diff --git a/basedpyright-code-budget.json b/basedpyright-code-budget.json index 1b0650c70d8..9b59480a0dc 100644 --- a/basedpyright-code-budget.json +++ b/basedpyright-code-budget.json @@ -93,7 +93,7 @@ "limit": 181 }, "reportTypedDictNotRequiredAccess": { - "limit": 24 + "limit": 22 }, "reportUndefinedVariable": { "limit": 0 diff --git a/litellm-proxy-extras/litellm_proxy_extras/migrations/20260901000001_add_guardrail_usage_units_cost/migration.sql b/litellm-proxy-extras/litellm_proxy_extras/migrations/20260901000001_add_guardrail_usage_units_cost/migration.sql new file mode 100644 index 00000000000..a89b7c4c6f8 --- /dev/null +++ b/litellm-proxy-extras/litellm_proxy_extras/migrations/20260901000001_add_guardrail_usage_units_cost/migration.sql @@ -0,0 +1,3 @@ +-- AlterTable +ALTER TABLE "LiteLLM_DailyGuardrailUsageUnits" ADD COLUMN IF NOT EXISTS "cost" DOUBLE PRECISION; +ALTER TABLE "LiteLLM_DailyGuardrailUsageUnits" ADD COLUMN IF NOT EXISTS "untracked_units" BIGINT NOT NULL DEFAULT 0; diff --git a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma index 2a2665f9731..e638ad68b6f 100644 --- a/litellm-proxy-extras/litellm_proxy_extras/schema.prisma +++ b/litellm-proxy-extras/litellm_proxy_extras/schema.prisma @@ -1124,6 +1124,8 @@ model LiteLLM_DailyGuardrailUsageUnits { api_key String // hashed virtual key; empty string when unknown usage_unit String // provider counter name, e.g. Bedrock's contentPolicyUnits units BigInt @default(0) + cost Float? // USD for the priced share of units; null only on rows written before this column existed + untracked_units BigInt @default(0) // units recorded with no known price, the share cost leaves out created_at DateTime @default(now()) updated_at DateTime @updatedAt diff --git a/litellm/litellm_core_utils/llm_cost_calc/guardrail_cost.py b/litellm/litellm_core_utils/llm_cost_calc/guardrail_cost.py index ad1880d4cc2..54cdf2cb8ff 100644 --- a/litellm/litellm_core_utils/llm_cost_calc/guardrail_cost.py +++ b/litellm/litellm_core_utils/llm_cost_calc/guardrail_cost.py @@ -1,8 +1,8 @@ import math from collections.abc import Mapping -from typing import Final +from typing import Annotated, Final -from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError +from pydantic import BaseModel, ConfigDict, Field, TypeAdapter, ValidationError import litellm from litellm._logging import verbose_logger @@ -30,6 +30,31 @@ class GuardrailCostEntry(BaseModel): _GUARDRAIL_COST_ENTRY_ADAPTER: Final[TypeAdapter[GuardrailCostEntry]] = TypeAdapter(GuardrailCostEntry) +class GuardrailCostByUnitEntry(BaseModel): + """The rollup-side view of a ``guardrail_information`` entry, validated apart from + ``GuardrailCostEntry`` so a forged per-counter map can never zero the spend path.""" + + model_config = ConfigDict(extra="ignore", frozen=True) + + guardrail_cost_by_unit: Mapping[str, Annotated[float, Field(ge=0, allow_inf_nan=False)] | None] | None = None + guardrail_cost_in_spend: bool | None = True + + +_GUARDRAIL_COST_BY_UNIT_ADAPTER: Final[TypeAdapter[GuardrailCostByUnitEntry]] = TypeAdapter(GuardrailCostByUnitEntry) + + +def billed_guardrail_cost_by_unit(raw: object) -> Mapping[str, float | None] | None: + """Per-counter USD the daily rollup may record for one raw ``guardrail_information`` + entry; None when the entry is unpriced, report-only, or malformed, and None per + counter the hook had no price for.""" + try: + entry: Final = _GUARDRAIL_COST_BY_UNIT_ADAPTER.validate_python(raw) + except ValidationError as e: + verbose_logger.warning("Ignoring malformed guardrail_information entry for guardrail cost rollup: %s", e) + return None + return None if entry.guardrail_cost_in_spend is False else entry.guardrail_cost_by_unit + + def _bedrock_guardrail_pricing(aws_region_name: str | None) -> GuardrailPricing | None: regional_key: Final = f"bedrock/{aws_region_name}/guardrails" if aws_region_name else None for key in (regional_key, BEDROCK_GUARDRAIL_PRICING_KEY): @@ -42,11 +67,32 @@ def _bedrock_guardrail_pricing(aws_region_name: str | None) -> GuardrailPricing return None -def bedrock_guardrail_cost(usage_units: Mapping[str, int], aws_region_name: str | None) -> float: +def _priced_units(units: int, price_per_unit: float | None) -> float | None: + return None if price_per_unit is None else units * price_per_unit + + +def bedrock_guardrail_cost_by_unit( + usage_units: Mapping[str, int], aws_region_name: str | None +) -> Mapping[str, float | None] | None: + """USD per counter, keyed like ``usage_units``; None when no pricing entry exists, + and None for a counter the entry has no price for, since only an explicit 0.0 means free.""" pricing: Final = _bedrock_guardrail_pricing(aws_region_name) if pricing is None: - return 0.0 - return sum(units * pricing.guardrail_cost_per_unit.get(counter, 0.0) for counter, units in usage_units.items()) + return None + return { # mutable-ok: stamped into guardrail_information, which safe_dumps only serializes as a plain dict + counter: _priced_units(units, pricing.guardrail_cost_per_unit.get(counter)) + for counter, units in usage_units.items() + } + + +def guardrail_cost_total(cost_by_unit: Mapping[str, float | None] | None) -> float: + """The scalar the spend path bills: unknown-priced counters count as 0 here, the + rollup keeps them unknown.""" + return sum(cost for cost in cost_by_unit.values() if cost is not None) if cost_by_unit is not None else 0.0 + + +def bedrock_guardrail_cost(usage_units: Mapping[str, int], aws_region_name: str | None) -> float: + return guardrail_cost_total(bedrock_guardrail_cost_by_unit(usage_units, aws_region_name)) AZURE_PROMPT_SHIELD_TEXT_RECORD_UNIT: Final = "text_records" diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index e8e1f53b473..316bfb8cf92 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -13050,6 +13050,59 @@ ], "title": "Avgscore" }, + "cost": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ], + "title": "Cost" + }, + "cost_by_key": { + "additionalProperties": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ] + }, + "title": "Cost By Key", + "type": "object" + }, + "cost_by_team": { + "additionalProperties": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ] + }, + "title": "Cost By Team", + "type": "object" + }, + "cost_by_unit": { + "additionalProperties": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ] + }, + "title": "Cost By Unit", + "type": "object" + }, "description": { "anyOf": [ { @@ -13100,6 +13153,13 @@ "title": "Type", "type": "string" }, + "untracked_usage_units": { + "additionalProperties": { + "type": "integer" + }, + "title": "Untracked Usage Units", + "type": "object" + }, "usage_units": { "additionalProperties": { "type": "integer" @@ -13151,7 +13211,12 @@ "usage_units", "usage_units_daily", "usage_units_by_team", - "usage_units_by_key" + "usage_units_by_key", + "cost", + "cost_by_unit", + "cost_by_team", + "cost_by_key", + "untracked_usage_units" ], "title": "UsageDetailResponse", "type": "object" @@ -13306,10 +13371,28 @@ "title": "Totalblocked", "type": "integer" }, + "totalCost": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ], + "title": "Totalcost" + }, "totalRequests": { "title": "Totalrequests", "type": "integer" }, + "totalUntrackedUsageUnits": { + "additionalProperties": { + "type": "integer" + }, + "title": "Totaluntrackedusageunits", + "type": "object" + }, "totalUsageUnits": { "additionalProperties": { "type": "integer" @@ -13324,7 +13407,9 @@ "totalRequests", "totalBlocked", "passRate", - "totalUsageUnits" + "totalUsageUnits", + "totalCost", + "totalUntrackedUsageUnits" ], "title": "UsageOverviewResponse", "type": "object" @@ -13353,6 +13438,18 @@ ], "title": "Avgscore" }, + "cost": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ], + "description": "USD for the priced share of usageUnits over the window; null when no unit was priced", + "title": "Cost" + }, "failRate": { "title": "Failrate", "type": "number" @@ -13385,6 +13482,14 @@ "title": "Type", "type": "string" }, + "untrackedUsageUnits": { + "additionalProperties": { + "type": "integer" + }, + "description": "The share of usageUnits that cost leaves out: units recorded with no known price, per counter", + "title": "Untrackedusageunits", + "type": "object" + }, "usageUnits": { "additionalProperties": { "type": "integer" @@ -13404,13 +13509,26 @@ "avgLatency", "status", "trend", - "usageUnits" + "usageUnits", + "cost", + "untrackedUsageUnits" ], "title": "UsageOverviewRow", "type": "object" }, "UsageUnitsDailyPoint": { "properties": { + "cost": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ], + "title": "Cost" + }, "date": { "title": "Date", "type": "string" @@ -13425,7 +13543,8 @@ }, "required": [ "date", - "units" + "units", + "cost" ], "title": "UsageUnitsDailyPoint", "type": "object" @@ -28784,10 +28903,28 @@ "title": "Totalblocked", "type": "integer" }, + "totalCost": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ], + "title": "Totalcost" + }, "totalRequests": { "title": "Totalrequests", "type": "integer" }, + "totalUntrackedUsageUnits": { + "additionalProperties": { + "type": "integer" + }, + "title": "Totaluntrackedusageunits", + "type": "object" + }, "totalUsageUnits": { "additionalProperties": { "type": "integer" @@ -28802,7 +28939,9 @@ "totalRequests", "totalBlocked", "passRate", - "totalUsageUnits" + "totalUsageUnits", + "totalCost", + "totalUntrackedUsageUnits" ], "title": "UsageOverviewResponse", "type": "object" @@ -28831,6 +28970,18 @@ ], "title": "Avgscore" }, + "cost": { + "anyOf": [ + { + "type": "number" + }, + { + "type": "null" + } + ], + "description": "USD for the priced share of usageUnits over the window; null when no unit was priced", + "title": "Cost" + }, "failRate": { "title": "Failrate", "type": "number" @@ -28863,6 +29014,14 @@ "title": "Type", "type": "string" }, + "untrackedUsageUnits": { + "additionalProperties": { + "type": "integer" + }, + "description": "The share of usageUnits that cost leaves out: units recorded with no known price, per counter", + "title": "Untrackedusageunits", + "type": "object" + }, "usageUnits": { "additionalProperties": { "type": "integer" @@ -28882,7 +29041,9 @@ "avgLatency", "status", "trend", - "usageUnits" + "usageUnits", + "cost", + "untrackedUsageUnits" ], "title": "UsageOverviewRow", "type": "object" diff --git a/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py b/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py index 7f2616c2fb8..7204839f6d3 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py +++ b/litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py @@ -36,7 +36,10 @@ from litellm.litellm_core_utils.core_helpers import redact_nested_match_and_rege from litellm.litellm_core_utils.litellm_logging import ( _get_masked_values, # pyright: ignore[reportPrivateUsage] # the shared header-masking helper has no public name ) -from litellm.litellm_core_utils.llm_cost_calc.guardrail_cost import bedrock_guardrail_cost +from litellm.litellm_core_utils.llm_cost_calc.guardrail_cost import ( + bedrock_guardrail_cost_by_unit, + guardrail_cost_total, +) from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler from litellm.llms.base_llm.guardrail_translation.utils import ( effective_scan_only_tool_results_for_guardrail, @@ -109,6 +112,7 @@ _BEDROCK_TOO_LARGE_ERROR_SUBSTRINGS: Final = ( _BEDROCK_APPLY_GUARDRAIL_MAX_THROTTLE_RETRIES: Final = 3 _BEDROCK_APPLY_GUARDRAIL_BASE_BACKOFF_SECONDS: Final = 0.5 _BEDROCK_WHITESPACE: Final = re.compile(r"\s") +_NO_TRACING_DETAIL: Final[GuardrailTracingDetail] = {} # Resource-less, detect-only InvokeGuardrailChecks API (no guardrail resource required). _BEDROCK_INVOKE_GUARDRAIL_CHECKS_PATH: Final = "/guardrail-checks/invoke" # InvokeGuardrailChecks accepts at most 10 content blocks per message. A message with @@ -2155,25 +2159,37 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM): OTEL integration can expose it as a queryable span attribute without re-parsing the redacted guardrail_response blob. """ - tracing_detail: Final[GuardrailTracingDetail] = {} violation_categories: Final = self._extract_violation_category_names(response) - if violation_categories: - tracing_detail["violation_categories"] = violation_categories bedrock_action: Final = response.get("action") - if isinstance(bedrock_action, str): - tracing_detail["guardrail_action"] = bedrock_action - usage: Final = response.get("usage") - if isinstance(usage, dict): - usage_units: Final = { # mutable-ok: json.dumps'd into spend log metadata downstream - key: value for key, value in usage.items() if isinstance(value, int) - } - if usage_units: - tracing_detail["guardrail_usage"] = usage_units - tracing_detail["guardrail_cost"] = bedrock_guardrail_cost( - usage_units=usage_units, aws_region_name=aws_region_name - ) + categories_detail: Final[GuardrailTracingDetail] = {"violation_categories": violation_categories} + action_detail: Final[GuardrailTracingDetail] = {"guardrail_action": bedrock_action} + tracing_detail: Final[GuardrailTracingDetail] = { + **(categories_detail if violation_categories else _NO_TRACING_DETAIL), + **(action_detail if isinstance(bedrock_action, str) else _NO_TRACING_DETAIL), + **self._usage_tracing_detail(response.get("usage"), aws_region_name), + } return tracing_detail + @staticmethod + def _usage_tracing_detail( + usage: BedrockGuardrailUsage | None, aws_region_name: str | None + ) -> GuardrailTracingDetail: + if not isinstance(usage, dict): + return _NO_TRACING_DETAIL + usage_units: Final = { # mutable-ok: json.dumps'd into spend log metadata downstream + key: value for key, value in usage.items() if isinstance(value, int) + } + if not usage_units: + return _NO_TRACING_DETAIL + cost_by_unit: Final = bedrock_guardrail_cost_by_unit(usage_units=usage_units, aws_region_name=aws_region_name) + priced_detail: Final[GuardrailTracingDetail] = {"guardrail_cost_by_unit": cost_by_unit} + usage_detail: Final[GuardrailTracingDetail] = { + "guardrail_usage": usage_units, + "guardrail_cost": guardrail_cost_total(cost_by_unit), + **(priced_detail if cost_by_unit is not None else _NO_TRACING_DETAIL), + } + return usage_detail + def _extract_violation_category_names(self, response: BedrockGuardrailResponse) -> list[str]: """ Flatten the BLOCKED assessments into a list of human-readable category diff --git a/litellm/proxy/guardrails/usage_endpoints.py b/litellm/proxy/guardrails/usage_endpoints.py index 7a0edbddca8..0390a2b5013 100644 --- a/litellm/proxy/guardrails/usage_endpoints.py +++ b/litellm/proxy/guardrails/usage_endpoints.py @@ -8,10 +8,10 @@ from collections.abc import Callable, Iterable, Mapping, Sequence from datetime import date, datetime, timedelta, timezone from itertools import groupby from types import MappingProxyType -from typing import TYPE_CHECKING, Any, Final, Literal, overload +from typing import TYPE_CHECKING, Any, Final, Literal, TypeVar, overload from fastapi import APIRouter, Depends, Query -from pydantic import BaseModel +from pydantic import BaseModel, Field from typing_extensions import NotRequired, ReadOnly, TypedDict from litellm._logging import verbose_proxy_logger @@ -42,6 +42,8 @@ router: Final = APIRouter() _EMPTY_UNITS: