diff --git a/litellm/router_strategy/complexity_router/README.md b/litellm/router_strategy/complexity_router/README.md index 4aa342ea59f..801d5149a24 100644 --- a/litellm/router_strategy/complexity_router/README.md +++ b/litellm/router_strategy/complexity_router/README.md @@ -68,6 +68,117 @@ still resolve to a deployment in `model_list`; this configuration does not creat - abc ``` +### Capability forecasting + +Set `classifier_type: capability` to use +[NVIDIA NeMo Switchyard's packaged capability classifier](https://github.com/NVIDIA-NeMo/Switchyard/blob/main/crates/libsy/src/prompts/capability-classifier/prompt.md). +The classifier forecasts the probability that an efficient model completes +the whole task, identifies the capability-card boundary that applies, and leaves the +route choice to a deterministic threshold policy + +```yaml +model_list: + - model_name: smart-router + litellm_params: + model: auto_router/complexity_router + complexity_router_config: + classifier_type: capability + classifier_llm_config: + model: classifier-model + capability_classifier_config: + efficient_tier: SIMPLE + capable_tier: REASONING + base_threshold: 0.5 + threshold_step: 0.1 + tiers: + SIMPLE: + - efficient-model-a + - efficient-model-b + REASONING: capable-model +``` + +The structured classifier verdict contains `crux`, `primary_rule`, +`capability_boundary`, and `p_solve`. The policy computes the required solve +probability as follows + +- `supported`: `base_threshold` +- `uncertain` or `unmatched`: `base_threshold + threshold_step` +- `unsupported`: `base_threshold + 2 * threshold_step` + +The efficient tier is selected when `p_solve` is greater than or equal to the +adjusted threshold. Otherwise the capable tier is selected. A malformed, +inconsistent, empty, or unavailable verdict always fails closed to the capable +tier. `base_threshold` is required, `threshold_step` defaults to `0`, and their +maximum adjusted threshold must not exceed `1` + +The classifier receives the packaged Switchyard system prompt, the opening user +task, and the latest user follow-up when present. Caller system messages, +assistant turns, and intermediate tool results are not sent. The classifier call +uses strict JSON Schema output and the existing classifier timeout, circuit +breaker, attribution, redaction, reasoning-effort, and optional vision settings + +`efficient_tier` and `capable_tier` name built-in complexity tiers with configured +model pools. The forecast still makes one binary quality decision, while the +ordinary tier pool may contain multiple equivalent deployments. Session affinity, +keyword overrides, plan-mode floors, modality checks, and other post-classification +complexity-router controls continue to apply + +Routing decisions record the adjusted threshold and the complete valid forecast: +`classifier_p_solve`, `classifier_capability_boundary`, `classifier_primary_rule`, +and `classifier_crux`. Prompt redaction removes `classifier_crux` while retaining +the derived fields needed to audit the decision + +#### Calibrating solve probabilities + +Supply a fitted monotone logit calibration under `capability_classifier_config` +to transform the forecast before applying the threshold. Calibration is opt-in; +without it the router uses the raw probability. Fit coefficients on benchmark +outcomes from separate training repositories, select thresholds on a validation +split, and report quality and cost on an untouched evaluation split + +```yaml +capability_classifier_config: + efficient_tier: SIMPLE + capable_tier: REASONING + base_threshold: 0.66 + threshold_step: 0 + max_output_tokens: 512 + response_format: json_object + calibration: + version: your-benchmark-artifact-v1 + slope: 1.0 + intercept: 0.0 +``` + +The example coefficients are an identity mapping, not a trained calibration. +The mapping is `sigmoid(slope * logit(clip(p_solve, 1e-6, 1-1e-6)) + intercept)`. +The slope must be nonnegative, so calibration cannot improve ranking. It can +make probabilities more accurate and thresholds easier to interpret. The version +is recorded for auditing; the router does not check whether an artifact matches +the judge, capability card, efficient solver, or agent harness. Operators must +keep those aligned and refit when they change + +Logs retain `classifier_p_solve` and add `classifier_calibrated_p_solve` and +`classifier_calibration_version`. `classifier_threshold` is compared to the +calibrated probability. Invalid verdicts still route to the capable tier + +`response_format` defaults to `json_schema`. For endpoints that support JSON +objects but not strict schemas, `json_object` appends the same schema to the +unchanged capability prompt and retains strict local validation. Set +`classifier_llm_config.timeout_ms` to cover the measured judge latency; a local +judge may need longer than the default 3000 ms. `max_output_tokens` still defaults +to 4096; 512 is an explicit benchmark setting for a short, non-reasoning judge + +For a controlled whole-task benchmark, use `adaptive: false`, +`session_affinity: true`, and a unique session ID for every task and policy arm. +Disable keyword, plan-mode, housekeeping, and other optional overrides when +measuring only the capability policy. When adaptive selection is enabled, it +cannot select below the capability decision, including a capable-tier fallback + +Configure capability forecasting through YAML or the model-management API. +The dashboard preserves its classifier and calibration on an untouched save; +it does not provide a capability-card editor + ### Heuristic v2 Set `classifier_type: heuristic_v2` to classify with the bundled calibrated diff --git a/litellm/router_strategy/complexity_router/__init__.py b/litellm/router_strategy/complexity_router/__init__.py index fa21f2eee10..5447e4e0f6d 100644 --- a/litellm/router_strategy/complexity_router/__init__.py +++ b/litellm/router_strategy/complexity_router/__init__.py @@ -16,6 +16,8 @@ from litellm.router_strategy.complexity_router.complexity_router import ( from litellm.router_strategy.complexity_router.config import ( DEFAULT_CLASSIFIER_CONTEXT_WINDOW_SIZE, DEFAULT_COMPLEXITY_CONFIG, + CapabilityCalibrationConfig, + CapabilityClassifierConfig, ClassificationRubric, ComplexityRouterConfig, ComplexityTier, @@ -28,6 +30,8 @@ from litellm.router_strategy.complexity_router.config import ( __all__ = [ "DEFAULT_CLASSIFIER_CONTEXT_WINDOW_SIZE", "DEFAULT_COMPLEXITY_CONFIG", + "CapabilityCalibrationConfig", + "CapabilityClassifierConfig", "ClassificationRubric", "ComplexityRouter", "ComplexityRouterConfig", diff --git a/litellm/router_strategy/complexity_router/capability_classifier.py b/litellm/router_strategy/complexity_router/capability_classifier.py new file mode 100644 index 00000000000..66ed9c36ed8 --- /dev/null +++ b/litellm/router_strategy/complexity_router/capability_classifier.py @@ -0,0 +1,211 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 + +"""Capability forecast contract and routing policy adapted from NVIDIA NeMo Switchyard.""" + +import json +from collections.abc import Mapping +from sys import float_info +from types import MappingProxyType +from typing import Final, Literal, NamedTuple, TypeAlias + +from pydantic import BaseModel, ConfigDict, Field, StrictFloat, TypeAdapter, model_validator + +CapabilityBoundary: TypeAlias = Literal["supported", "uncertain", "unsupported", "unmatched"] +CapabilityRule: TypeAlias = Literal[ + "SUP-1", + "SUP-2", + "SUP-3", + "SUP-4", + "SUP-5", + "UNC-1", + "UNC-2", + "LIM-1", + "LIM-2", + "none", +] + +CAPABILITY_CLASSIFIER_SYSTEM_PROMPT: Final = """You are a task-level probability forecaster for a model router. You receive the +task's opening instruction and, when present, its latest user follow-up, plus +the qualitative capability card below. + +Forecast one binary event: + +SUCCESS means that the efficient agent completes the whole task correctly on +one fresh run under the actual harness, tools, and budget, as judged by the +final verifier. FAILURE means any other outcome. The two outcomes are +exhaustive. + +Use only evidence in the instruction and the capability card. Do not assume +hidden repository state, unmentioned tools, validators, documentation, access, +or future work habits. Do not invent empirical counts, success rates, or base +rates. The capability card is qualitative evidence, not a measured prior. + +# Assessment procedure + +1. State the crux: the hardest material requirement for whole-task success. +2. Select the one capability rule that best describes the crux. Use + primary_rule=none and capability_boundary=unmatched when no rule applies. + Rule ids are opaque labels. Do not infer a boundary from an id's spelling. +3. Privately identify the strongest instruction-visible reasons for SUCCESS + and FAILURE, then imagine the most likely concrete failure. +4. Privately consider material unknowns. Missing information should limit + extreme estimates, but it is not evidence that p_solve must equal 0.50. +5. Estimate p_solve last. It is the probability of whole-task SUCCESS, not + confidence in this assessment, a route recommendation, or a cost judgment. + +Interpret probabilities as natural frequencies. If p_solve is 0.70 for 100 +comparable fresh runs, about 70 should succeed and 30 should fail. Use the full +range when justified. Reserve 0.00 and 1.00 for outcomes that are logically +impossible or certain under the visible contract. Supported does not mean 1.00, +and unsupported does not mean 0.00. The downstream routing threshold is not +part of this forecast. + +# Efficient-agent capability card + +The route verbs in this source card are inherited qualitative descriptions. +They do not ask you to output a route and do not assign a fixed probability to +any boundary. + +- SUP-1 [supported]: Route to the Efficient model when the task provides a complete output contract and a deterministic local validator that covers the material requirements. +- SUP-2 [supported]: Route to the Efficient model when all required inputs are available, the target environment can be inspected, and correctness can be verified end-to-end without inaccessible external state. +- SUP-3 [supported]: Route to the Efficient model when mathematical behavior, interfaces, shapes, data types, tolerances, and performance requirements are explicit and exercised by a representative harness. +- SUP-4 [supported]: Route to the Efficient model when the required mechanism is identified, the relevant search space is bounded, and the success condition is executable. Do not infer this rule merely from the task's technical domain. +- SUP-5 [supported]: Route to the Efficient model when reconstruction or behavioral reproduction is constrained by an executable reference, parser, format specification, or checker strong enough to distinguish correct from merely plausible output. +- UNC-1 [uncertain]: Treat the route as uncertain when multiple reasonable interpretations of preprocessing, representation, indexing, naming, or output placement would produce different results and neither the instructions nor a validator resolve the choice. +- UNC-2 [uncertain]: Treat the route as uncertain when success requires finding every relevant item across heterogeneous inputs or environment state, but the task does not define the search boundary or provide a completeness check. +- LIM-1 [unsupported]: Prefer the Capable model when correctness depends primarily on extracting precise information from noisy visual, temporal, or rendered media and no machine-checkable extraction or replay mechanism is available. +- LIM-2 [unsupported]: Prefer the Capable model when success depends on reproducing undocumented reference behavior, hidden intermediate state, or an unknown configuration, and small deviations fail despite satisfying the visible specification. + +# Output + +Return exactly one JSON object matching the response schema supplied with the +request. Do not include markdown or commentary. + +p_solve must be between 0.00 and 1.00. p_fail is exactly 1.00 - p_solve and +must not be emitted separately. Do not output recommended_route, confidence, +abstain, counts, task totals, empirical rates, or any other field.""" + +_BOUNDARY_STEPS: Final = MappingProxyType( + { + "supported": 0, + "uncertain": 1, + "unmatched": 1, + "unsupported": 2, + } +) + +_RULE_BOUNDARIES: Final = MappingProxyType( + { + "SUP-1": "supported", + "SUP-2": "supported", + "SUP-3": "supported", + "SUP-4": "supported", + "SUP-5": "supported", + "UNC-1": "uncertain", + "UNC-2": "uncertain", + "LIM-1": "unsupported", + "LIM-2": "unsupported", + "none": "unmatched", + } +) + + +class CapabilityClassifierVerdict(BaseModel): + """Strict structured verdict returned by the capability forecaster.""" + + model_config = ConfigDict(extra="forbid", frozen=True) + + crux: str = Field(min_length=1) + primary_rule: CapabilityRule + capability_boundary: CapabilityBoundary + p_solve: StrictFloat = Field(ge=0.0, le=1.0) + + @model_validator(mode="after") + def _validate_rule_boundary_pair(self) -> "CapabilityClassifierVerdict": + if not self.crux.strip(): + raise ValueError("crux must contain non-whitespace text") + expected: Final = _RULE_BOUNDARIES[self.primary_rule] + if self.capability_boundary != expected: + raise ValueError( + f"primary_rule {self.primary_rule!r} requires capability_boundary {expected!r}, " + f"got {self.capability_boundary!r}" + ) + return self + + def routing_threshold(self, base_threshold: float, threshold_step: float) -> float: + """Required efficient-model solve probability for this boundary.""" + return base_threshold + _BOUNDARY_STEPS[self.capability_boundary] * threshold_step + + def meets_routing_threshold(self, threshold: float) -> bool: + """Inclusive comparison with Switchyard's one-epsilon rounding guard.""" + return self.p_solve >= threshold or abs(threshold - self.p_solve) <= float_info.epsilon + + +class CapabilityClassifierForecast(NamedTuple): + verdict: CapabilityClassifierVerdict + threshold: float + p_solve: float + calibration_version: str | None + + def meets_routing_threshold(self) -> bool: + return self.p_solve >= self.threshold or abs(self.threshold - self.p_solve) <= float_info.epsilon + + +_CAPABILITY_CLASSIFIER_RESPONSE_FORMAT_JSON: Final = """{ + "type": "json_schema", + "json_schema": { + "name": "CapabilityClassifierDecision", + "strict": true, + "schema": { + "type": "object", + "additionalProperties": false, + "required": ["crux", "primary_rule", "capability_boundary", "p_solve"], + "properties": { + "crux": {"type": "string", "minLength": 1}, + "primary_rule": { + "type": "string", + "enum": ["SUP-1", "SUP-2", "SUP-3", "SUP-4", "SUP-5", "UNC-1", "UNC-2", "LIM-1", "LIM-2", "none"] + }, + "capability_boundary": { + "type": "string", + "enum": ["supported", "uncertain", "unsupported", "unmatched"] + }, + "p_solve": {"type": "number", "minimum": 0.0, "maximum": 1.0} + } + } + } +}""" + +_RESPONSE_FORMAT_ADAPTER: Final = TypeAdapter(Mapping[str, object]) + + +def capability_classifier_response_format( + mode: Literal["json_schema", "json_object"] = "json_schema", +) -> Mapping[str, object]: + """Fresh copy of Switchyard's packaged strict JSON Schema wrapper.""" + return ( + _RESPONSE_FORMAT_ADAPTER.validate_json('{"type": "json_object"}') + if mode == "json_object" + else _RESPONSE_FORMAT_ADAPTER.validate_json(_CAPABILITY_CLASSIFIER_RESPONSE_FORMAT_JSON) + ) + + +def capability_classifier_system_prompt(mode: Literal["json_schema", "json_object"]) -> str: + if mode == "json_schema": + return CAPABILITY_CLASSIFIER_SYSTEM_PROMPT + wrapper: Final = _RESPONSE_FORMAT_ADAPTER.validate_python(capability_classifier_response_format()["json_schema"]) + return ( + CAPABILITY_CLASSIFIER_SYSTEM_PROMPT + + "\n\nReturn exactly one JSON object matching this JSON Schema:\n" + + json.dumps(wrapper["schema"], indent=2, sort_keys=True) + ) + + +def parse_capability_classifier_verdict(content: str) -> CapabilityClassifierVerdict: + """Parse raw JSON or the fenced JSON shape tolerated by Switchyard.""" + text: Final = content.strip() + if not text.startswith("```"): + return CapabilityClassifierVerdict.model_validate_json(text) + unfenced: Final = text.removeprefix("```").removeprefix("json").lstrip("\n\r") + return CapabilityClassifierVerdict.model_validate_json(unfenced.removesuffix("```").strip()) diff --git a/litellm/router_strategy/complexity_router/complexity_router.py b/litellm/router_strategy/complexity_router/complexity_router.py index 1000f0f479f..d20abefbb2a 100644 --- a/litellm/router_strategy/complexity_router/complexity_router.py +++ b/litellm/router_strategy/complexity_router/complexity_router.py @@ -5,8 +5,9 @@ A rule-based routing strategy that uses weighted scoring across multiple dimensi to classify requests by complexity and route them to appropriate models. By default, scoring is local (regex/keyword-based) with no external API calls and <1ms -latency. Optionally, classifier_type="llm" routes classification through a configured -model instead, trading that latency/cost guarantee for potentially better accuracy. +latency. Optionally, classifier_type="llm" selects a tier through a configured model, +while classifier_type="capability" forecasts efficient-model success and applies a +Switchyard-compatible threshold policy. keyword_tier_rules (lexical or, with semantic_keyword_matching, embedding-based) are evaluated before either classification strategy and force a tier outright when matched. @@ -73,6 +74,12 @@ from litellm.types.utils import ( StandardLoggingRoutingDecisionTierBoundaries, ) +from .capability_classifier import ( + CapabilityClassifierForecast, + capability_classifier_response_format, + capability_classifier_system_prompt, + parse_capability_classifier_verdict, +) from .classification_rubrics import BUSINESS_TIER_CRITERIA, calibration_examples_section from .config import ( CALIBRATION_EXAMPLES_HEADING, @@ -994,20 +1001,49 @@ class ClassificationOutcome(NamedTuple): "heuristic_v2", "reasoning_override", "llm_classifier", + "capability_classifier", "heuristic_first_short_circuit", "hybrid_short_circuit", "housekeeping", "classifier_plugin", "classifier_fallback", + "capability_classifier_fallback", "default_model_fallback", ] classifier_cost: float | None = None + capability_forecast: CapabilityClassifierForecast | None = None def _with_signal(outcome: ClassificationOutcome, signal: str | None) -> ClassificationOutcome: return outcome if signal is None else outcome._replace(signals=(*outcome.signals, signal)) +def _with_capability_forecast( + decision: StandardLoggingRoutingDecision, outcome: ClassificationOutcome +) -> StandardLoggingRoutingDecision: + """Attach the validated capability verdict and applied threshold to its decision record.""" + forecast: Final = outcome.capability_forecast + if forecast is None: + return decision + verdict: Final = forecast.verdict + enriched: Final[StandardLoggingRoutingDecision] = { # mutable-ok: routing decisions are JSON TypedDict records + **decision, + "classifier_crux": verdict.crux, + "classifier_primary_rule": verdict.primary_rule, + "classifier_capability_boundary": verdict.capability_boundary, + "classifier_p_solve": verdict.p_solve, + "classifier_threshold": forecast.threshold, + } + if forecast.calibration_version is None: + return enriched + calibrated: Final[StandardLoggingRoutingDecision] = { + **enriched, + "classifier_calibrated_p_solve": forecast.p_solve, + "classifier_calibration_version": forecast.calibration_version, + } + return calibrated + + class _ClassifierCircuitBreaker: """Process-local timeout breaker for one complexity-router classifier. @@ -1276,8 +1312,15 @@ class ComplexityRouter(CustomLogger): self._classifier_system_prompt: str | None = ( self._build_classifier_system_prompt() if llm_classifier_configured else None ) + capability_config: Final = self.config.capability_classifier_config self._classifier_response_format: Mapping[str, object] | None = ( - type_to_response_format_param(_tier_classification_model(self.config.classifier_wire_labels())) + ( + capability_classifier_response_format( + capability_config.response_format if capability_config is not None else "json_schema" + ) + if self.config.classifier_type == "capability" + else type_to_response_format_param(_tier_classification_model(self.config.classifier_wire_labels())) + ) if llm_classifier_configured else None ) @@ -1303,6 +1346,11 @@ class ComplexityRouter(CustomLogger): llm_config: Final = self.config.classifier_llm_config if llm_config is None: raise ValueError("classifier_llm_config is not set") + if self.config.classifier_type == "capability": + capability: Final = self.config.capability_classifier_config + return capability_classifier_system_prompt( + capability.response_format if capability is not None else "json_schema" + ) definitions: Final = self.config.tier_definitions if definitions is not None: return custom_tier_classification_prompt( @@ -1720,6 +1768,8 @@ class ComplexityRouter(CustomLogger): return await self._classify_heuristic_first(prompt, system_prompt, request_kwargs, messages) if self.config.classifier_type == "hybrid" and self.config.classifier_llm_config is not None: return await self._classify_hybrid(prompt, system_prompt, request_kwargs, messages) + if