yunchan/vignette
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

관리자 사용량 근거를 정교화

Browse source
This commit is contained in:
Yun Chan 2026-08-29 23:59:17 +09:00
parent ccdcfcd2f5
commit 707dba4f8f
10 changed files with 1136 additions and 209 deletions
Download patch file Download diff file Expand all files Collapse all files

78
apps/api/app/services/llm_pricing.py
View file

@ -9,10 +9,14 @@
from __future__ import annotations
from dataclasses import dataclass
from datetime import date, datetime, timezone
MILLION = 1_000_000
RATE_CARD_VERSION = "2026-07-31"
GOOGLE_36_RELEASE_START = date(2026, 7, 21)
GOOGLE_FLASH_INTRO_START = date(2026, 8, 13)
GOOGLE_FLASH_INTRO_END = date(2027, 1, 1)
GOOGLE_PRICING_URL = "https://ai.google.dev/gemini-api/docs/pricing"
OPENAI_CODEX_RATE_URL = "https://help.openai.com/en/articles/20001106-codex-rate-card"
@ -40,11 +44,65 @@ class CostEstimate:
source_url: str
def _pricing_date(value: date | datetime | int | float | str) -> date:
if isinstance(value, datetime):
if value.tzinfo is not None:
value = value.astimezone(timezone.utc)
return value.date()
if isinstance(value, date):
return value
if isinstance(value, (int, float)):
return datetime.fromtimestamp(value, timezone.utc).date()
return date.fromisoformat(value[:10])
def _google_flash_rate(*, model_key: str, priced_on: date) -> ModelRate | None:
if model_key == "gemini-3.6-flash" and priced_on < GOOGLE_36_RELEASE_START:
return None
if model_key == "gemini-3.7-flash" and priced_on < GOOGLE_FLASH_INTRO_START:
return None
if model_key not in {"gemini-3.6-flash", "gemini-3.7-flash"}:
return None
model_name = "Gemini 3.7 Flash" if model_key == "gemini-3.7-flash" else "Gemini 3.6 Flash"
if GOOGLE_FLASH_INTRO_START <= priced_on < GOOGLE_FLASH_INTRO_END:
return ModelRate(
0.75,
3.75,
0.075,
f"google-{model_key}-standard-intro@2026-08-13",
(
f"Google {model_name} 프로모션 표준 단가(2026-12-31까지)"
" · 입력 $0.75/M · 캐시 $0.075/M · 출력 $3.75/M"
),
GOOGLE_PRICING_URL,
)
effective_from = "2027-01-01" if priced_on >= GOOGLE_FLASH_INTRO_END else "2026-07-21"
period_label = (
"2027-01-01부터"
if priced_on >= GOOGLE_FLASH_INTRO_END
else "2026-07-21~2026-08-12"
)
return ModelRate(
1.50,
7.50,
0.15,
f"google-{model_key}-standard@{effective_from}",
(
f"Google {model_name} 표준 단가({period_label})"
" · 입력 $1.50/M · 캐시 $0.15/M · 출력 $7.50/M"
),
GOOGLE_PRICING_URL,
)
def _rate(
*,
provider: str,
model: str,
tokens_in: int,
priced_on: date,
) -> ModelRate | None:
provider_key = provider.strip().lower()
model_key = model.strip().lower()
@ -56,15 +114,9 @@ def _rate(
break
if provider_key == "agy_cli":
if model_key == "gemini-3.6-flash":
return ModelRate(
1.50,
7.50,
0.15,
f"google-gemini-3.6-flash-standard@{RATE_CARD_VERSION}",
"Google Gemini 3.6 Flash 표준 단가 · 입력 $1.50/M · 캐시 $0.15/M · 출력 $7.50/M",
GOOGLE_PRICING_URL,
)
flash_rate = _google_flash_rate(model_key=model_key, priced_on=priced_on)
if flash_rate is not None:
return flash_rate
if model_key == "gemini-3.5-flash":
return ModelRate(
1.50,
@ -182,6 +234,7 @@ def estimate_reference_cost(
model: str,
tokens_in: int,
tokens_out: int,
priced_at: date | datetime | int | float | str,
cached_input_tokens: int = 0,
) -> CostEstimate | None:
"""공식 공개 단가로 USD 상당액을 계산한다.
@ -195,7 +248,12 @@ def estimate_reference_cost(
if safe_input == 0 and safe_output == 0:
return None
cached = min(safe_input, max(0, int(cached_input_tokens or 0)))
rate = _rate(provider=provider, model=model, tokens_in=safe_input)
rate = _rate(
provider=provider,
model=model,
tokens_in=safe_input,
priced_on=_pricing_date(priced_at),
)
if rate is None:
return None
uncached = safe_input - cached

