54 lines
2.7 KiB
Python
"""LLM 调用成本计算(纯逻辑,无 DB):按 model 分桶累加 token × 单价,返回总成本(元)+ 价格快照。
|
|
|
|
用量取自 comparison_record.llm_calls[].usage(pricebot 已归一为 prompt/completion_tokens);
|
|
error / 无 usage 的调用跳过。price_cfg = {per_model:{model:{input_per_1m,output_per_1m}}, default:{...}}。
|
|
成本单位「元」——单次亚分级,用 float(不用 *_cents);snapshot 只含本次用到的模型的价(审计用,
|
|
不存整张价表)。用到但没配价(既无 per_model 又无 default)的模型 → 快照标 unpriced,成本按 0 计。
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
_PRICE_KEY = "llm_token_price"
|
|
|
|
|
|
def get_llm_prices(db) -> dict:
|
|
"""读 LLM 单价配置(app_config;表内无则回退 CONFIG_DEFS 默认)。返回 compute_llm_cost 的 price_cfg。"""
|
|
from app.repositories import app_config # 延迟 import:compute_llm_cost 纯逻辑不牵连 DB 层
|
|
return app_config.get_value(db, _PRICE_KEY)
|
|
|
|
|
|
def compute_llm_cost(calls: list[dict], price_cfg: dict) -> tuple[float | None, dict | None]:
|
|
"""遍历 calls 按 model 分桶,cost = Σ(入/1e6*入价 + 出/1e6*出价);无有效调用 → (None, None)。"""
|
|
if not calls:
|
|
return None, None
|
|
per_model = price_cfg.get("per_model") or {}
|
|
default = price_cfg.get("default")
|
|
buckets: dict[str, list[int]] = {} # model -> [Σprompt_tokens, Σcompletion_tokens]
|
|
for c in calls:
|
|
if c.get("error"):
|
|
continue
|
|
usage = c.get("usage") or {}
|
|
model = c.get("model") or "unknown"
|
|
b = buckets.setdefault(model, [0, 0])
|
|
b[0] += usage.get("prompt_tokens") or 0
|
|
b[1] += usage.get("completion_tokens") or 0
|
|
if not buckets: # 全是 error / 无 usage
|
|
return None, None
|
|
total = 0.0
|
|
prices: dict[str, dict] = {}
|
|
for model, (tin, tout) in buckets.items():
|
|
price = per_model.get(model, default)
|
|
in_p = price.get("input_per_1m") if isinstance(price, dict) else None
|
|
out_p = price.get("output_per_1m") if isinstance(price, dict) else None
|
|
# 没配价 / 无 default / 单价残缺或非法(配置页手改 JSON 可能存出脏数据)→ 标记待补价、
|
|
# 不计入成本;绝不抛异常,以免连累同一回填里的 token/llm_calls 落库。
|
|
if not isinstance(in_p, (int, float)) or not isinstance(out_p, (int, float)):
|
|
prices[model] = {"input_per_1m": in_p, "output_per_1m": out_p, "unpriced": True}
|
|
continue
|
|
total += tin / 1e6 * in_p + tout / 1e6 * out_p
|
|
prices[model] = {
|
|
"input_per_1m": in_p,
|
|
"output_per_1m": out_p,
|
|
"_source": "per_model" if model in per_model else "default",
|
|
}
|
|
return round(total, 6), {"mode": "per_model", "prices": prices}
|