self.config.classifier_type == "capability" and self.config.classifier_llm_config is not None: + return await self._capability_classifier_outcome(prompt, request_kwargs, messages) if self.config.classifier_type != "llm" or self.config.classifier_llm_config is None: tier, score, signals, cause = self._score_and_classify(prompt, system_prompt) return ClassificationOutcome(tier=tier, score=score, signals=signals, cause=cause) @@ -1831,6 +1881,66 @@ class ComplexityRouter(CustomLogger): ) ) + async def _capability_classifier_outcome( + self, + prompt: str, + request_kwargs: Mapping[str, object] | None, + messages: Sequence[Mapping[str, object]] | None, + ) -> ClassificationOutcome: + """Forecast efficient-tier success, then apply the deterministic boundary policy.""" + breaker: Final = self._classifier_circuit_breaker + permit: Final = breaker.acquire_permit() if breaker is not None else None + if breaker is not None and permit is None: + return self._capability_classifier_failure_outcome( + "capability classifier circuit is open", signal=_CLASSIFIER_CIRCUIT_OPEN_SIGNAL + ) + try: + tier, classifier_cost, forecast = await self._classify_with_capability_llm(prompt, request_kwargs, messages) + if breaker is not None and permit is not None: + breaker.record_success(permit) + return ClassificationOutcome( + tier=tier, + score=None, + signals=( + f"capability-boundary:{forecast.verdict.capability_boundary}", + f"capability-rule:{forecast.verdict.primary_rule}", + ), + cause="capability_classifier", + classifier_cost=classifier_cost, + capability_forecast=forecast, + ) + except asyncio.CancelledError: + if breaker is not None and permit is not None: + breaker.record_failure(permit, is_timeout=False) + raise + except Exception as e: # noqa: BLE001 -- every unavailable or invalid judge verdict must fail closed + if breaker is not None and permit is not None: + breaker.record_failure(permit, is_timeout=_is_classifier_timeout(e)) + return self._capability_classifier_failure_outcome(f"capability classifier failed ({e})") + + def _capability_classifier_failure_outcome(self, reason: str, signal: str | None = None) -> ClassificationOutcome: + """Fail closed to the configured capable tier without consulting another taxonomy.""" + capability: Final = self.config.capability_classifier_config + if capability is None: + raise ValueError("capability_classifier_config is not set") + verbose_router_logger.warning( + "ComplexityRouter: %s, routing to capable_tier %s", reason, capability.capable_tier + ) + signals: Final = ( + ("capability-classifier-fallback",) + if signal is None + else ( + "capability-classifier-fallback", + signal, + ) + ) + return ClassificationOutcome( + tier=ComplexityTier(capability.capable_tier), + score=None, + signals=signals, + cause="capability_classifier_fallback", + ) + async def _llm_classifier_outcome( self, prompt: str, @@ -2066,13 +2176,6 @@ class ComplexityRouter(CustomLogger): label_roles=include_assistant, ) - request_metadata = (request_kwargs or {}).get("litellm_metadata") or (request_kwargs or {}).get("metadata") - metadata: Final = { # mutable-ok: SDK metadata kwarg is enriched by the request pipeline - **forwarded_internal_call_metadata(request_metadata, AUTOROUTER_CLASSIFIER_CALL_ORIGIN), - INTERNAL_CALL_ORIGIN_METADATA_KEY: AUTOROUTER_CLASSIFIER_CALL_ORIGIN, - } - turn_off_message_logging: Final = _effective_turn_off_message_logging(request_kwargs) - image_parts: Final = self._classifier_image_parts(messages) user_content: Final[str | Sequence[ChatCompletionTextObject | ChatCompletionImageObject]] = ( [ # mutable-ok: SDK request payload content list is built once @@ -2086,18 +2189,125 @@ class ComplexityRouter(CustomLogger): {"role": "system", "content": classifier_system_prompt}, {"role": "user", "content": user_content}, ] - response_format: Final = classifier_response_format - classifier_call_params: Mapping[str, str] = EMPTY_MAPPING - if llm_config.reasoning_effort is not None: - classifier_call_params = MappingProxyType({"reasoning_effort": llm_config.reasoning_effort}) + content, classifier_cost = await self._call_classifier_model( + messages_for_call, request_kwargs, encrypted_task=encrypted_task + ) + raw_tier: Final = _LabeledTierClassification.model_validate_json(content).tier + tier: Final = self.config.resolve_classified_tier(raw_tier) + if tier is None: + raise ValueError(f"LLM classifier returned an unrecognized tier: {raw_tier!r}") + return tier, classifier_cost - payload: Final = ( + async def _classify_with_capability_llm( + self, + prompt: str, + request_kwargs: Mapping[str, object] | None, + messages: Sequence[Mapping[str, object]] | None, + ) -> tuple[ComplexityTier, float | None, CapabilityClassifierForecast]: + """Call the packaged capability forecaster and apply its two-tier policy.""" + capability: Final = self.config.capability_classifier_config + classifier_system_prompt: Final = self._classifier_system_prompt + if capability is None or classifier_system_prompt is None: + raise ValueError("capability classifier is not configured") + + markers: Final = self._reminder_markers_for_request(request_kwargs or EMPTY_MAPPING) + encrypted_task: Final = _encrypted_classifier_task(request_kwargs, markers) + asks_newest_first: Final = ( + () if encrypted_task is not None else tuple(_iter_human_asks_newest_first(messages or (), markers)) + ) + opening_task: Final = ( + "The delegated task in the following agent_message." + if encrypted_task is not None + else asks_newest_first[-1] + if asks_newest_first + else prompt + ) + latest_follow_up: Final = asks_newest_first[0] if len(asks_newest_first) > 1 else None + task_messages: list[AllMessageValues] = [ # mutable-ok: the latest message gains optional image parts below + {"role": "user", "content": opening_task}, # mutable-ok: SDK messages are dict-shaped + ] + if latest_follow_up is not None: + task_messages.append( # mutable-ok: the provider SDK requires a concrete message list + {"role": "user", "content": latest_follow_up} # mutable-ok: SDK messages are dict-shaped + ) + + image_parts: Final = self._classifier_image_parts(messages) + if image_parts: + latest_text: Final = latest_follow_up or opening_task + task_messages[-1] = { # mutable-ok: SDK messages are dict-shaped + "role": "user", + "content": [ # mutable-ok: multimodal SDK content is a JSON array + {"type": "text", "text": latest_text}, # mutable-ok: SDK content parts are dict-shaped + *image_parts, + ], + } + messages_for_call: Final[list[AllMessageValues]] = [ # mutable-ok: provider SDK requires a concrete list + {"role": "system", "content": classifier_system_prompt}, # mutable-ok: SDK messages are dict-shaped + *task_messages, + ] + content, classifier_cost = await self._call_classifier_model( + messages_for_call, + request_kwargs, + max_output_tokens=capability.max_output_tokens, + encrypted_task=encrypted_task, + ) + verdict: Final = parse_capability_classifier_verdict(content) + threshold: Final = verdict.routing_threshold(capability.base_threshold, capability.threshold_step) + calibration: Final = capability.calibration + forecast: Final = CapabilityClassifierForecast( + verdict=verdict, + threshold=threshold, + p_solve=calibration.calibrate(verdict.p_solve) if calibration is not None else verdict.p_solve, + calibration_version=calibration.version if calibration is not None else None, + ) + selected_tier: Final = ( + capability.efficient_tier if forecast.meets_routing_threshold() else capability.capable_tier + ) + return ComplexityTier(selected_tier), classifier_cost, forecast + + async def _call_classifier_model( + self, + messages_for_call: list[AllMessageValues], # mutable-ok: provider SDK requires a concrete message list + request_kwargs: Mapping[str, object] | None, + max_output_tokens: int | None = None, + encrypted_task: Mapping[str, object] | None = None, + ) -> tuple[str, float | None]: + """Execute one structured classifier call with the router's shared safeguards.""" + llm_config: Final = self.config.classifier_llm_config + response_format: Final = self._classifier_response_format + if llm_config is None or response_format is None: + raise ValueError("classifier_llm_config is not set") + + request_values: Final = request_kwargs or EMPTY_MAPPING + request_metadata = request_values.get("litellm_metadata") or request_values.get("metadata") + metadata: Final = { # mutable-ok: SDK metadata kwarg is enriched by the request pipeline + **forwarded_internal_call_metadata(request_metadata, AUTOROUTER_CLASSIFIER_CALL_ORIGIN), + INTERNAL_CALL_ORIGIN_METADATA_KEY: AUTOROUTER_CLASSIFIER_CALL_ORIGIN, + } + classifier_call_params: Final = ( + MappingProxyType({"reasoning_effort": llm_config.reasoning_effort}) + if llm_config.reasoning_effort is not None + else EMPTY_MAPPING + ) + classifier_payload: Final = ( self._native_classifier_payload(messages_for_call, response_format, encrypted_task) if encrypted_task is not None else MappingProxyType( {"messages": messages_for_call, "response_format": response_format, **classifier_call_params} ) ) + payload: Final = MappingProxyType( + { + **classifier_payload, + **( + MappingProxyType( + {"max_output_tokens" if encrypted_task is not None else "max_tokens": max_output_tokens} + ) + if max_output_tokens is not None + else EMPTY_MAPPING + ), + } + ) proxy_server_request: Final = { "originating_request_masked": masked_originating_request(request_kwargs), "body": {"model": llm_config.model, **payload}, @@ -2118,7 +2328,7 @@ class ComplexityRouter(CustomLogger): disable_fallbacks=True, metadata=metadata, proxy_server_request=proxy_server_request, - turn_off_message_logging=turn_off_message_logging, + turn_off_message_logging=_effective_turn_off_message_logging(request_kwargs), **payload, **_parent_session_kwargs(request_kwargs), ), @@ -2129,11 +2339,7 @@ class ComplexityRouter(CustomLogger): ) if not content: raise ValueError("LLM classifier returned empty content") - raw_tier: Final = _LabeledTierClassification.model_validate_json(content).tier - tier: Final = self.config.resolve_classified_tier(raw_tier) - if tier is None: - raise ValueError(f"LLM classifier returned an unrecognized tier: {raw_tier!r}") - return tier, _response_cost_or_none(response) + return content, _response_cost_or_none(response) def _native_classifier_payload( self, @@ -4088,7 +4294,12 @@ class ComplexityRouter(CustomLogger): housekeeping_ceiling: Final = tier if outcome.cause == "housekeeping" else None # A context-escalated tier becomes the hard floor: a floor the bandit can slide # under is not a floor. - adaptive_floor: Final = tier if context_original_tier is