Final[Mapping[str, int]] = MappingProxyType({}) +_T = TypeVar("_T") + _USAGE_MAX_RANGE_DAYS: Final = 366 @@ -154,6 +156,16 @@ def _counter_name(row: "prisma_models.LiteLLM_DailyGuardrailUsageUnits") -> str: return row.usage_unit +def _row_untracked_units(row: "prisma_models.LiteLLM_DailyGuardrailUsageUnits") -> int: + """A row written before the cost column carries NULL cost and is untracked in full.""" + return int(row.units) if row.cost is None else int(row.untracked_units) + + +def _row_tracked_cost(row: "prisma_models.LiteLLM_DailyGuardrailUsageUnits") -> float | None: + """The row's cost when it prices at least one unit; None when every unit is untracked.""" + return None if row.cost is None or _row_untracked_units(row) >= int(row.units) else row.cost + + def _sum_counter_units(rows: "Iterable[prisma_models.LiteLLM_DailyGuardrailUsageUnits]") -> Mapping[str, int]: ordered: Final = sorted(rows, key=_counter_name) return MappingProxyType( @@ -161,12 +173,31 @@ def _sum_counter_units(rows: "Iterable[prisma_models.LiteLLM_DailyGuardrailUsage ) -def _units_by( +def _sum_untracked_units(rows: "Iterable[prisma_models.LiteLLM_DailyGuardrailUsageUnits]") -> Mapping[str, int]: + ordered: Final = sorted(rows, key=_counter_name) + per_counter: Final = tuple( + (name, sum(map(_row_untracked_units, group))) for name, group in groupby(ordered, key=_counter_name) + ) + return MappingProxyType({name: units for name, units in per_counter if units}) + + +def _sum_tracked_cost(rows: "Iterable[prisma_models.LiteLLM_DailyGuardrailUsageUnits]") -> float | None: + """Sum over rows that price at least one unit; None when no row does.""" + tracked: Final = tuple(cost for cost in map(_row_tracked_cost, rows) if cost is not None) + return sum(tracked) if tracked else None + + +def _by( rows: "Sequence[prisma_models.LiteLLM_DailyGuardrailUsageUnits]", key_of: "Callable[[prisma_models.LiteLLM_DailyGuardrailUsageUnits], str]", -) -> Mapping[str, Mapping[str, int]]: + reduce: "Callable[[Iterable[prisma_models.LiteLLM_DailyGuardrailUsageUnits]], _T]", +) -> Mapping[str, _T]: ordered: Final = sorted(rows, key=key_of) - return MappingProxyType({key: _sum_counter_units(group) for key, group in groupby(ordered, key=key_of)}) + return MappingProxyType({key: reduce(group) for key, group in groupby(ordered, key=key_of)}) + + +def _first_match(lookup_keys: Sequence[str], mapping: Mapping[str, _T], default: _T) -> _T: + return next((mapping[k] for k in lookup_keys if k in mapping), default) # --- Response models --- @@ -218,6 +249,12 @@ class UsageOverviewRow(BaseModel): status: str # healthy | warning | critical trend: str # up | down | stable usageUnits: Mapping[str, int] + cost: float | None = Field( + description="USD for the priced share of usageUnits over the window; null when no unit was priced" + ) + untrackedUsageUnits: Mapping[str, int] = Field( + description="The share of usageUnits that cost leaves out: units recorded with no known price, per counter" + ) class UsageOverviewResponse(BaseModel): @@ -227,11 +264,26 @@ class UsageOverviewResponse(BaseModel): totalBlocked: int passRate: float totalUsageUnits: Mapping[str, int] + totalCost: float | None + totalUntrackedUsageUnits: Mapping[str, int] + + +_EMPTY_OVERVIEW: Final = UsageOverviewResponse( + rows=[], + chart=[], + totalRequests=0, + totalBlocked=0, + passRate=100.0, + totalUsageUnits=_EMPTY_UNITS, + totalCost=None, + totalUntrackedUsageUnits=_EMPTY_UNITS, +) class UsageUnitsDailyPoint(BaseModel): date: str units: Mapping[str, int] + cost: float | None class UsageDetailResponse(BaseModel): @@ -251,6 +303,11 @@ class UsageDetailResponse(BaseModel): usage_units_daily: Sequence[UsageUnitsDailyPoint] usage_units_by_team: Mapping[str, Mapping[str, int]] usage_units_by_key: Mapping[str, Mapping[str, int]] + cost: float | None + cost_by_unit: Mapping[str, float | None] + cost_by_team: Mapping[str, float | None] + cost_by_key: Mapping[str, float | None] + untracked_usage_units: Mapping[str, int] class UsageLogEntry(BaseModel): @@ -367,6 +424,8 @@ def _guardrail_overview_rows( agg: Mapping[str, _MetricTotals], prev_agg: Mapping[str, float], units_agg: Mapping[str, Mapping[str, int]], + cost_agg: Mapping[str, float | None], + untracked_agg: Mapping[str, Mapping[str, int]], ) -> list[UsageOverviewRow]: rows: Final[list[UsageOverviewRow]] = [] covered_keys: Final[set[str]] = set() @@ -392,7 +451,6 @@ def _guardrail_overview_rows( prev_fail = float(prev_agg.get(k, 0.0) or 0.0) break trend = _trend_from_comparison(fail_rate, prev_fail) - row_units: Mapping[str, int] = next((units_agg[k] for k in lookup_keys if k in units_agg), _EMPTY_UNITS) rows.append( UsageOverviewRow( id=gid, @@ -405,7 +463,9 @@ def _guardrail_overview_rows( avgLatency=None, status=_status_from_fail_rate(fail_rate), trend=trend, - usageUnits=row_units, + usageUnits=_first_match(lookup_keys, units_agg, _EMPTY_UNITS), + cost=_first_match(lookup_keys, cost_agg, None), + untrackedUsageUnits=_first_match(lookup_keys, untracked_agg, _EMPTY_UNITS), ) ) # Add rows for guardrails with metrics but not in guardrails table (e.g. MCP, config) @@ -429,6 +489,8 @@ def _guardrail_overview_rows( status=_status_from_fail_rate(fail_rate), trend=trend, usageUnits=units_agg.get(agg_key, _EMPTY_UNITS), + cost=cost_agg.get(agg_key), + untrackedUsageUnits=untracked_agg.get(agg_key, _EMPTY_UNITS), ) ) return rows @@ -459,6 +521,8 @@ def _policy_overview_rows( status=_status_from_fail_rate(fail_rate), trend=trend, usageUnits=_EMPTY_UNITS, + cost=None, + untrackedUsageUnits=_EMPTY_UNITS, ) ) return rows @@ -479,9 +543,7 @@ async def guardrails_usage_overview( from litellm.proxy.proxy_server import prisma_client if prisma_client is None: - return UsageOverviewResponse( - rows=[], chart=[], totalRequests=0, totalBlocked=0, passRate=100.0, totalUsageUnits=_EMPTY_UNITS - ) + return _EMPTY_OVERVIEW start, end = _resolve_usage_window(start_date, end_date) @@ -515,12 +577,14 @@ async def guardrails_usage_overview( agg: Final = _aggregate_daily_metrics(metrics, "guardrail_id") prev_agg: Final = _prev_fail_rates(metrics_prev, "guardrail_id") - units_agg: Final = _units_by(units_rows, lambda r: r.guardrail_id) + units_agg: Final = _by(units_rows, lambda r: r.guardrail_id, _sum_counter_units) + cost_agg: Final = _by(units_rows, lambda r: r.guardrail_id, _sum_tracked_cost) + untracked_agg: Final = _by(units_rows, lambda r: r.guardrail_id, _sum_untracked_units) chart: Final = _chart_from_metrics(metrics) total_requests: Final = sum(a["requests"] for a in agg.values()) total_blocked: Final = sum(a["blocked"] for a in agg.values()) pass_rate: Final = (100.0 * (total_requests - total_blocked) / total_requests) if total_requests else 100.0 - rows: Final = _guardrail_overview_rows(guardrails, agg, prev_agg, units_agg) + rows: Final = _guardrail_overview_rows(guardrails, agg, prev_agg, units_agg, cost_agg, untracked_agg) return UsageOverviewResponse( rows=rows, chart=chart, @@ -528,6 +592,8 @@ async def guardrails_usage_overview( totalBlocked=total_blocked, passRate=round(pass_rate, 1), totalUsageUnits=_sum_counter_units(units_rows), + totalCost=_sum_tracked_cost(units_rows), + totalUntrackedUsageUnits=_sum_untracked_units(units_rows), ) except Exception as e: from litellm.proxy.utils import handle_exception_on_proxy @@ -618,8 +684,11 @@ async def guardrails_usage_detail( litellm_params: Final = _to_dict(_get_guardrail_field(guardrail, "litellm_params")) guardrail_info: Final = _to_dict(_get_guardrail_field(guardrail, "guardrail_info")) _guardrail_name: Final = _get_guardrail_field(guardrail, "guardrail_name") - daily_unit_sums: Final = sorted(_units_by(units_rows, lambda r: r.date).items()) - units_daily: Final = tuple(UsageUnitsDailyPoint(date=d, units=units) for d, units in daily_unit_sums) + daily_unit_sums: Final = sorted(_by(units_rows, lambda r: r.date, _sum_counter_units).items()) + daily_cost: Final = _by(units_rows, lambda r: r.date, _sum_tracked_cost) + units_daily: Final = tuple( + UsageUnitsDailyPoint(date=d, units=units, cost=daily_cost.get(d)) for d, units in daily_unit_sums + ) return UsageDetailResponse( guardrail_id=guardrail_id, @@ -636,8 +705,13 @@ async def guardrails_usage_detail( time_series=time_series, usage_units=_sum_counter_units(units_rows), usage_units_daily=units_daily, - usage_units_by_team=_units_by(units_rows, lambda r: r.team_id), - usage_units_by_key=_units_by(units_rows, lambda r: r.api_key), + usage_units_by_team=_by(units_rows, lambda r: r.team_id, _sum_counter_units), + usage_units_by_key=_by(units_rows, lambda r: r.api_key, _sum_counter_units), + cost=_sum_tracked_cost(units_rows), + cost_by_unit=_by(units_rows, _counter_name, _sum_tracked_cost), + cost_by_team=_by(units_rows, lambda r: r.team_id, _sum_tracked_cost), + cost_by_key=_by(units_rows, lambda r: r.api_key, _sum_tracked_cost), + untracked_usage_units=_sum_untracked_units(units_rows), ) @@ -857,9 +931,7 @@ async def policies_usage_overview( from litellm.proxy.proxy_server import prisma_client if prisma_client is None: - return UsageOverviewResponse( - rows=[], chart=[], totalRequests=0, totalBlocked=0, passRate=100.0, totalUsageUnits=_EMPTY_UNITS - ) + return _EMPTY_OVERVIEW start, end = _resolve_usage_window(start_date, end_date) @@ -891,6 +963,8 @@ async def policies_usage_overview( totalBlocked=total_blocked, passRate=round(pass_rate, 1), totalUsageUnits=_EMPTY_UNITS, + totalCost=None, + totalUntrackedUsageUnits=_EMPTY_UNITS, ) except Exception as e: from litellm.proxy.utils import handle_exception_on_proxy diff --git a/litellm/proxy/guardrails/usage_tracking.py b/litellm/proxy/guardrails/usage_tracking.py index b8ae09afc00..cb6aec14f8c 100644 --- a/litellm/proxy/guardrails/usage_tracking.py +++ b/litellm/proxy/guardrails/usage_tracking.py @@ -6,7 +6,7 @@ insert into SpendLogGuardrailIndex when spend logs are written. import asyncio import json from collections import defaultdict -from collections.abc import Awaitable, Callable, Iterator, Mapping, Sequence +from collections.abc import Awaitable, Callable, Iterable, Iterator, Mapping, Sequence from datetime import datetime, timezone from functools import partial from itertools import groupby @@ -17,6 +17,7 @@ from typing import TYPE_CHECKING, Any, Final, NamedTuple, TypeVar from typing_extensions import ReadOnly, TypedDict from litellm._logging import verbose_proxy_logger +from litellm.litellm_core_utils.llm_cost_calc.guardrail_cost import billed_guardrail_cost_by_unit from litellm.proxy._types import DB_RETRY_SAFE_ERROR_TYPES from litellm.proxy.utils import PrismaClient from litellm.repositories.table_repositories import ( @@ -44,6 +45,20 @@ class _UsageUnitKey(NamedTuple): usage_unit: str +class _UsageUnitIncrement(NamedTuple): + units: int + cost: float + """USD for the priced share of units.""" + untracked_units: int + """Units recorded with no known price, the share cost leaves out.""" + + +def _usage_unit_increment(units: int, cost: float | None) -> _UsageUnitIncrement: + if cost is None: + return _UsageUnitIncrement(units=units, cost=0.0, untracked_units=units) + return _UsageUnitIncrement(units=units, cost=cost, untracked_units=0) + + class _MetricsKey(NamedTuple): guardrail_id: str date: str @@ -67,22 +82,37 @@ class PendingRollups: def __init__(self) -> None: self.lock: Final = asyncio.Lock() self.metrics: Mapping[_MetricsKey, Mapping[str, int]] = MappingProxyType({}) - self.units: Mapping[_UsageUnitKey, int] = MappingProxyType({}) + self.units: Mapping[_UsageUnitKey, _UsageUnitIncrement] = MappingProxyType({}) _PENDING_ROLLUPS: Final = PendingRollups() _NO_COUNTERS: Final[Mapping[str, int]] = MappingProxyType({}) +_NO_INCREMENT: Final = _UsageUnitIncrement(units=0, cost=0.0, untracked_units=0) def _merged_keys(base: Mapping[_RowKey, object], extra: Mapping[_RowKey, object]) -> tuple[_RowKey, ...]: return (*base, *(key for key in extra if key not in base)) +def _summed_increments(increments: Iterable[_UsageUnitIncrement]) -> _UsageUnitIncrement: + materialized: Final = tuple(increments) + return _UsageUnitIncrement( + units=sum(i.units for i in materialized), + cost=sum(i.cost for i in materialized), + untracked_units=sum(i.untracked_units for i in materialized), + ) + + def _merged_unit_rows( - base: Mapping[_UsageUnitKey, int], extra: Mapping[_UsageUnitKey, int] -) -> Mapping[_UsageUnitKey, int]: - return MappingProxyType({key: base.get(key, 0) + extra.get(key, 0) for key in _merged_keys(base, extra)}) + base: Mapping[_UsageUnitKey, _UsageUnitIncrement], extra: Mapping[_UsageUnitKey, _UsageUnitIncrement] +) -> Mapping[_UsageUnitKey, _UsageUnitIncrement]: + return MappingProxyType( + { + key: _summed_increments((base.get(key, _NO_INCREMENT), extra.get(key, _NO_INCREMENT))) + for key in _merged_keys(base, extra) + } + ) def _merged_metric_rows( @@ -209,7 +239,9 @@ def _parse_payload_start_time(payload: Mapping[str, Any]) -> datetime | None: return None -def _iter_usage_unit_increments(logs_to_process: Sequence[Mapping[str, Any]]) -> Iterator[tuple[_UsageUnitKey, int]]: +def _iter_usage_unit_increments( + logs_to_process: Sequence[Mapping[str, Any]], +) -> Iterator[tuple[_UsageUnitKey, _UsageUnitIncrement]]: for payload in logs_to_process: start_time = _parse_payload_start_time(payload) if not payload.get("request_id") or start_time is None: @@ -222,26 +254,38 @@ def _iter_usage_unit_increments(logs_to_process: Sequence[Mapping[str, Any]]) -> usage = entry.get("guardrail_usage") if not guardrail_id or not isinstance(usage, dict): continue + cost_by_unit = billed_guardrail_cost_by_unit(entry) for unit_name, units in usage.items(): if isinstance(units, int) and not isinstance(units, bool) and units > 0: - yield _UsageUnitKey(guardrail_id, date_key, team_id, api_key, str(unit_name)), units + key = _UsageUnitKey(guardrail_id, date_key, team_id, api_key, str(unit_name)) + cost = cost_by_unit.get(str(unit_name)) if cost_by_unit is not None else None + yield key, _usage_unit_increment(units=units, cost=cost) -def _sum_usage_unit_increments(logs_to_process: Sequence[Mapping[str, Any]]) -> Mapping[_UsageUnitKey, int]: +def _sum_usage_unit_increments( + logs_to_process: Sequence[Mapping[str, Any]], +) -> Mapping[_UsageUnitKey, _UsageUnitIncrement]: ordered: Final = sorted(_iter_usage_unit_increments(logs_to_process), key=itemgetter(0)) return MappingProxyType( - {key: sum(units for _, units in group) for key, group in groupby(ordered, key=itemgetter(0))} + { + key: _summed_increments(increment for _, increment in group) + for key, group in groupby(ordered, key=itemgetter(0)) + } ) -async def _upsert_usage_unit_row(prisma_client: PrismaClient, key: _UsageUnitKey, units: int) -> None: +async def _upsert_usage_unit_row( + prisma_client: PrismaClient, key: _UsageUnitKey, increment: _UsageUnitIncrement +) -> None: row: Final[prisma_types.LiteLLM_DailyGuardrailUsageUnitsCreateInput] = { "guardrail_id": key.guardrail_id, "date": key.date, "team_id": key.team_id, "api_key": key.api_key, "usage_unit": key.usage_unit, - "units": units, + "units": increment.units, + "cost": increment.cost, + "untracked_units": increment.untracked_units, } where: Final[_UsageUnitWhereUnique] = { "guardrail_id_date_team_id_api_key_usage_unit": { @@ -252,9 +296,14 @@ async def _upsert_usage_unit_row(prisma_client: PrismaClient, key: _UsageUnitKey "usage_unit": key.usage_unit, } } + # A row written before the cost column has NULL cost, and NULL + x stays NULL, so it keeps reading as unknown data: Final[prisma_types.LiteLLM_DailyGuardrailUsageUnitsUpsertInput] = { "create": row, - "update": {"units": {"increment": units}}, + "update": { + "units": {"increment": increment.units}, + "cost": {"increment": increment.cost}, + "untracked_units": {"increment": increment.untracked_units}, + }, } await DailyGuardrailUsageUnitsRepository(prisma_client).table.upsert(where=where, data=data) diff --git a/litellm/proxy/schema.prisma b/litellm/proxy/schema.prisma index 2a2665f9731..e638ad68b6f 100644 --- a/litellm/proxy/schema.prisma +++ b/litellm/proxy/schema.prisma @@ -1124,6 +1124,8 @@ model LiteLLM_DailyGuardrailUsageUnits { api_key String // hashed virtual key; empty string when unknown usage_unit String // provider counter name, e.g. Bedrock's contentPolicyUnits units BigInt @default(0) + cost Float? // USD for the priced share of units; null only on rows written before this column existed + untracked_units BigInt @default(0) // units recorded with no known price, the share cost leaves out created_at DateTime @default(now()) updated_at DateTime @updatedAt diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 26a132cec2d..44c312e2e77 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -3155,6 +3155,12 @@ class StandardLoggingGuardrailInformation(TypedDict, total=False): provider hook. Summed into the request's ``response_cost`` so it counts against spend and budgets like token cost, unless ``guardrail_cost_in_spend`` is False.""" + guardrail_cost_by_unit: ReadOnly[Mapping[str, float | None] | None] + """``guardrail_cost`` split per ``guardrail_usage`` counter, so the daily + per-counter usage rollup can carry cost at its own grain. Absent when the + hook had no pricing for the invocation; a counter is None when the pricing + entry has no price for it, which the rollup stores as unknown rather than $0.""" + guardrail_cost_in_spend: ReadOnly[bool | None] """Whether ``guardrail_cost`` participates in the request's ``response_cost`` and the spend/budget aggregates built from it. Absent, None, or True keeps the default @@ -3206,6 +3212,7 @@ class GuardrailTracingDetail(TypedDict, total=False): guardrail_action: str | None guardrail_usage: ReadOnly[Mapping[str, int] | None] guardrail_cost: ReadOnly[float | None] + guardrail_cost_by_unit: ReadOnly[Mapping[str, float | None] | None] guardrail_cost_in_spend: ReadOnly[bool | None] diff --git a/schema.prisma b/schema.prisma index 2a2665f9731..e638ad68b6f 100644 --- a/schema.prisma +++ b/schema.prisma @@ -1124,6 +1124,8 @@ model LiteLLM_DailyGuardrailUsageUnits { api_key String // hashed virtual key; empty string when unknown usage_unit String // provider counter name, e.g. Bedrock's contentPolicyUnits units BigInt @default(0) + cost Float? // USD for the priced share of units; null only on rows written before this column existed + untracked_units BigInt @default(0) // units recorded with no known price, the share cost leaves out created_at DateTime @default(now()) updated_at DateTime @updatedAt diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_guardrail_cost.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_guardrail_cost.py index baaef31036c..af2f169157e 100644 --- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_guardrail_cost.py +++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_guardrail_cost.py @@ -5,7 +5,10 @@ import pytest import litellm from litellm.litellm_core_utils.llm_cost_calc.guardrail_cost import ( bedrock_guardrail_cost, + bedrock_guardrail_cost_by_unit, + billed_guardrail_cost_by_unit, cost_breakdown_with_guardrail, + guardrail_cost_total, guardrail_information_cost, ) @@ -56,6 +59,68 @@ def test_bedrock_guardrail_cost_no_pricing_entry(monkeypatch): assert bedrock_guardrail_cost(usage_units={"contentPolicyUnits": 1}, aws_region_name="us-east-1") == 0.0 +def test_bedrock_guardrail_cost_by_unit_prices_every_counter_it_was_given(synthetic_cost_map): + """LIT-5652: the daily rollup stores one row per counter, so pricing must come + back at that grain, keyed exactly like the usage. An explicit 0.0 in the cost + map is free; a counter the map does not list is unknown (None), never free, + while the scalar the spend path bills still sums only the known prices.""" + usage = {"contentPolicyUnits": 2, "topicPolicyUnits": 1, "wordPolicyUnits": 5, "someFutureCounter": 3} + by_unit = bedrock_guardrail_cost_by_unit(usage_units=usage, aws_region_name="us-east-1") + assert by_unit is not None + assert by_unit.keys() == usage.keys() + assert by_unit["contentPolicyUnits"] == pytest.approx(0.0003) + assert by_unit["topicPolicyUnits"] == pytest.approx(0.00015) + assert by_unit["wordPolicyUnits"] == 0.0 + assert by_unit["someFutureCounter"] is None + assert guardrail_cost_total(by_unit) == pytest.approx(0.00045) + assert guardrail_cost_total(by_unit) == pytest.approx( + bedrock_guardrail_cost(usage_units=usage, aws_region_name="us-east-1") + ) + + +def test_bedrock_guardrail_cost_by_unit_is_none_without_pricing_so_unpriced_is_not_free(monkeypatch): + """The scalar keeps returning 0.0 for the spend path; the per-unit view must + say "unknown" instead so the rollup stores NULL rather than a $0 that would + hide the exact silent-spend problem this feature exists to surface.""" + monkeypatch.setattr(litellm, "model_cost", {}) + assert bedrock_guardrail_cost_by_unit(usage_units={"contentPolicyUnits": 1}, aws_region_name="us-east-1") is None + + +def test_billed_guardrail_cost_by_unit_reads_the_hook_stamp(): + entry = { + "guardrail_name": "bedrock", + "guardrail_cost_by_unit": {"contentPolicyUnits": 0.15, "wordPolicyUnits": 0, "someFutureCounter": None}, + } + assert billed_guardrail_cost_by_unit(entry) == { + "contentPolicyUnits": 0.15, + "wordPolicyUnits": 0.0, + "someFutureCounter": None, + } + + +@pytest.mark.parametrize( + "entry", + [ + {"guardrail_name": "no-pricing", "guardrail_usage": {"contentPolicyUnits": 1}}, + {"guardrail_cost_by_unit": {"text_records": 0.5}, "guardrail_cost_in_spend": False}, + {"guardrail_cost_by_unit": {"contentPolicyUnits": -0.5}}, + {"guardrail_cost_by_unit": {"contentPolicyUnits": float("nan")}}, + {"guardrail_cost_by_unit": {"contentPolicyUnits": float("inf")}}, + {"guardrail_cost_by_unit": {"contentPolicyUnits": "bad"}}, + {"guardrail_cost_by_unit": "not-a-map"}, + {"guardrail_cost_by_unit": {"contentPolicyUnits": 0.1}, "guardrail_cost_in_spend": "maybe"}, + "not-an-entry", + ], +) +def test_billed_guardrail_cost_by_unit_is_none_when_unpriced_report_only_or_forged(entry): + assert billed_guardrail_cost_by_unit(entry) is None + + +def test_billed_guardrail_cost_by_unit_treats_none_in_spend_as_billed(): + entry = {"guardrail_cost_by_unit": {"contentPolicyUnits": 0.15}, "guardrail_cost_in_spend": None} + assert billed_guardrail_cost_by_unit(entry) == {"contentPolicyUnits": 0.15} + + def test_shipped_bedrock_guardrail_prices_match_aws_pricing_page(monkeypatch): monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") litellm.model_cost = litellm.get_model_cost_map(url="") diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py index 1e44d29b610..9842d88e8d1 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_bedrock_guardrails.py @@ -2961,9 +2961,7 @@ async def test_streaming_hook_reraises_guardrail_service_failures(): guardrail = _sse_guardrail() with patch.object(guardrail, "make_bedrock_api_request", new_callable=AsyncMock) as mock_api: - mock_api.side_effect = HTTPException( - status_code=500, detail="Bedrock guardrail throttle retries exhausted" - ) + mock_api.side_effect = HTTPException(status_code=500, detail="Bedrock guardrail throttle retries exhausted") with pytest.raises(HTTPException) as exc: await _drain_streaming_hook(guardrail) @@ -5097,13 +5095,46 @@ def test_build_tracing_detail_surfaces_usage_counters_and_cost(monkeypatch): detail = guardrail._build_tracing_detail( { "action": "GUARDRAIL_INTERVENED", - "usage": {"topicPolicyUnits": 1, "contentPolicyUnits": 2, "wordPolicyUnits": 0, "oddball": "not-an-int"}, + "usage": { + "topicPolicyUnits": 1, + "contentPolicyUnits": 2, + "wordPolicyUnits": 0, + "someFutureCounter": 3, + "oddball": "not-an-int", + }, }, aws_region_name="us-east-1", ) - assert detail["guardrail_usage"] == {"topicPolicyUnits": 1, "contentPolicyUnits": 2, "wordPolicyUnits": 0} + assert detail["guardrail_usage"] == { + "topicPolicyUnits": 1, + "contentPolicyUnits": 2, + "wordPolicyUnits": 0, + "someFutureCounter": 3, + } assert detail["guardrail_cost"] == pytest.approx(0.00045) + by_unit = detail["guardrail_cost_by_unit"] + assert by_unit is not None and by_unit.keys() == detail["guardrail_usage"].keys() + assert by_unit["topicPolicyUnits"] == pytest.approx(0.00015) + assert by_unit["contentPolicyUnits"] == pytest.approx(0.0003) + assert by_unit["wordPolicyUnits"] == 0.0 + assert by_unit["someFutureCounter"] is None + assert by_unit["wordPolicyUnits"] == 0.0 + + +def test_build_tracing_detail_omits_cost_by_unit_when_unpriced_but_keeps_scalar_zero(monkeypatch): + """LIT-5652: without a cost-map entry the spend path still bills 0.0, but the + per-counter stamp must be absent so the rollup records NULL, not $0.""" + monkeypatch.setattr(litellm, "model_cost", {}) + guardrail = BedrockGuardrail(guardrailIdentifier="test-guardrail", guardrailVersion="DRAFT") + + detail = guardrail._build_tracing_detail( + {"action": "NONE", "usage": {"contentPolicyUnits": 5}}, aws_region_name="us-east-1" + ) + + assert detail["guardrail_usage"] == {"contentPolicyUnits": 5} + assert detail["guardrail_cost"] == 0.0 + assert "guardrail_cost_by_unit" not in detail def test_build_tracing_detail_omits_guardrail_usage_when_bedrock_reports_none(): @@ -5115,6 +5146,7 @@ def test_build_tracing_detail_omits_guardrail_usage_when_bedrock_reports_none(): ): assert "guardrail_usage" not in detail assert "guardrail_cost" not in detail + assert "guardrail_cost_by_unit" not in detail @pytest.mark.asyncio @@ -5478,7 +5510,7 @@ async def test_unbuffered_end_of_stream_hook_yields_chunks_before_scan(): scan_index = events.index("scan") chunk_events = [e for e in events if e != "scan"] assert events.count("scan") == 1 - assert [e for e in events[:scan_index] if e != "scan"] == chunk_events[: scan_index] + assert [e for e in events[:scan_index] if e != "scan"] == chunk_events[:scan_index] assert ("chunk", "Hello") in events[:scan_index] assert ("chunk", " world") in events[:scan_index] assert len(chunk_events) == 3 diff --git a/tests/test_litellm/proxy/guardrails/test_usage_endpoints.py b/tests/test_litellm/proxy/guardrails/test_usage_endpoints.py index 1665fa03639..ebb2be6edc2 100644 --- a/tests/test_litellm/proxy/guardrails/test_usage_endpoints.py +++ b/tests/test_litellm/proxy/guardrails/test_usage_endpoints.py @@ -85,7 +85,10 @@ def _units_row( api_key: str = "", usage_unit: str = "contentPolicyUnits", units: int = 1, + cost: float | None = None, + untracked_units: int = 0, ) -> Any: + """cost=None is a row written before the cost column existed (untracked in full).""" r = MagicMock() r.guardrail_id = guardrail_id r.date = date @@ -93,6 +96,8 @@ def _units_row( r.api_key = api_key r.usage_unit = usage_unit r.units = units + r.cost = cost + r.untracked_units = untracked_units return r @@ -279,8 +284,8 @@ async def test_detail_breaks_units_down_by_day_team_and_key(): ) assert resp.usage_units == {"contentPolicyUnits": 3, "topicPolicyUnits": 1} assert [p.model_dump() for p in resp.usage_units_daily] == [ - {"date": "2026-04-24", "units": {"topicPolicyUnits": 1}}, - {"date": "2026-04-25", "units": {"contentPolicyUnits": 3}}, + {"date": "2026-04-24", "units": {"topicPolicyUnits": 1}, "cost": None}, + {"date": "2026-04-25", "units": {"contentPolicyUnits": 3}, "cost": None}, ] assert resp.usage_units_by_team == { "team-a": {"contentPolicyUnits": 2, "topicPolicyUnits": 1}, @@ -311,6 +316,120 @@ async def test_overview_degrades_units_to_empty_when_units_table_is_missing(): row = next(r for r in resp.rows if r.id == "yaml-uuid") assert (row.requestsEvaluated, row.usageUnits) == (4, {}) assert (resp.totalRequests, resp.totalBlocked, resp.totalUsageUnits) == (4, 1, {}) + assert (row.cost, resp.totalCost) == (None, None) + assert (row.untrackedUsageUnits, resp.totalUntrackedUsageUnits) == ({}, {}) + + +@pytest.mark.asyncio +async def test_overview_reports_cost_per_row_and_total_summing_only_tracked_days(): + """LIT-5652: cost rides the units rollup. Rows written before the cost column + carry NULL and rows whose every unit was unpriced carry 0.0 with + untracked_units == units; both must drop out of the sum rather than read as + $0, and a guardrail with only such rows reports None, not 0.0.""" + prisma = _prisma( + find_many=[], + metrics=[_metric("yaml-pii", requests=4, passed=3, blocked=1)], + units=[ + _units_row("yaml-pii", usage_unit="contentPolicyUnits", units=1000, cost=0.15), + _units_row("yaml-pii", team_id="team-a", usage_unit="contentPolicyUnits", units=2000, cost=0.3), + _units_row("yaml-pii", date="2026-04-24", usage_unit="contentPolicyUnits", units=5000, cost=None), + _units_row( + "yaml-pii", date="2026-04-23", usage_unit="topicPolicyUnits", units=9, cost=0.0, untracked_units=9 + ), + _units_row("legacy-guard", usage_unit="topicPolicyUnits", units=7, cost=None), + ], + ) + handler = _config_handler( + _yaml_guardrail(guardrail_id="yaml-uuid", name="yaml-pii"), + _yaml_guardrail(guardrail_id="legacy-uuid", name="legacy-guard"), + ) + p1, p2 = _patches(prisma, handler) + with p1, p2: + resp = await guardrails_usage_overview(start_date=START, end_date=END, user_api_key_dict=ADMIN) + by_id = {r.id: r for r in resp.rows} + assert by_id["yaml-uuid"].cost == pytest.approx(0.45) + assert by_id["legacy-uuid"].cost is None + assert resp.totalCost == pytest.approx(0.45) + + +@pytest.mark.asyncio +async def test_overview_reports_the_units_its_cost_leaves_out_per_row_and_total(): + """A row's cost covers only the units that had a price, so the response must + say exactly which units (per counter) that cost excludes: the row's own + untracked_units, or all of its units when it predates the cost column. A + guardrail whose rows are all priced reports none, one whose rows are all + unpriced reports all of its units, and a mixed row keeps its priced subtotal + while reporting just the unpriced share.""" + prisma = _prisma( + find_many=[], + metrics=[_metric("yaml-pii", requests=4, passed=3, blocked=1)], + units=[ + _units_row("yaml-pii", usage_unit="contentPolicyUnits", units=1000, cost=0.15, untracked_units=200), + _units_row("yaml-pii", date="2026-04-24", usage_unit="contentPolicyUnits", units=5000, cost=None), + _units_row( + "yaml-pii", date="2026-04-24", usage_unit="topicPolicyUnits", units=40, cost=0.0, untracked_units=40 + ), + _units_row("yaml-pii", usage_unit="wordPolicyUnits", units=9, cost=0.0), + _units_row("legacy-guard", usage_unit="topicPolicyUnits", units=7, cost=None), + _units_row("priced-guard", usage_unit="contentPolicyUnits", units=3, cost=0.0003), + ], + ) + handler = _config_handler( + _yaml_guardrail(guardrail_id="yaml-uuid", name="yaml-pii"), + _yaml_guardrail(guardrail_id="legacy-uuid", name="legacy-guard"), + _yaml_guardrail(guardrail_id="priced-uuid", name="priced-guard"), + ) + p1, p2 = _patches(prisma, handler) + with p1, p2: + resp = await guardrails_usage_overview(start_date=START, end_date=END, user_api_key_dict=ADMIN) + by_id = {r.id: r for r in resp.rows} + assert by_id["yaml-uuid"].usageUnits == {"contentPolicyUnits": 6000, "topicPolicyUnits": 40, "wordPolicyUnits": 9} + assert by_id["yaml-uuid"].cost == pytest.approx(0.15) + assert by_id["yaml-uuid"].untrackedUsageUnits == {"contentPolicyUnits": 5200, "topicPolicyUnits": 40} + assert by_id["legacy-uuid"].untrackedUsageUnits == {"topicPolicyUnits": 7} + assert by_id["priced-uuid"].untrackedUsageUnits == {} + assert resp.totalUntrackedUsageUnits == {"contentPolicyUnits": 5200, "topicPolicyUnits": 47} + + +@pytest.mark.asyncio +async def test_detail_breaks_cost_down_by_unit_day_team_and_key(): + """Every cost breakdown keeps the same keys as its units twin so the UI can + render them side by side, with None where that group has no tracked cost.""" + prisma = _prisma( + find_unique=None, + units=[ + _units_row("yaml-pii", date="2026-04-25", team_id="team-a", api_key="hash-1", units=1000, cost=0.15), + _units_row( + "yaml-pii", date="2026-04-25", team_id="", api_key="hash-2", units=200, cost=0.03, untracked_units=50 + ), + _units_row( + "yaml-pii", + date="2026-04-24", + team_id="team-a", + api_key="hash-1", + usage_unit="topicPolicyUnits", + units=10, + cost=None, + ), + ], + ) + handler = _config_handler(_yaml_guardrail()) + p1, p2 = _patches(prisma, handler) + with p1, p2: + resp = await guardrails_usage_detail( + guardrail_id="yaml-1", start_date=START, end_date=END, user_api_key_dict=ADMIN + ) + assert resp.cost == pytest.approx(0.18) + assert resp.cost_by_unit == {"contentPolicyUnits": pytest.approx(0.18), "topicPolicyUnits": None} + assert [p.model_dump() for p in resp.usage_units_daily] == [ + {"date": "2026-04-24", "units": {"topicPolicyUnits": 10}, "cost": None}, + {"date": "2026-04-25", "units": {"contentPolicyUnits": 1200}, "cost": pytest.approx(0.18)}, + ] + assert resp.cost_by_team == {"team-a": pytest.approx(0.15), "": pytest.approx(0.03)} + assert resp.cost_by_key == {"hash-1": pytest.approx(0.15), "hash-2": pytest.approx(0.03)} + assert resp.cost_by_team.keys() == resp.usage_units_by_team.keys() + assert resp.cost_by_key.keys() == resp.usage_units_by_key.keys() + assert resp.untracked_usage_units == {"contentPolicyUnits": 50, "topicPolicyUnits": 10} @pytest.mark.asyncio @@ -330,6 +449,8 @@ async def test_detail_degrades_units_to_empty_when_units_table_is_missing(): {}, {}, ) + assert (resp.cost, resp.cost_by_unit, resp.cost_by_team, resp.cost_by_key) == (None, {}, {}, {}) + assert resp.untracked_usage_units == {} # ---- logs ------------------------------------------------------------------- @@ -411,6 +532,29 @@ async def test_detail_rejects_reversed_dates(): assert exc.value.status_code == 400 +@pytest.mark.asyncio +async def test_policies_overview_returns_a_full_row_and_totals(): + """Regression: the policies overview shares the guardrail response model, so + every field added there (usage units, cost, untracked units) must be filled + here too or the endpoint 500s on model validation.""" + policy = MagicMock(spec=["policy_id", "policy_name"]) + policy.policy_id = "pol-1" + policy.policy_name = "block-pii" + metric = _metric("pol-1", requests=10, passed=8, blocked=2) + metric.policy_id = "pol-1" + prisma = _prisma() + prisma.db.litellm_policytable.find_many = AsyncMock(return_value=[policy]) + prisma.db.litellm_dailypolicymetrics.find_many = AsyncMock(return_value=[metric]) + p1, p2 = _patches(prisma, _config_handler()) + with p1, p2: + resp = await policies_usage_overview(start_date=START, end_date=END, user_api_key_dict=ADMIN) + row = next(r for r in resp.rows if r.id == "pol-1") + assert (row.name, row.type, row.requestsEvaluated, row.failRate) == ("block-pii", "Policy", 10, 20.0) + assert (row.usageUnits, row.cost, row.untrackedUsageUnits) == ({}, None, {}) + assert (resp.totalRequests, resp.totalBlocked, resp.passRate) == (10, 2, 80.0) + assert (resp.totalUsageUnits, resp.totalCost, resp.totalUntrackedUsageUnits) == ({}, None, {}) + + @pytest.mark.asyncio async def test_policies_overview_rejects_range_over_max_days(): prisma = _prisma() diff --git a/tests/test_litellm/proxy/guardrails/test_usage_tracking.py b/tests/test_litellm/proxy/guardrails/test_usage_tracking.py index 6da121703d7..ae360b281cb 100644 --- a/tests/test_litellm/proxy/guardrails/test_usage_tracking.py +++ b/tests/test_litellm/proxy/guardrails/test_usage_tracking.py @@ -30,6 +30,8 @@ def _payload( api_key: str = "hashed-key-1", usage: dict[str, Any] | None = None, guardrail_status: str = "success", + cost_by_unit: dict[str, Any] | None = None, + cost_in_spend: bool | None = None, ) -> dict[str, Any]: entry: dict[str, Any] = { "guardrail_id": "bedrock-guard", @@ -37,6 +39,10 @@ def _payload( } if usage is not None: entry["guardrail_usage"] = usage + if cost_by_unit is not None: + entry["guardrail_cost_by_unit"] = cost_by_unit + if cost_in_spend is not None: + entry["guardrail_cost_in_spend"] = cost_in_spend return { "request_id": request_id, "startTime": datetime(2026, 8, 17, 12, 0, tzinfo=timezone.utc), @@ -58,6 +64,19 @@ def _units_upserts(prisma: MagicMock) -> dict[tuple, int]: return out +def _cost_upserts(prisma: MagicMock) -> dict[str, tuple[float, int]]: + """usage_unit -> (cost, untracked_units) written on create; the update path must increment by the same.""" + calls = prisma.db.litellm_dailyguardrailusageunits.upsert.call_args_list + out: dict[str, tuple[float, int]] = {} + for c in calls: + create = c.kwargs["data"]["create"] + update = c.kwargs["data"]["update"] + assert update["cost"] == {"increment": create["cost"]} + assert update["untracked_units"] == {"increment": create["untracked_units"]} + out[create["usage_unit"]] = (create["cost"], create["untracked_units"]) + return out + + @pytest.mark.asyncio async def test_usage_units_rolled_up_by_guardrail_team_key_and_date(): """ @@ -181,7 +200,9 @@ async def test_retry_exhausted_rows_are_requeued_and_land_on_the_next_flush(): down, [_payload("r1", usage={"topicPolicyUnits": 2})], sleep=sleep, pending=pending ) - assert dict(pending.units) == {("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "topicPolicyUnits"): 2} + assert dict(pending.units) == { + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "topicPolicyUnits"): (2, 0.0, 2) + } recovered = _prisma() await process_spend_logs_guardrail_usage( @@ -320,3 +341,152 @@ async def test_payload_without_request_id_is_skipped_like_the_metrics_path(): ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "topicPolicyUnits"): 1, } assert prisma.db.litellm_dailyguardrailmetrics.upsert.call_args.kwargs["data"]["create"]["requests_evaluated"] == 1 + + +@pytest.mark.asyncio +async def test_cost_rolled_up_per_counter_alongside_units(): + """LIT-5652: the hook's per-counter cost lands on the same daily row as the + units it priced, summed across payloads exactly like the units are, and the + update path increments it so a second flush on the same day keeps adding.""" + prisma = _prisma() + logs = [ + _payload( + "r1", + usage={"contentPolicyUnits": 1000, "wordPolicyUnits": 50}, + cost_by_unit={"contentPolicyUnits": 