66
apps/api/app/services/usage_report.py
View file

@ -66,6 +66,8 @@ def build_model_cost_report(usage: Mapping[str, Any]) -> dict[str, Any]:
by_provider = list(usage.get("by_provider") or [])
budget = dict(usage.get("budget") or {})
evaluator_cache = dict(usage.get("evaluator_cache") or {})
total_cost_basis = str(usage.get("cost_basis") or "provider_reported")
exact_cost_bases = {"provider_estimate", "provider_reported", "reference_rate"}
models: list[dict[str, Any]] = []
for item in by_provider:
@ -87,6 +89,12 @@ def build_model_cost_report(usage: Mapping[str, Any]) -> dict[str, Any]:
_number(row.get("recorded_cost_usd"), row_cost), 6
)
row_estimated_cost = round(_number(row.get("estimated_cost_usd")), 6)
row_cost_basis = str(row.get("cost_basis") or "provider_reported")
row_cost_complete = row_cost_basis not in {
"partial",
"partial_upper_bound",
"unavailable",
}
models.append(
{
"provider": str(row.get("provider") or "unknown"),
@ -100,13 +108,24 @@ def build_model_cost_report(usage: Mapping[str, Any]) -> dict[str, Any]:
"cost_usd": row_cost,
"recorded_cost_usd": row_recorded_cost,
"estimated_cost_usd": row_estimated_cost,
"cost_basis": str(row.get("cost_basis") or "provider_reported"),
"cost_basis": row_cost_basis,
"rate_label": row.get("rate_label"),
"rate_source_url": row.get("rate_source_url"),
"cost_share": _ratio(row_cost, total_cost),
"cost_share": (
_ratio(row_cost, total_cost)
if total_cost_basis in exact_cost_bases
and row_cost_basis in exact_cost_bases
else None
),
"token_share": _ratio(float(row_tokens), float(total_tokens)),
"cost_per_turn_usd": _cost_per_turn(row_cost, turns),
"cost_per_1k_tokens_usd": _cost_per_1k_tokens(row_cost, row_tokens),
"cost_per_turn_usd": (
_cost_per_turn(row_cost, turns) if row_cost_complete else None
),
"cost_per_1k_tokens_usd": (
_cost_per_1k_tokens(row_cost, row_tokens)
if row_cost_complete
else None
),
}
)
models.sort(key=lambda item: (-float(item["cost_usd"]), item["provider"], item["model"]))
@ -120,15 +139,28 @@ def build_model_cost_report(usage: Mapping[str, Any]) -> dict[str, Any]:
warnings.append("zero_cost_metered_usage")
if estimated_cost > 0:
warnings.append("reference_rate_cost")
if any(item["cost_basis"] == "unavailable" for item in models):
if any(item["cost_basis"] == "partial" for item in models):
warnings.append("partial_model_cost")
if any(item["cost_basis"] == "partial_upper_bound" for item in models):
warnings.append("partial_upper_bound_model_cost")
if any(item["cost_basis"] == "reference_upper_bound" for item in models):
warnings.append("reference_upper_bound_cost")
if any(
item["cost_basis"] in {"partial", "partial_upper_bound", "unavailable"}
for item in models
):
warnings.append("unavailable_model_cost")
if str(budget.get("status") or "") in {"warn", "exceeded"}:
if str(budget.get("status") or "") in {"warn", "exceeded", "indeterminate"}:
warnings.append(f"budget_{budget.get('status')}")
cache_hit_rate = _number(evaluator_cache.get("hit_rate"))
if bool(evaluator_cache.get("enabled")) and _integer(evaluator_cache.get("requests")) > 0:
if cache_hit_rate < 0.25:
warnings.append("low_evaluator_cache_hit_rate")
if models and models[0]["cost_share"] >= 0.8:
if (
models
and isinstance(models[0]["cost_share"], (float, int))
and models[0]["cost_share"] >= 0.8
):
warnings.append("dominant_model_cost")
return {
@ -152,14 +184,28 @@ def build_model_cost_report(usage: Mapping[str, Any]) -> dict[str, Any]:
"cost_usd": total_cost,
"recorded_cost_usd": recorded_cost,
"estimated_cost_usd": estimated_cost,
"cost_per_turn_usd": _cost_per_turn(total_cost, metered_turns),
"cost_per_1k_tokens_usd": _cost_per_1k_tokens(total_cost, total_tokens),
"cost_basis": total_cost_basis,
"cost_per_turn_usd": (
_cost_per_turn(total_cost, metered_turns)
if total_cost_basis not in {"partial", "partial_upper_bound", "unavailable"}
else None
),
"cost_per_1k_tokens_usd": (
_cost_per_1k_tokens(total_cost, total_tokens)
if total_cost_basis not in {"partial", "partial_upper_bound", "unavailable"}
else None
),
},
"budget": {
"status": str(budget.get("status") or "disabled"),
"limit_usd": round(_number(budget.get("limit_usd")), 6),
"used_ratio": round(_number(budget.get("used_ratio")), 6),
"remaining_usd": round(_number(budget.get("remaining_usd")), 6),
"remaining_usd": (
round(_number(budget.get("remaining_usd")), 6)
if budget.get("remaining_usd") is not None
else None
),
"cost_basis": str(budget.get("cost_basis") or total_cost_basis),
},
"evaluator_cache": {
"enabled": bool(evaluator_cache.get("enabled")),