not None else plan_floor + adaptive_floor: Final = ( + tier + if context_original_tier is not None + or outcome.cause in ("capability_classifier", "capability_classifier_fallback") + else plan_floor + ) adaptive_fit: Final = context_placement.holdable_models if context_placement is not None else None sampled_model: Final = self._soft_floor_pick( tier, @@ -4138,7 +4349,8 @@ class ComplexityRouter(CustomLogger): tier_litellm_params: Final = self._litellm_params_for_model(tier, routed_model) classifier_model: Final = ( self.config.classifier_llm_config.model - if outcome.cause == "llm_classifier" and self.config.classifier_llm_config is not None + if outcome.cause in ("llm_classifier", "capability_classifier") + and self.config.classifier_llm_config is not None else None ) # cause=default_model_fallback means no tier was decided: the classifier failed and the @@ -4161,23 +4373,24 @@ class ComplexityRouter(CustomLogger): decision_keyword: Final = ( plan_mode_sentinel if plan_floored else (housekeeping_sentinel if outcome.cause == "housekeeping" else None) ) + routing_decision: Final = self._build_routing_decision( + routed_model=routed_model, + conversation_continuing=conversation_continuing, + cause=decision_cause, + tier=classified_pool_tier, + score=score, + signals=decision_signals, + matched_keyword=decision_keyword, + escalation_keyword=escalation_keyword, + escalated=escalated, + classifier_model=classifier_model, + classifier_cost=outcome.classifier_cost, + tier_litellm_params=tier_litellm_params, + context_escalation_original_tier=context_original_tier, + ) return PreRoutingHookResponse( model=routed_model, messages=messages if has_original_messages else None, litellm_params=tier_litellm_params, - routing_decision=self._build_routing_decision( - routed_model=routed_model, - conversation_continuing=conversation_continuing, - cause=decision_cause, - tier=classified_pool_tier, - score=score, - signals=decision_signals, - matched_keyword=decision_keyword, - escalation_keyword=escalation_keyword, - escalated=escalated, - classifier_model=classifier_model, - classifier_cost=outcome.classifier_cost, - tier_litellm_params=tier_litellm_params, - context_escalation_original_tier=context_original_tier, - ), + routing_decision=_with_capability_forecast(routing_decision, outcome), ) diff --git a/litellm/router_strategy/complexity_router/config.py b/litellm/router_strategy/complexity_router/config.py index bfc83f8dcab..7c47bac68da 100644 --- a/litellm/router_strategy/complexity_router/config.py +++ b/litellm/router_strategy/complexity_router/config.py @@ -13,7 +13,16 @@ from enum import Enum from types import MappingProxyType from typing import Annotated, Final, Literal, NamedTuple -from pydantic import BaseModel, ConfigDict, Field, SkipValidation, field_serializer, field_validator, model_validator +from pydantic import ( + BaseModel, + ConfigDict, + Field, + SkipValidation, + StrictFloat, + field_serializer, + field_validator, + model_validator, +) with warnings.catch_warnings(): warnings.simplefilter("ignore", DeprecationWarning) @@ -53,7 +62,7 @@ DEFAULT_CLASSIFICATION_RUBRIC: Final[ClassificationRubric] = ClassificationRubri # The classifier_type values that can call classifier_llm_config.model. Every consumer asking # "is the classifier model a real dependency of this router" resolves it here, including the ones # that only hold the raw config mapping and cannot reach ComplexityRouterConfig.uses_llm_classifier. -LLM_CLASSIFIER_TYPES: Final[frozenset[str]] = frozenset({"llm", "heuristic_first", "hybrid"}) +LLM_CLASSIFIER_TYPES: Final[frozenset[str]] = frozenset({"llm", "capability", "heuristic_first", "hybrid"}) TIER_SEVERITY_ORDER: Final[tuple[ComplexityTier, ...]] = ( @@ -591,6 +600,78 @@ class ClassifierLLMConfig(BaseModel): return self +class CapabilityCalibrationConfig(BaseModel): + model_config = ConfigDict(extra="forbid", frozen=True) + + version: str = Field(min_length=1, max_length=128, pattern=r"^\S(?:.*\S)?$") + slope: StrictFloat = Field(ge=0.0, le=20.0, allow_inf_nan=False) + intercept: StrictFloat = Field(ge=-20.0, le=20.0, allow_inf_nan=False) + + def calibrate(self, p_solve: float) -> float: + clipped: Final = min(max(p_solve, 1e-6), 1.0 - 1e-6) + log_odds: Final = self.slope * (math.log(clipped) - math.log1p(-clipped)) + self.intercept + return 1.0 / (1.0 + math.exp(-log_odds)) + + +class CapabilityClassifierConfig(BaseModel): + """Switchyard-compatible probability threshold policy for two model tiers.""" + + model_config = ConfigDict(extra="forbid", frozen=True) + + efficient_tier: str = Field( + description="Tier used when the efficient model's forecasted solve probability meets the adjusted threshold", + ) + capable_tier: str = Field( + description=( + "Higher, fail-closed tier used below the adjusted threshold or when the classifier verdict is unavailable" + ), + ) + base_threshold: StrictFloat = Field( + ge=0.0, + le=1.0, + description="Lowest p_solve that routes a supported task to efficient_tier", + ) + threshold_step: StrictFloat = Field( + default=0.0, + ge=0.0, + description=("Amount added once for uncertain or unmatched verdicts and twice for unsupported verdicts"), + ) + max_output_tokens: int = Field( + default=4096, + ge=1, + description="Maximum completion tokens available to the capability classifier verdict", + ) + calibration: CapabilityCalibrationConfig | None = Field( + default=None, + description=( + "Optional versioned sigmoid calibration fitted for this judge, capability card, efficient model, " + "and execution setup. Applies sigmoid(slope * logit(clip(p_solve, 1e-6, 1-1e-6)) + intercept) " + "before the threshold policy. Omit to route on the raw forecast." + ), + ) + response_format: Literal["json_schema", "json_object"] = Field( + default="json_schema", + description=( + "Use json_object for judges without strict JSON Schema support. This appends the verdict schema " + "to the packaged system prompt; both modes validate the returned verdict identically." + ), + ) + + @field_validator("efficient_tier", "capable_tier") + @classmethod + def _normalize_tier(cls, value: str) -> str: + normalized: Final = value.strip() + if not normalized: + raise ValueError("tier must be non-empty") + return normalized + + @model_validator(mode="after") + def _validate_threshold_range(self) -> "CapabilityClassifierConfig": + if self.base_threshold + 2 * self.threshold_step > 1.0: + raise ValueError("base_threshold + 2 * threshold_step must be at most 1") + return self + + MAX_CUSTOM_PATTERN_REPEAT: Final[int] = 64 MAX_CUSTOM_PATTERN_WORK: Final[int] = 2048 MAX_CUSTOM_DIMENSIONS_WORK: Final[int] = 8192 @@ -882,13 +963,16 @@ class ComplexityRouterConfig(BaseModel): ) # Classifier strategy - classifier_type: Literal["heuristic", "heuristic_v2", "llm", "custom", "heuristic_first", "hybrid"] = Field( + classifier_type: Literal[ + "heuristic", "heuristic_v2", "llm", "capability", "custom", "heuristic_first", "hybrid" + ] = Field( default="heuristic", description=( "Classification strategy: local regex/keyword scoring, the bundled trained four-tier heuristic, " - "an LLM call, a custom classifier plugin, 'heuristic_first', which scores locally and only pays " - "for the LLM classifier when the local scorer does not confidently land a cheap tier, or 'hybrid', " - "which trusts the local scorer everywhere except when its score lands near a tier boundary" + "an LLM tier-selection call, a Switchyard-compatible capability forecast, a custom classifier " + "plugin, 'heuristic_first', which scores locally and only pays for the LLM classifier when the " + "local scorer does not confidently land a cheap tier, or 'hybrid', which trusts the local scorer " + "everywhere except when its score lands near a tier boundary" ), ) heuristic_v2_artifact: TrainedTierArtifact | Literal["ultrafeedback"] = Field( @@ -902,7 +986,15 @@ class ComplexityRouterConfig(BaseModel): default=None, description=( "Configuration for the LLM classifier; required when classifier_type is 'llm', " - "'heuristic_first' or 'hybrid'" + "'capability', 'heuristic_first' or 'hybrid'" + ), + ) + capability_classifier_config: CapabilityClassifierConfig | None = Field( + default=None, + description=( + "Probability threshold policy required when classifier_type is 'capability'. The classifier " + "forecasts p_solve for efficient_tier, adjusts base_threshold using the capability-card boundary, " + "and otherwise routes to capable_tier" ), ) heuristic_first_max_tier: str | None = Field( @@ -1427,6 +1519,66 @@ class ComplexityRouterConfig(BaseModel): ) return self + @model_validator(mode="after") + def _validate_capability_classifier_config(self) -> "ComplexityRouterConfig": + capability: Final = self.capability_classifier_config + if self.classifier_type != "capability": + if capability is not None: + raise ValueError( + "capability_classifier_config requires classifier_type 'capability'; otherwise it has no effect" + ) + return self + if capability is None: + raise ValueError("capability_classifier_config is required when classifier_type is 'capability'") + return self + + @model_validator(mode="after") + def _validate_capability_classifier_tiers(self) -> "ComplexityRouterConfig": + capability: Final = self.capability_classifier_config + if self.classifier_type != "capability" or capability is None: + return self + if self.tier_definitions is not None: + raise ValueError( + "classifier_type 'capability' uses the built-in tier map and cannot be combined with tier_definitions" + ) + for field, tier in ( + ("efficient_tier", capability.efficient_tier), + ("capable_tier", capability.capable_tier), + ): + if tier not in self.tier_names(): + raise ValueError( + f"{field} {tier!r} is not an active tier: it must name one of {', '.join(self.tier_names())}" + ) + if not self.tiers.get(tier): + raise ValueError(f"{field} {tier!r} has no model configured in tiers") + names: Final = self.tier_names() + if names.index(capability.capable_tier) <= names.index(capability.efficient_tier): + raise ValueError("capable_tier must be a higher tier than efficient_tier") + return self + + @model_validator(mode="after") + def _validate_capability_classifier_prompt_policy(self) -> "ComplexityRouterConfig": + if self.classifier_type != "capability": + return self + llm_config: Final = self.classifier_llm_config + if llm_config is not None and ( + llm_config.system_prompt is not None or llm_config.classification_rubric is not None + ): + raise ValueError( + "classifier_type 'capability' uses the packaged capability card; classifier_llm_config.system_prompt " + "and classification_rubric are not supported" + ) + if