0.15, "wordPolicyUnits": 0.0}, + ), + _payload( + "r2", + usage={"contentPolicyUnits": 2000, "wordPolicyUnits": 10}, + cost_by_unit={"contentPolicyUnits": 0.3, "wordPolicyUnits": 0.0}, + ), + ] + + await process_spend_logs_guardrail_usage(prisma, logs) + + assert _units_upserts(prisma) == { + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "contentPolicyUnits"): 3000, + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "wordPolicyUnits"): 60, + } + costs = _cost_upserts(prisma) + assert costs["contentPolicyUnits"] == (pytest.approx(0.45), 0) + assert costs["wordPolicyUnits"] == (0.0, 0) + + +@pytest.mark.asyncio +async def test_counter_the_hook_could_not_price_is_stored_as_untracked_units_not_free(): + """A counter the cost map does not list arrives stamped as None. Its units + must land in untracked_units with no cost, so the row never reads as free, + while the priced counter on the same request keeps its cost.""" + prisma = _prisma() + logs = [ + _payload( + "r1", + usage={"contentPolicyUnits": 1000, "someFutureCounter": 3}, + cost_by_unit={"contentPolicyUnits": 0.15, "someFutureCounter": None}, + ) + ] + + await process_spend_logs_guardrail_usage(prisma, logs) + + assert _units_upserts(prisma) == { + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "contentPolicyUnits"): 1000, + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "someFutureCounter"): 3, + } + costs = _cost_upserts(prisma) + assert costs["contentPolicyUnits"] == (pytest.approx(0.15), 0) + assert costs["someFutureCounter"] == (0.0, 3) + + +@pytest.mark.asyncio +async def test_mixed_priced_and_unpriced_increments_keep_the_subtotal_and_count_the_rest_untracked(): + """Priced and unpriced increments on the same row (a hook without pricing, + a pre-upgrade proxy in a mixed fleet) must keep the priced subtotal and + count exactly the unpriced units as untracked. Nulling the cost would throw + away a known number; keeping it alone would look exact while understating.""" + prisma = _prisma() + logs = [ + _payload("r1", usage={"contentPolicyUnits": 1000}, cost_by_unit={"contentPolicyUnits": 0.15}), + _payload("r2", usage={"contentPolicyUnits": 700}), + _payload("r3", usage={"contentPolicyUnits": 300}, cost_by_unit={"contentPolicyUnits": None}), + ] + + await process_spend_logs_guardrail_usage(prisma, logs) + + assert _units_upserts(prisma) == { + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "contentPolicyUnits"): 2000, + } + assert _cost_upserts(prisma) == {"contentPolicyUnits": (pytest.approx(0.15), 1000)} + + +@pytest.mark.asyncio +async def test_report_only_and_forged_costs_are_not_rolled_up_but_units_are(): + """guardrail_cost_in_spend=False (Azure Prompt Shield) keeps its cost out of + spend, so the rollup must not record it either or the dashboard would show + a number the budget never charged. A negative or non-finite per-counter cost + is treated the same way rather than subtracting from the day.""" + prisma = _prisma() + logs = [ + _payload("r1", usage={"text_records": 3}, cost_by_unit={"text_records": 0.5}, cost_in_spend=False), + _payload("r2", usage={"contentPolicyUnits": 10}, cost_by_unit={"contentPolicyUnits": -0.5}), + _payload("r3", usage={"topicPolicyUnits": 10}, cost_by_unit={"topicPolicyUnits": float("inf")}), + ] + + await process_spend_logs_guardrail_usage(prisma, logs) + + assert _units_upserts(prisma) == { + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "text_records"): 3, + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "contentPolicyUnits"): 10, + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "topicPolicyUnits"): 10, + } + assert _cost_upserts(prisma) == { + "text_records": (0.0, 3), + "contentPolicyUnits": (0.0, 10), + "topicPolicyUnits": (0.0, 10), + } + + +@pytest.mark.asyncio +async def test_requeued_cost_is_added_to_the_next_flush(): + """Cost and untracked units must survive the connection-error requeue the + same way units do, or a DB blip would silently drop dollars (or the record + that some units had no price) while keeping the units themselves.""" + pending = PendingRollups() + down = _prisma() + down.db.litellm_dailyguardrailmetrics.upsert.side_effect = httpx.ConnectError("db down") + down.db.litellm_dailyguardrailusageunits.upsert.side_effect = httpx.ConnectError("db down") + sleep, _ = _fake_sleep() + + await process_spend_logs_guardrail_usage( + down, + [ + _payload( + "r1", + usage={"contentPolicyUnits": 1000, "someFutureCounter": 3}, + cost_by_unit={"contentPolicyUnits": 0.15, "someFutureCounter": None}, + ) + ], + sleep=sleep, + pending=pending, + ) + recovered = _prisma() + await process_spend_logs_guardrail_usage( + recovered, + [ + _payload( + "r2", + usage={"contentPolicyUnits": 2000, "someFutureCounter": 4}, + cost_by_unit={"contentPolicyUnits": 0.3, "someFutureCounter": None}, + ) + ], + sleep=sleep, + pending=pending, + ) + + assert _units_upserts(recovered) == { + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "contentPolicyUnits"): 3000, + ("bedrock-guard", "2026-08-17", "team-a", "hashed-key-1", "someFutureCounter"): 7, + } + costs = _cost_upserts(recovered) + assert costs["contentPolicyUnits"] == (pytest.approx(0.45), 0) + assert costs["someFutureCounter"] == (0.0, 7) diff --git a/type-discipline-budget.json b/type-discipline-budget.json index 094b9749d98..8589a9451cf 100644 --- a/type-discipline-budget.json +++ b/type-discipline-budget.json @@ -3,7 +3,7 @@ "limit": 22328 }, "LIT002": { - "limit": 26750 + "limit": 26748 }, "LIT003": { "limit": 261 diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index fe1cc2772b4..d94b5425ba9 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -38347,6 +38347,20 @@ export interface components { avgLatency: number | null; /** Avgscore */ avgScore: number | null; + /** Cost */ + cost: number | null; + /** Cost By Key */ + cost_by_key: { + [key: string]: number | null; + }; + /** Cost By Team */ + cost_by_team: { + [key: string]: number | null; + }; + /** Cost By Unit */ + cost_by_unit: { + [key: string]: number | null; + }; /** Description */ description: string | null; /** Failrate */ @@ -38367,6 +38381,10 @@ export interface components { trend: string; /** Type */ type: string; + /** Untracked Usage Units */ + untracked_usage_units: { + [key: string]: number; + }; /** Usage Units */ usage_units: { [key: string]: number; @@ -38428,8 +38446,14 @@ export interface components { rows: components["schemas"]["UsageOverviewRow"][]; /** Totalblocked */ totalBlocked: number; + /** Totalcost */ + totalCost: number | null; /** Totalrequests */ totalRequests: number; + /** Totaluntrackedusageunits */ + totalUntrackedUsageUnits: { + [key: string]: number; + }; /** Totalusageunits */ totalUsageUnits: { [key: string]: number; @@ -38441,6 +38465,11 @@ export interface components { avgLatency: number | null; /** Avgscore */ avgScore: number | null; + /** + * Cost + * @description USD for the priced share of usageUnits over the window; null when no unit was priced + */ + cost: number | null; /** Failrate */ failRate: number; /** Id */ @@ -38457,6 +38486,13 @@ export interface components { trend: string; /** Type */ type: string; + /** + * Untrackedusageunits + * @description The share of usageUnits that cost leaves out: units recorded with no known price, per counter + */ + untrackedUsageUnits: { + [key: string]: number; + }; /** Usageunits */ usageUnits: { [key: string]: number; @@ -38464,6 +38500,8 @@ export interface components { }; /** UsageUnitsDailyPoint */ UsageUnitsDailyPoint: { + /** Cost */ + cost: number | null; /** Date */ date: string; /** Units */