self.classification_prompt is not None or self.classification_examples is not None: + raise ValueError( + "classifier_type 'capability' uses the packaged capability card; classification_prompt and " + "classification_examples are not supported" + ) + if self.classifier_fallback != "heuristic": + raise ValueError( + "classifier_type 'capability' always fails closed to capable_tier; classifier_fallback cannot override it" + ) + return self + @model_validator(mode="after") def _validate_custom_dimensions(self) -> "ComplexityRouterConfig": if not self.custom_dimensions: @@ -1690,7 +1842,7 @@ class ComplexityRouterConfig(BaseModel): ) if duplicated: raise ValueError(f"tier_definitions names must be unique (case-insensitive): {', '.join(duplicated)}") - if self.classifier_type in ("heuristic", "heuristic_v2", "heuristic_first", "hybrid"): + if self.classifier_type in ("heuristic", "heuristic_v2", "capability", "heuristic_first", "hybrid"): raise ValueError( "tier_definitions requires classifier_type 'llm' or 'custom': the heuristic scorer only " "produces the built-in tiers from SIMPLE up, as does heuristic_v2" diff --git a/litellm/router_utils/auto_router_model_naming.py b/litellm/router_utils/auto_router_model_naming.py index 190c4921d5f..9af8a9a1180 100644 --- a/litellm/router_utils/auto_router_model_naming.py +++ b/litellm/router_utils/auto_router_model_naming.py @@ -113,7 +113,7 @@ def strategy_router_dependencies( """The model names a strategy-router deployment must reach, in no particular order. A field is a dependency only under the condition the runtime itself reads it: the - classifier model needs `classifier_type: llm`, and the complexity embedding model needs + classifier model needs an LLM-backed classifier type, and the complexity embedding model needs `semantic_keyword_matching`. Listing one the router never calls reds a working deployment. The two default-model spellings are not symmetric. A quality router falls back to its diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 8bab7c349ff..b723248bb93 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -2891,6 +2891,7 @@ RoutingDecisionCause = Literal[ # meant anything that filtered `signals` silently changed what the row claimed. "reasoning_override", "llm_classifier", + "capability_classifier", # classifier_type 'heuristic_first': the local scorer produced at least one signal and landed at # or below heuristic_first_max_tier, so it decided the tier and the LLM classifier was never # called. Distinct from "heuristic_scorer", which is a router whose only classifier IS the @@ -2903,6 +2904,9 @@ RoutingDecisionCause = Literal[ # The LLM classifier or classifier plugin failed on a router with an operator-defined # tier set, so the request routed to the configured fallback_tier without being classified. "classifier_fallback", + # The capability judge failed or returned an invalid verdict, so its fail-closed policy + # routed to capable_tier without consulting the unrelated complexity heuristic. + "capability_classifier_fallback", # The LLM classifier or classifier plugin failed and classifier_fallback is # 'default_model', so the request went to default_model without being classified. # Distinct from "default_fallback", @@ -2978,6 +2982,13 @@ class StandardLoggingRoutingDecision(TypedDict, total=False): escalation_keyword: str classifier_model: str classifier_cost: float + classifier_crux: str # writable-ok: added only when a capability verdict is available + classifier_primary_rule: str # writable-ok: added only when a capability verdict is available + classifier_capability_boundary: str # writable-ok: added only when a capability verdict is available + classifier_p_solve: float # writable-ok: added only when a capability verdict is available + classifier_calibrated_p_solve: ReadOnly[float] + classifier_calibration_version: ReadOnly[str] + classifier_threshold: float # writable-ok: added only when a capability verdict is available escalated: bool context_escalated: bool # writable-ok: Pydantic warns on ReadOnly TypedDict fields context_escalation_original_tier: str # writable-ok: Pydantic warns on ReadOnly TypedDict fields @@ -2993,7 +3004,9 @@ class StandardLoggingRoutingDecision(TypedDict, total=False): # logging off. Every other field aggregates the prompt without reproducing it and is kept, # so a redacted row stays explainable. `test_every_routing_decision_field_is_classified` # fails if a field is added to the record without being placed in one set or the other. -PROMPT_QUOTING_ROUTING_DECISION_FIELDS: frozenset[str] = frozenset({"signals", "matched_keyword", "escalation_keyword"}) +PROMPT_QUOTING_ROUTING_DECISION_FIELDS: frozenset[str] = frozenset( + {"signals", "matched_keyword", "escalation_keyword", "classifier_crux"} +) DERIVED_ROUTING_DECISION_FIELDS: Final[frozenset[str]] = frozenset( { "router_model_name", @@ -3006,6 +3019,12 @@ DERIVED_ROUTING_DECISION_FIELDS: Final[frozenset[str]] = frozenset( "score", "classifier_model", "classifier_cost", + "classifier_primary_rule", + "classifier_capability_boundary", + "classifier_p_solve", + "classifier_calibrated_p_solve", + "classifier_calibration_version", + "classifier_threshold", "escalated", "context_escalated", "context_escalation_original_tier", diff --git a/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py index fb7b515eca1..067f30c2fd7 100644 --- a/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py +++ b/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py @@ -338,6 +338,15 @@ def test_a_request_must_carry_exactly_one_usable_conversation(body: dict): "config_overrides", [ {"classifier_type": "llm", "classifier_llm_config": {"model": "classifier-model"}}, + { + "classifier_type": "capability", + "classifier_llm_config": {"model": "classifier-model"}, + "capability_classifier_config": { + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.5, + }, + }, { "semantic_keyword_matching": True, "embedding_model": "classifier-model", diff --git a/tests/test_litellm/router_strategy/test_complexity_router.py b/tests/test_litellm/router_strategy/test_complexity_router.py index cd72388ea21..0931b9d01a7 100644 --- a/tests/test_litellm/router_strategy/test_complexity_router.py +++ b/tests/test_litellm/router_strategy/test_complexity_router.py @@ -7,6 +7,7 @@ Tests the rule-based complexity scoring and tier assignment logic. import asyncio import json import logging +import math import sys import time from collections.abc import AsyncIterator, Mapping, Sequence @@ -52,7 +53,13 @@ from litellm.router_strategy.complexity_router.complexity_router import ( classification_system_prompt, custom_tier_classification_prompt, ) +from litellm.router_strategy.complexity_router.capability_classifier import ( + CAPABILITY_CLASSIFIER_SYSTEM_PROMPT, + CapabilityClassifierVerdict, +) from litellm.router_strategy.complexity_router.config import ( + CapabilityCalibrationConfig, + CapabilityClassifierConfig, DEFAULT_CLASSIFICATION_RUBRIC, DEFAULT_CLASSIFIER_CONTEXT_WINDOW_SIZE, DEFAULT_COMPLEXITY_CONFIG, @@ -2363,6 +2370,515 @@ class TestLLMClassifierConfig: ) +CAPABILITY_TIERS: Dict[str, str] = { + "SIMPLE": "efficient-model", + "REASONING": "capable-model", +} + + +def _capability_router_config(**overrides): + return { + "tiers": dict(CAPABILITY_TIERS), + "classifier_type": "capability", + "classifier_llm_config": {"model": "judge-model", "timeout_ms": 400}, + "capability_classifier_config": { + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.5, + "threshold_step": 0.1, + }, + **overrides, + } + + +def _capability_reply( + *, + p_solve: float, + primary_rule: str = "SUP-1", + capability_boundary: str = "supported", + crux: str = "complete the requested change", +) -> str: + return json.dumps( + { + "crux": crux, + "primary_rule": primary_rule, + "capability_boundary": capability_boundary, + "p_solve": p_solve, + } + ) + + +class TestCapabilityClassifierConfig: + @pytest.mark.parametrize( + "calibration", + ( + {"version": "v1", "slope": -1.0, "intercept": 0.0}, + {"version": "v1", "slope": float("nan"), "intercept": 0.0}, + {"version": "v1", "slope": 1.0, "intercept": float("inf")}, + {"version": "v1", "slope": True, "intercept": 0.0}, + {"version": " ", "slope": 1.0, "intercept": 0.0}, + {"version": "v1", "slope": 1.0, "intercept": 0.0, "typo": 1}, + ), + ) + def test_rejects_invalid_calibration(self, calibration: dict[str, object]) -> None: + with pytest.raises(ValidationError): + CapabilityCalibrationConfig.model_validate(calibration) + + def test_calibration_round_trip_and_probability_endpoints(self) -> None: + calibration: Final = CapabilityCalibrationConfig(version="held-out-v1", slope=0.0, intercept=0.0) + config: Final = CapabilityClassifierConfig( + efficient_tier="SIMPLE", capable_tier="REASONING", base_threshold=0.6, calibration=calibration + ) + restored: Final = CapabilityClassifierConfig.model_validate_json(config.model_dump_json()) + assert restored.calibration == calibration + assert tuple(calibration.calibrate(p) for p in (0.0, 0.5, 1.0)) == (0.5, 0.5, 0.5) + steep: Final = CapabilityCalibrationConfig(version="endpoints", slope=20.0, intercept=-20.0) + values: Final = tuple(steep.calibrate(p) for p in (0.0, 0.5, 1.0)) + assert all(math.isfinite(p) and 0.0 <= p <= 1.0 for p in values) + assert values[0] < values[1] < values[2] + + @pytest.mark.parametrize( + "patch,error_match", + [ + ({"classifier_llm_config": None}, "classifier_llm_config is required"), + ({"capability_classifier_config": None}, "capability_classifier_config is required"), + ( + { + "capability_classifier_config": { + "efficient_tier": "SIMPLE", + "capable_tier": "SIMPLE", + "base_threshold": 0.5, + } + }, + "must be a higher tier", + ), + ( + { + "capability_classifier_config": { + "efficient_tier": "REASONING", + "capable_tier": "SIMPLE", + "base_threshold": 0.5, + } + }, + "must be a higher tier", + ), + ( + { + "capability_classifier_config": { + "efficient_tier": "MEDIUM", + "capable_tier": "REASONING", + "base_threshold": 0.5, + } + }, + "has no model configured", + ), + ( + { + "capability_classifier_config": { + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.9, + "threshold_step": 0.1, + } + }, + r"base_threshold \+ 2 \* threshold_step must be at most 1", + ), + ({"classifier_fallback": "default_model", "default_model": "fallback"}, "always fails closed"), + ( + {"classifier_llm_config": {"model": "judge-model", "system_prompt": "pick one"}}, + "uses the packaged capability card", + ), + ({"classification_examples": "example"}, "uses the packaged capability card"), + ], + ) + def test_rejects_incoherent_configuration(self, patch, error_match): + with pytest.raises(ValidationError, match=error_match): + ComplexityRouterConfig(**{**_capability_router_config(), **patch}) + + def test_capability_config_is_rejected_on_other_classifier_types(self): + config = _capability_router_config(classifier_type="llm") + with pytest.raises(ValidationError, match="requires classifier_type 'capability'"): + ComplexityRouterConfig(**config) + + def test_rejects_misspelled_optional_policy_instead_of_using_defaults(self) -> None: + with pytest.raises(ValidationError, match="threshold_steps"): + CapabilityClassifierConfig.model_validate( + { + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.5, + "threshold_steps": 0.2, + } + ) + + def test_threshold_defaults_match_switchyard(self): + config = CapabilityClassifierConfig(efficient_tier=" SIMPLE ", capable_tier=" REASONING ", base_threshold=0.5) + assert config.efficient_tier == "SIMPLE" + assert config.capable_tier == "REASONING" + assert config.threshold_step == 0.0 + assert config.max_output_tokens == 4096 + + def test_classifier_model_is_registered_as_a_dependency(self): + assert ComplexityRouterConfig(**_capability_router_config()).uses_llm_classifier is True + + +class TestCapabilityClassifierVerdict: + @pytest.mark.parametrize( + "primary_rule,capability_boundary", + [ + *((f"SUP-{index}", "supported") for index in range(1, 6)), + *((f"UNC-{index}", "uncertain") for index in range(1, 3)), + *((f"LIM-{index}", "unsupported") for index in range(1, 3)), + ("none", "unmatched"), + ], + ) + def test_accepts_every_valid_rule_boundary_pair(self, primary_rule, capability_boundary): + verdict = CapabilityClassifierVerdict( + crux="the hard part", + primary_rule=primary_rule, + capability_boundary=capability_boundary, + p_solve=0.5, + ) + assert verdict.primary_rule == primary_rule + assert verdict.capability_boundary == capability_boundary + + @pytest.mark.parametrize( + "payload,error_match", + [ + ( + { + "crux": "x", + "primary_rule": "SUP-1", + "capability_boundary": "unsupported", + "p_solve": 0.5, + }, + "requires capability_boundary", + ), + ( + {"crux": " ", "primary_rule": "none", "capability_boundary": "unmatched", "p_solve": 0.5}, + "non-whitespace", + ), + ( + { + "crux": "x", + "primary_rule": "none", + "capability_boundary": "unmatched", + "p_solve": 0.5, + "recommended_route": "efficient", + }, + "Extra inputs are not permitted", + ), + ( + {"crux": "x", "primary_rule": "none", "capability_boundary": "unmatched", "p_solve": True}, + "valid number", + ), + ], + ) + def test_rejects_invalid_or_inconsistent_verdicts(self, payload, error_match): + with pytest.raises(ValidationError, match=error_match): + CapabilityClassifierVerdict.model_validate(payload) + + +class TestCapabilityClassifier: + @staticmethod + def _router(mock_router_instance, **overrides): + return ComplexityRouter( + model_name="capability-router", + litellm_router_instance=mock_router_instance, + complexity_router_config=_capability_router_config(**overrides), + ) + + @pytest.mark.asyncio + async def test_encrypted_task_is_not_replaced_by_plaintext_envelope(self, mock_router_instance: MagicMock) -> None: + mock_router_instance.aresponses = AsyncMock( + return_value=_native_classifier_response(_capability_reply(p_solve=0.8)) + ) + router: Final = self._router(mock_router_instance) + task: Final = _encrypted_agent_task() + request: Final = {"input": [task]} + original: Final = deepcopy(request) + result: Final = await router.async_pre_routing_hook(model="capability-router", request_kwargs=request) + assert result is not None and result.model == "efficient-model" + assert result.routing_decision is not None + assert result.routing_decision["cause"] == "capability_classifier" + mock_router_instance.aresponses.assert_awaited_once() + call: Final = mock_router_instance.aresponses.call_args.kwargs + assert call["input"][-1] == task + plaintext: Final = json.dumps(call["input"][:-1]) + assert "The delegated task in the following agent_message." in plaintext + assert "Message Type: NEW_TASK" not in plaintext + assert "opaque-provider-task" not in plaintext + assert request == original + + @pytest.mark.asyncio + @pytest.mark.parametrize("custom_markers", (False, True)) + async def test_task_forecast_uses_request_scoped_codex_markers( + self, mock_router_instance: MagicMock, custom_markers: bool + ) -> None: + completion: Final = AsyncMock(return_value=_llm_response(_capability_reply(p_solve=0.8))) + mock_router_instance.acompletion = completion + router: Final = self._router( + mock_router_instance, + escalation_keywords=[], + **({"reminder_markers": [{"open": "", "close": ""}]} if custom_markers else {}), + ) + envelope: Final = "\n".join(_CODEX_ENVELOPES) + opening: Final = f"{envelope}\nFix nested behavior" + messages: Final = [ + {"role": "user", "content": opening}, + {"role": "user", "content": "Preserve empty inputs"}, + {"role": "user", "content": envelope}, + ] + original: Final = deepcopy(messages) + for user_agent in ("codex-tui", "curl/8.7.1", "codex_cli_rs/0.62.0"): + result: Final = await router.async_pre_routing_hook( + model="capability-router", messages=messages, request_kwargs={"metadata": {"user_agent": user_agent}} + ) + assert result is not None and result.model == "efficient-model" + sent: Final = completion.call_args.kwargs["messages"] + if user_agent.startswith("codex") and not custom_markers: + assert [message["content"] for message in sent[1:]] == ["Fix nested behavior", "Preserve empty inputs"] + else: + assert [message["content"] for message in sent[1:]] == [opening, envelope] + assert result.messages == original + assert completion.await_count == 3 + assert messages == original + + @pytest.mark.asyncio + @pytest.mark.parametrize("p_solve,expected_model", ((0.95, "capable-model"), (0.98, "efficient-model"))) + async def test_fitted_probability_controls_routing_and_preserves_raw_score( + self, mock_router_instance: MagicMock, p_solve: float, expected_model: str + ) -> None: + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response(_capability_reply(p_solve=p_solve))) + router: Final = self._router( + mock_router_instance, + capability_classifier_config={ + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.66, + "threshold_step": 0.1, + "calibration": { + "version": "qwen3-haiku45-mini-swe-v1", + "slope": 0.1482462649948327, + "intercept": 0.1895438369492216, + }, + }, + ) + result: Final = await router.async_pre_routing_hook( + model="capability-router", request_kwargs={}, messages=[{"role": "user", "content": "Fix the issue"}] + ) + assert result is not None and result.model == expected_model + decision: Final = result.routing_decision + assert decision is not None + assert decision["classifier_p_solve"] == p_solve + assert decision["classifier_threshold"] == 0.66 + assert decision["classifier_calibration_version"] == "qwen3-haiku45-mini-swe-v1" + assert 0.65 < decision["classifier_calibrated_p_solve"] < 0.69 + assert (decision["classifier_calibrated_p_solve"] >= 0.66) == (expected_model == "efficient-model") + + @pytest.mark.asyncio + @pytest.mark.parametrize("mode", ("json_schema", "json_object")) + async def test_response_modes_preserve_the_card_and_validate_the_same_verdict( + self, mock_router_instance: MagicMock, mode: str + ) -> None: + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response(_capability_reply(p_solve=0.8))) + router: Final = self._router( + mock_router_instance, + capability_classifier_config={ + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.5, + "response_format": mode, + }, + ) + outcome: Final = await router.aclassify("Fix the issue") + assert outcome.tier == ComplexityTier.SIMPLE + call: Final = mock_router_instance.acompletion.call_args.kwargs + system_prompt: Final = call["messages"][0]["content"] + assert call["response_format"]["type"] == mode + if mode == "json_object": + marker: Final = "\n\nReturn exactly one JSON object matching this JSON Schema:\n" + assert system_prompt.startswith(CAPABILITY_CLASSIFIER_SYSTEM_PROMPT + marker) + schema: Final = json.loads(system_prompt.split(marker)[1]) + assert schema["required"] == ["crux", "primary_rule", "capability_boundary", "p_solve"] + assert schema["additionalProperties"] is False + else: + assert system_prompt == CAPABILITY_CLASSIFIER_SYSTEM_PROMPT + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response("invalid JSON")) + assert (await router.aclassify("Fix another issue")).tier == ComplexityTier.REASONING + + @pytest.mark.asyncio + @pytest.mark.parametrize("reply", ("invalid JSON", _capability_reply(p_solve=0.0))) + async def test_adaptive_selection_cannot_undo_a_capable_verdict( + self, mock_router_instance: MagicMock, reply: str + ) -> None: + from litellm.router_strategy.adaptive_router.bandit import BanditCell + from litellm.types.router import RequestType + + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response(reply)) + mock_router_instance.model_list = [ + {"model_name": "efficient-model", "litellm_params": {"input_cost_per_token": 0.000001}}, + {"model_name": "capable-model", "litellm_params": {"input_cost_per_token": 0.00001}}, + ] + mock_router_instance.model_name_to_deployment_indices = {"efficient-model": [0], "capable-model": [1]} + router: Final = self._router( + mock_router_instance, + adaptive=True, + adaptive_eligible="all", + adaptive_weights={"quality": 0.0, "cost": 1.0}, + tier_distance_penalty=0.0, + tiers={"SIMPLE": ["efficient-model"], "REASONING": ["capable-model"]}, + ) + adaptive: Final = router._ensure_adaptive_router() + assert adaptive is not None + for model in ("efficient-model", "capable-model"): + adaptive._cells[(RequestType.GENERAL, model)] = BanditCell(alpha=20.0, beta=1.0) + assert router._soft_floor_pick(ComplexityTier.REASONING, "Fix the issue") == "efficient-model" + result: Final = await router.async_pre_routing_hook( + model="capability-router", request_kwargs={}, messages=[{"role": "user", "content": "Fix the issue"}] + ) + assert result is not None and result.model == "capable-model" + assert result.routing_decision is not None + assert result.routing_decision["tier"] == "REASONING" + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "p_solve,primary_rule,boundary,expected_tier,expected_threshold", + [ + (0.5, "SUP-1", "supported", ComplexityTier.SIMPLE, 0.5), + (0.59, "UNC-1", "uncertain", ComplexityTier.REASONING, 0.6), + (0.6, "UNC-1", "uncertain", ComplexityTier.SIMPLE, 0.6), + (0.59, "none", "unmatched", ComplexityTier.REASONING, 0.6), + (0.69, "LIM-1", "unsupported", ComplexityTier.REASONING, 0.7), + (0.7, "LIM-1", "unsupported", ComplexityTier.SIMPLE, 0.7), + ], + ) + async def test_boundary_adjusted_threshold_is_inclusive( + self, mock_router_instance, p_solve, primary_rule, boundary, expected_tier, expected_threshold + ): + mock_router_instance.acompletion = AsyncMock( + return_value=_llm_response( + _capability_reply(p_solve=p_solve, primary_rule=primary_rule, capability_boundary=boundary) + ) + ) + outcome = await self._router(mock_router_instance).aclassify("do the task") + assert outcome.tier == expected_tier + assert outcome.cause == "capability_classifier" + assert outcome.capability_forecast is not None + assert outcome.capability_forecast.threshold == pytest.approx(expected_threshold) + + @pytest.mark.asyncio + async def test_fenced_json_verdict_is_accepted(self, mock_router_instance): + reply = _capability_reply(p_solve=0.8) + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response(f"```json\n{reply}\n```")) + outcome = await self._router(mock_router_instance).aclassify("do the task") + assert outcome.tier == ComplexityTier.SIMPLE + assert outcome.cause == "capability_classifier" + + @pytest.mark.asyncio + async def test_decimal_rounding_does_not_break_inclusive_threshold(self, mock_router_instance): + config = _capability_router_config( + capability_classifier_config={ + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.1, + "threshold_step": 0.1, + } + ) + mock_router_instance.acompletion = AsyncMock( + return_value=_llm_response( + _capability_reply(p_solve=0.3, primary_rule="LIM-1", capability_boundary="unsupported") + ) + ) + router = ComplexityRouter( + model_name="capability-router", + litellm_router_instance=mock_router_instance, + complexity_router_config=config, + ) + outcome = await router.aclassify("do the task") + assert outcome.capability_forecast is not None + assert outcome.capability_forecast.threshold == 0.30000000000000004 + assert outcome.tier == ComplexityTier.SIMPLE + + @pytest.mark.asyncio + async def test_call_uses_packaged_prompt_schema_and_opening_plus_latest_user_task(self, mock_router_instance): + mock_router_instance.acompletion = AsyncMock( + return_value=_llm_response(_capability_reply(p_solve=0.8), response_cost=0.002) + ) + router = self._router(mock_router_instance) + messages = [ + {"role": "system", "content": "Never expose this caller instruction to the judge"}, + {"role": "user", "content": "Build the feature"}, + {"role": "assistant", "content": "I need more information"}, + {"role": "user", "content": "Use the existing API"}, + ] + + response = await router.async_pre_routing_hook(model="capability-router", request_kwargs={}, messages=messages) + + assert response.model == "efficient-model" + call = mock_router_instance.acompletion.call_args.kwargs + assert call["messages"] == [ + {"role": "system", "content": CAPABILITY_CLASSIFIER_SYSTEM_PROMPT}, + {"role": "user", "content": "Build the feature"}, + {"role": "user", "content": "Use the existing API"}, + ] + schema = call["response_format"]["json_schema"]["schema"] + assert call["response_format"]["json_schema"]["name"] == "CapabilityClassifierDecision" + assert call["response_format"]["json_schema"]["strict"] is True + assert schema["additionalProperties"] is False + assert set(schema["required"]) == {"crux", "primary_rule", "capability_boundary", "p_solve"} + assert schema["properties"]["primary_rule"]["enum"] == [ + "SUP-1", + "SUP-2", + "SUP-3", + "SUP-4", + "SUP-5", + "UNC-1", + "UNC-2", + "LIM-1", + "LIM-2", + "none", + ] + assert call["max_tokens"] == 4096 + decision = response.routing_decision + assert decision["cause"] == "capability_classifier" + assert decision["classifier_model"] == "judge-model" + assert decision["classifier_cost"] == 0.002 + assert decision["classifier_crux"] == "complete the requested change" + assert decision["classifier_primary_rule"] == "SUP-1" + assert decision["classifier_capability_boundary"] == "supported" + assert decision["classifier_p_solve"] == 0.8 + assert decision["classifier_threshold"] == 0.5 + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "reply", + [ + "not json", + _capability_reply(p_solve=0.9, primary_rule="SUP-1", capability_boundary="unsupported"), + '{"crux":"x","primary_rule":"SUP-1","capability_boundary":"supported","p_solve":0.9,"route":"efficient"}', + ], + ids=["malformed", "inconsistent-pair", "extra-field"], + ) + async def test_invalid_verdict_fails_closed_to_capable_tier(self, mock_router_instance, reply): + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response(reply)) + outcome = await self._router(mock_router_instance).aclassify("do the task") + assert outcome.tier == ComplexityTier.REASONING + assert outcome.cause == "capability_classifier_fallback" + assert outcome.signals == ("capability-classifier-fallback",) + + @pytest.mark.asyncio + async def test_classifier_call_failure_fails_closed_to_capable_model(self, mock_router_instance): + mock_router_instance.acompletion = AsyncMock(side_effect=TimeoutError("judge unavailable")) + response = await self._router(mock_router_instance).async_pre_routing_hook( + model="capability-router", + request_kwargs={}, + messages=[{"role": "user", "content": "do the task"}], + ) + assert response.model == "capable-model" + assert response.routing_decision["cause"] == "capability_classifier_fallback" + + CUSTOM_TIER_LABELS: Dict[str, str] = { "SIMPLE": "Cheap", "MEDIUM": "Standard", @@ -8342,6 +8858,13 @@ class TestRedactedLoggingDropsPromptText: "score": 0.8, "tier_boundaries": {"simple_medium": 0.15, "medium_complex": 0.35, "complex_reasoning": 0.6}, "classifier_model": "claude-haiku", + "classifier_crux": "deploy the requested service to k8s", + "classifier_primary_rule": "SUP-2", + "classifier_capability_boundary": "supported", + "classifier_p_solve": 0.8, + "classifier_calibrated_p_solve": 0.65, + "classifier_calibration_version": "fitted-v1", + "classifier_threshold": 0.5, "escalated": True, "tier_litellm_params": {"reasoning_effort": "xhigh"}, "signals": ["code (python)"], @@ -8349,7 +8872,15 @@ class TestRedactedLoggingDropsPromptText: "escalation_keyword": "LITELLM ESCALATE", } kept = Router._redact_prompt_text_if_needed(request_kwargs={}, routing_decision=full) - assert set(full) - set(kept) == {"signals", "matched_keyword", "escalation_keyword"} + assert set(full) - set(kept) == { + "signals", + "matched_keyword", + "escalation_keyword", + "classifier_crux", + } + assert kept["classifier_p_solve"] == 0.8 + assert kept["classifier_calibrated_p_solve"] == 0.65 + assert kept["classifier_calibration_version"] == "fitted-v1" assert kept["tier_litellm_params"] == {"reasoning_effort": "xhigh"} @pytest.mark.asyncio @@ -8463,8 +8994,10 @@ class TestContextAwareClassifier: assert messages == original_messages assert (claude_kwargs, compared_kwargs) == original_kwargs calls: Final = tuple(call.kwargs["messages"] for call in dependency.acompletion.await_args_list) - assert calls[0][0]["content"] == calls[1][0]["content"] == classification_system_prompt( - router.config.classifier_context_window_size + assert ( + calls[0][0]["content"] + == calls[1][0]["content"] + == classification_system_prompt(router.config.classifier_context_window_size) ) payloads: Final = (calls[0][1]["content"], calls[1][1]["content"]) for payload, expected_system in zip(payloads, (False, forwards_system)): @@ -13653,11 +14186,7 @@ class TestHealthFallbackDispatch: "api_key": "test-only", "api_base": f"https://{name}.test{base_suffix}", **({"tags": [name]} if tagged else {}), - **( - {"max_budget": 1.0, "budget_duration": "1d"} - if budgeted and name == "primary" - else {} - ), + **({"max_budget": 1.0, "budget_duration": "1d"} if budgeted and name == "primary" else {}), }, "model_info": {"id": f"{name}-id"}, } diff --git a/tests/test_litellm/router_utils/test_auto_router_model_naming.py b/tests/test_litellm/router_utils/test_auto_router_model_naming.py index 8dede941a14..61e31255d12 100644 --- a/tests/test_litellm/router_utils/test_auto_router_model_naming.py +++ b/tests/test_litellm/router_utils/test_auto_router_model_naming.py @@ -214,6 +214,22 @@ def test_config_check_ignores_the_model_entirely(): }, (("a", "tier"), ("clf", "classifier")), ), + ( + { + "model": "auto_router/complexity_router", + "complexity_router_config": { + "tiers": {"SIMPLE": "a", "REASONING": "b"}, + "classifier_type": "capability", + "classifier_llm_config": {"model": "clf"}, + "capability_classifier_config": { + "efficient_tier": "SIMPLE", + "capable_tier": "REASONING", + "base_threshold": 0.5, + }, + }, + }, + (("a", "tier"), ("b", "tier"), ("clf", "classifier")), + ), ( { "model": "auto_router/complexity_router", diff --git a/ui/litellm-dashboard/src/components/add_model/ClassificationMethodConfig.tsx b/ui/litellm-dashboard/src/components/add_model/ClassificationMethodConfig.tsx index 93ccf561387..ffd7468152f 100644 --- a/ui/litellm-dashboard/src/components/add_model/ClassificationMethodConfig.tsx +++ b/ui/litellm-dashboard/src/components/add_model/ClassificationMethodConfig.tsx @@ -433,6 +433,15 @@ const ClassificationMethodConfig: React.FC = ({ }); }; + if (classifierType === "capability") { + return ( +

+ This router uses capability forecasting. Configure its classifier, threshold, and calibration through YAML or + the API. Saving preserves those settings +

+ ); + } + return ( <> diff --git a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx index febcde269f7..6da1133c57b 100644 --- a/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx +++ b/ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx @@ -143,7 +143,7 @@ export interface ClassifierLLMConfig { system_prompt?: string; } -export type ClassifierType = "heuristic" | "heuristic_v2" | "llm" | "heuristic_first" | "hybrid"; +export type ClassifierType = "heuristic" | "heuristic_v2" | "llm" | "heuristic_first" | "hybrid" | "capability"; /** * Whether this router can call classifier_llm_config.model. Mirrors the backend's @@ -151,7 +151,7 @@ export type ClassifierType = "heuristic" | "heuristic_v2" | "llm" | "heuristic_f * control and payload key, so a new chaining type cannot strip knobs the operator set. */ export const usesLlmClassifier = (classifierType: ClassifierType): boolean => - classifierType === "llm" || classifierType === "heuristic_first" || classifierType === "hybrid"; + (["llm", "heuristic_first", "hybrid", "capability"] as const).some((type) => type === classifierType); export type ClassifierFallback = "heuristic" | "default_model"; @@ -176,7 +176,7 @@ export const heuristicScoringRoleFor = ( classifierType: ClassifierType, classifierFallback: ClassifierFallback | undefined, ): HeuristicScoringRole => { - if (classifierType === "heuristic_v2") return "never"; + if (classifierType === "heuristic_v2" || classifierType === "capability") return "never"; if (classifierType === "heuristic" || classifierType === "heuristic_first" || classifierType === "hybrid") return "decides"; return (classifierFallback ?? DEFAULT_CLASSIFIER_FALLBACK) === "heuristic" ? "fallback_only" : "never"; diff --git a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts index 5b53941bc10..822af77e0db 100644 --- a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts +++ b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts @@ -470,7 +470,10 @@ const classifierWireFields = ( >, ): Partial => ({ ...(usesLlmClassifier(effectiveType) && - classifierLlmConfig && { classifier_llm_config: normalizeClassifierLlmConfig(classifierLlmConfig) }), + classifierLlmConfig && { + classifier_llm_config: + effectiveType === "capability" ? classifierLlmConfig : normalizeClassifierLlmConfig(classifierLlmConfig), + }), ...(usesLlmClassifier(effectiveType) && classifierFallback !== undefined && { classifier_fallback: classifierFallback }), ...(effectiveType === "heuristic_first" && diff --git a/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts b/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts index 3899361cd17..82785610646 100644 --- a/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts +++ b/ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts @@ -164,6 +164,32 @@ const STORED_LLM = { classifier_context_per_turn_chars: 300, }; +describe("capability classifier configuration", () => { + it("preserves the judge and calibrated policy through an untouched dashboard edit", () => { + const stored = { + tiers: { SIMPLE: ["efficient-model"], REASONING: ["capable-model"] }, + classifier_type: "capability" as const, + classifier_llm_config: { model: "judge", timeout_ms: 30000, temperature: 0 }, + capability_classifier_config: { + efficient_tier: "SIMPLE", + capable_tier: "REASONING", + base_threshold: 0.66, + max_output_tokens: 512, + response_format: "json_object", + calibration: { version: "fitted-v1", slope: 0.15, intercept: 0.19 }, + }, + }; + const hydrated = hydrateComplexityRouterConfig(stored, null); + const saved = buildUpdatedComplexityRouterConfig(stored, hydrated); + + expect(saved.classifier_type).toBe("capability"); + expect(saved.classifier_llm_config).toEqual(stored.classifier_llm_config); + expect(saved.capability_classifier_config).toEqual(stored.capability_classifier_config); + expect(saved).not.toHaveProperty("classification_prompt"); + expect(saved).not.toHaveProperty("custom_dimensions"); + }); +}); + describe("buildUpdatedComplexityRouterConfig classifier context window", () => { it("round-trips an untouched edit without changing the classifier context values", () => { const formValue = { diff --git a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.integration.test.tsx b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.integration.test.tsx index 646e83a773b..0bb3340ac09 100644 --- a/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.integration.test.tsx +++ b/ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.integration.test.tsx @@ -105,7 +105,7 @@ describe("EditAutoRouterModal keyword matching", () => { expect(screen.queryByText("Advanced: Compression")).not.toBeInTheDocument(); expect(screen.queryByText("Model Access Groups")).not.toBeInTheDocument(); await user.click(screen.getByText("Advanced: Affinity")); - await user.click(await screen.findByRole("switch", { name: "Pin a session to one deployment per model group" })); + await user.click(await screen.findByRole("switch", { name: "Pin one model deployment per tier" })); await user.click(screen.getByRole("button", { name: /save changes/i })); await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled()); expect(modelPatchUpdateCall).toHaveBeenLastCalledWith( diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 371fbbe67bd..ba7909b3ab1 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -25328,6 +25328,57 @@ export interface components { */ status: "cancelled"; }; + /** CapabilityCalibrationConfig */ + CapabilityCalibrationConfig: { + /** Intercept */ + intercept: number; + /** Slope */ + slope: number; + /** Version */ + version: string; + }; + /** + * CapabilityClassifierConfig + * @description Switchyard-compatible probability threshold policy for two model tiers. + */ + CapabilityClassifierConfig: { + /** + * Base Threshold + * @description Lowest p_solve that routes a supported task to efficient_tier + */ + base_threshold: number; + /** @description Optional versioned sigmoid calibration fitted for this judge, capability card, efficient model, and execution setup. Applies sigmoid(slope * logit(clip(p_solve, 1e-6, 1-1e-6)) + intercept) before the threshold policy. Omit to route on the raw forecast. */ + calibration?: components["schemas"]["CapabilityCalibrationConfig"] | null; + /** + * Capable Tier + * @description Higher, fail-closed tier used below the adjusted threshold or when the classifier verdict is unavailable + */ + capable_tier: string; + /** + * Efficient Tier + * @description Tier used when the efficient model's forecasted solve probability meets the adjusted threshold + */ + efficient_tier: string; + /** + * Max Output Tokens + * @description Maximum completion tokens available to the capability classifier verdict + * @default 4096 + */ + max_output_tokens: number; + /** + * Response Format + * @description Use json_object for judges without strict JSON Schema support. This appends the verdict schema to the packaged system prompt; both modes validate the returned verdict identically. + * @default json_schema + * @enum {string} + */ + response_format: "json_schema" | "json_object"; + /** + * Threshold Step + * @description Amount added once for uncertain or unmatched verdicts and twice for unsupported verdicts + * @default 0 + */ + threshold_step: number; + }; /** ChatCompletionAnnotation */ ChatCompletionAnnotation: { /** @@ -35542,6 +35593,8 @@ export interface components { adaptive_eligible: "all" | "classified_tier"; /** @description Quality vs cost weights for adaptive selection (used when adaptive=True) */ adaptive_weights?: components["schemas"]["AdaptiveRouterWeights"]; + /** @description Probability threshold policy required when classifier_type is 'capability'. The classifier forecasts p_solve for efficient_tier, adjusts base_threshold using the capability-card boundary, and otherwise routes to capable_tier */ + capability_classifier_config?: components["schemas"]["CapabilityClassifierConfig"] | null; /** * Classification Examples * @description Replaces the calibration examples of the LLM classifier rubric, and nothing else. Written as example lines only: the router renders the 'Calibration examples:' heading above them, after the per-tier bullets. Requires an LLM classifier and cannot be combined with classifier_llm_config.system_prompt. With built-in tiers the rubric preset still supplies the tier criteria and, unless classification_prompt replaces them, the classification instructions; a custom tier set ships no examples of its own, so the section renders only when this is set. @@ -35589,7 +35642,7 @@ export interface components { * @enum {string} */ classifier_fallback: "heuristic" | "default_model"; - /** @description Configuration for the LLM classifier; required when classifier_type is 'llm', 'heuristic_first' or 'hybrid' */ + /** @description Configuration for the LLM classifier; required when classifier_type is 'llm', 'capability', 'heuristic_first' or 'hybrid' */ classifier_llm_config?: components["schemas"]["ClassifierLLMConfig"] | null; /** * Classifier Plugin @@ -35604,11 +35657,11 @@ export interface components { classifier_plugin_timeout_ms: number; /** * Classifier Type - * @description Classification strategy: local regex/keyword scoring, the bundled trained four-tier heuristic, an LLM call, a custom classifier plugin, 'heuristic_first', which scores locally and only pays for the LLM classifier when the local scorer does not confidently land a cheap tier, or 'hybrid', which trusts the local scorer everywhere except when its score lands near a tier boundary + * @description Classification strategy: local regex/keyword scoring, the bundled trained four-tier heuristic, an LLM tier-selection call, a Switchyard-compatible capability forecast, a custom classifier plugin, 'heuristic_first', which scores locally and only pays for the LLM classifier when the local scorer does not confidently land a cheap tier, or 'hybrid', which trusts the local scorer everywhere except when its score lands near a tier boundary * @default heuristic * @enum {string} */ - classifier_type: "heuristic" | "heuristic_v2" | "llm" | "custom" | "heuristic_first" | "hybrid"; + classifier_type: "heuristic" | "heuristic_v2" | "llm" | "capability" | "custom" | "heuristic_first" | "hybrid"; /** * Code Keywords * @description Keywords indicating code-related content @@ -36957,11 +37010,25 @@ export interface components { * Cause * @enum {string} */ - cause?: "heuristic_scorer" | "heuristic_v2" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "hybrid_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "modality_escalation" | "modality_pin_override" | "health_failover" | "health_default_fallback" | "session_affinity_pin" | "session_affinity_escalation" | "user_turn_continuation" | "default_fallback" | "keyword" | "quality_tier" | "bandit"; + cause?: "heuristic_scorer" | "heuristic_v2" | "reasoning_override" | "llm_classifier" | "capability_classifier" | "heuristic_first_short_circuit" | "hybrid_short_circuit" | "classifier_plugin" | "classifier_fallback" | "capability_classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "modality_escalation" | "modality_pin_override" | "health_failover" | "health_default_fallback" | "session_affinity_pin" | "session_affinity_escalation" | "user_turn_continuation" | "default_fallback" | "keyword" | "quality_tier" | "bandit"; + /** Classifier Calibrated P Solve */ + classifier_calibrated_p_solve?: number; + /** Classifier Calibration Version */ + classifier_calibration_version?: string; + /** Classifier Capability Boundary */ + classifier_capability_boundary?: string; /** Classifier Cost */ classifier_cost?: number; + /** Classifier Crux */ + classifier_crux?: string; /** Classifier Model */ classifier_model?: string; + /** Classifier P Solve */ + classifier_p_solve?: number; + /** Classifier Primary Rule */ + classifier_primary_rule?: string; + /** Classifier Threshold */ + classifier_threshold?: number; /** Context Escalated */ context_escalated?: boolean; /** Context Escalation Original Tier */