From d8ca43a800d2c814655981e2c1a565a4a029c9e9 Mon Sep 17 00:00:00 2001 From: tin-berri Date: Fri, 4 Sep 2026 18:59:50 -0700 Subject: [PATCH] feat(complexity_router): let the LLM classifier see request images (#39825) The classifier scores extracted text, so a turn whose complexity lives in its image is invisible to it: a screenshot of a stack trace classifies on its caption, and an image-only turn flattens to empty text and never reaches the classifier at all. classifier_llm_config.vision opts in, off by default, with max_images bounding what one turn can add. Images are still dropped when the classifier model is declared supports_vision false. Anthropic and Responses image parts are rewritten into chat-completions dialect before they reach the classifier call, since /v1/messages hands the pre-routing hook its own dialect untranslated. The local scorer no longer short-circuits heuristic_first or hybrid on a turn carrying forwarded images, because it reads text alone and its confidence describes a request it has only partly seen. --- .../prompt_templates/common_utils.py | 39 +++ .../adapters/transformation.py | 16 +- .../complexity_router/complexity_router.py | 144 +++++++-- .../complexity_router/config.py | 35 +++ .../router_strategy/test_complexity_router.py | 274 ++++++++++++++++++ .../build_complexity_router_config.test.ts | 6 + .../build_complexity_router_config.ts | 11 +- ui/litellm-dashboard/src/lib/http/schema.d.ts | 24 ++ 8 files changed, 507 insertions(+), 42 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index 17ebde83eee..00b80839dde 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -229,6 +229,45 @@ def _content_parts_contain_image(parts: Sequence[object]) -> bool: return False +def anthropic_image_source_to_openai_url(image_source: Mapping[str, object]) -> str | None: + """Data or remote URL for an Anthropic ``source`` block, in the form chat completions expects.""" + source_type: Final = image_source.get("type") + if source_type == "base64": + media_type: Final = image_source.get("media_type") or "image/jpeg" + image_data: Final = image_source.get("data") or "" + return f"data:{media_type};base64,{image_data}" if image_data else None + if source_type == "url": + url: Final = image_source.get("url") + return url if isinstance(url, str) else "" + return None + + +def _image_part_url(part: Mapping[str, object]) -> str | None: + """The image URL carried by one content part, whichever of the three dialects wrote it.""" + part_type: Final = part.get("type") + if part_type == "image_url": + image_url: Final = part.get("image_url") + if isinstance(image_url, str): + return image_url + return image_url.get("url") if isinstance(image_url, Mapping) else None + if part_type == "input_image": + responses_url: Final = part.get("image_url") + return responses_url if isinstance(responses_url, str) else None + if part_type == "image": + source: Final = part.get("source") + return anthropic_image_source_to_openai_url(source) if isinstance(source, Mapping) else None + return None + + +def as_openai_image_part(part: Mapping[str, object]) -> ChatCompletionImageObject | None: + """One image content part rewritten into chat-completions dialect, or None when it is not one. + + Rebuilt rather than forwarded so no caller-controlled key beyond the URL rides along. + """ + url: Final = _image_part_url(part) + return {"type": "image_url", "image_url": {"url": url}} if url else None + + def request_contains_image_content(messages: Sequence[Mapping[str, object]]) -> bool: """Whether any message carries an image content part, across the dialects that reach pre-routing hooks untranslated: chat-completions ``image_url``, Responses ``input_image``, diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index 64f10046109..594fac512e6 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -99,6 +99,7 @@ def create_tool_name_mapping( from openai.types.chat.chat_completion_chunk import Choice as OpenAIStreamingChoice from litellm.litellm_core_utils.prompt_templates.common_utils import ( + anthropic_image_source_to_openai_url, parse_tool_call_arguments, reasoning_content_from_thinking_blocks, with_prompt_cache_breakpoint, @@ -1225,20 +1226,7 @@ class LiteLLMAnthropicMessagesAdapter: """ if not isinstance(image_source, dict): return None - - source_type: Final = image_source.get("type") - - if source_type == "base64": - # Base64 image format - media_type: Final = image_source.get("media_type", "image/jpeg") - image_data: Final = image_source.get("data", "") - if image_data: - return f"data:{media_type};base64,{image_data}" - elif source_type == "url": - # URL-referenced image format - return image_source.get("url", "") - - return None + return anthropic_image_source_to_openai_url(image_source) def _tool_result_content(self, raw_content: object) -> ToolResultContent: if isinstance(raw_content, str): diff --git a/litellm/router_strategy/complexity_router/complexity_router.py b/litellm/router_strategy/complexity_router/complexity_router.py index ec11c16eea0..98a1eb7ac9e 100644 --- a/litellm/router_strategy/complexity_router/complexity_router.py +++ b/litellm/router_strategy/complexity_router/complexity_router.py @@ -40,7 +40,10 @@ from litellm.litellm_core_utils.core_helpers import ( get_metadata_variable_name_from_kwargs, ) from litellm.litellm_core_utils.internal_call_metadata import forwarded_internal_call_metadata -from litellm.litellm_core_utils.prompt_templates.common_utils import request_contains_image_content +from litellm.litellm_core_utils.prompt_templates.common_utils import ( + as_openai_image_part, + request_contains_image_content, +) from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload from litellm.llms.base_llm.base_utils import type_to_response_format_param from litellm.router_strategy.adaptive_router.classifier import classify_prompt @@ -48,7 +51,11 @@ from litellm.router_strategy.complexity_router.tier_predictor import ( TierSuccessPredictor, resolve_tier_artifact, ) -from litellm.types.llms.openai import AllMessageValues +from litellm.types.llms.openai import ( + AllMessageValues, + ChatCompletionImageObject, + ChatCompletionTextObject, +) from litellm.types.utils import ( AUTOROUTER_CLASSIFIER_CALL_ORIGIN, ModelResponse, @@ -435,6 +442,23 @@ def _strip_reminder_blocks(text: str, marker_pairs: tuple[tuple[str, str], ...] return " ".join(kept for a, b in zip(keep_from, keep_to) if (kept := text[a:b].strip())) +def _inline_image_part(part: Mapping[str, object]) -> ChatCompletionImageObject | None: + """One image content part safe to hand the classifier, or None. + + Inline data URIs only. A remote URL is caller-controlled and provider adapters do not uniformly + delegate fetching to the provider: gigachat's file handler downloads any non-data URL with + `client.get` from the proxy host, so forwarding one would let a key scoped to this router aim a + proxy-side request at an internal address, on a call the caller never asked for. The routed + model still receives the original URL exactly as before. + """ + converted: Final = as_openai_image_part(part) + if converted is None: + return None + image_url: Final = converted["image_url"] + url: Final = image_url if isinstance(image_url, str) else image_url.get("url", "") + return converted if url.startswith("data:") else None + + def _human_text(content: object, marker_pairs: tuple[tuple[str, str], ...] = _DEFAULT_REMINDER_MARKERS) -> str: """Message content as the text a human wrote, with complete reminder blocks removed. @@ -1592,6 +1616,10 @@ class ComplexityRouter(CustomLogger): threshold check alone would hand that traffic to the cheapest model without ever consulting the classifier. Scores also go negative when simple indicators fire, so a score threshold would reject exactly the trivial prompts this path exists to serve. + + A turn carrying images the classifier would see is never decided cheaply: the scorer reads + text alone, so its confidence describes a request it has only partly seen, and a trivial + caption beside a screenshot is exactly the misrouting vision classification exists to stop. """ tier, score, signals, cause = self._score_and_classify(prompt, system_prompt) scored: Final = ClassificationOutcome(tier=tier, score=score, signals=signals, cause=cause) @@ -1599,6 +1627,7 @@ class ComplexityRouter(CustomLogger): decided_cheaply: Final = ( threshold is not None and bool(signals) + and not self._classifier_image_parts(messages) and self._active_tier_severity(tier) <= self._active_tier_severity(threshold) ) if decided_cheaply: @@ -1623,11 +1652,43 @@ class ComplexityRouter(CustomLogger): tier, score, signals, cause = self._score_and_classify(prompt, system_prompt) scored: Final = ClassificationOutcome(tier=tier, score=score, signals=signals, cause=cause) margin: Final = self.config.hybrid_boundary_margin - decided: Final = margin is not None and bool(signals) and not self._is_near_tier_boundary(score, margin) + decided: Final = ( + margin is not None + and bool(signals) + and not self._classifier_image_parts(messages) + and not self._is_near_tier_boundary(score, margin) + ) if decided: return ClassificationOutcome(tier=tier, score=score, signals=signals, cause="hybrid_short_circuit") return await self._llm_classifier_outcome(prompt, system_prompt, request_kwargs, messages, scored=scored) + def _classifier_image_parts( + self, messages: Sequence[Mapping[str, object]] | None + ) -> tuple[ChatCompletionImageObject, ...]: + """Images from the newest user turn to hand the classifier, capped by max_images. + + Empty unless the operator opted in AND the classifier model is declared vision-capable, so + every other deployment keeps today's text-only payload byte for byte. Only the newest user + turn is read: earlier turns are context the classifier already gets as quoted text, and an + image nested in a tool_result is tool output rather than the ask being classified. + Remote-URL images are left out entirely; `_inline_image_part` carries why. + """ + llm_config: Final = self.config.classifier_llm_config + if llm_config is None or not llm_config.vision.enabled or not self.config.uses_llm_classifier or not messages: + return () + if not self._model_declares_vision_support(llm_config.model): + return () + newest_user_turn: Final = next((msg for msg in reversed(messages) if msg.get("role") == "user"), None) + content: Final = newest_user_turn.get("content") if newest_user_turn is not None else None + if not isinstance(content, list): + return () + return tuple( + islice( + (part for raw in content if isinstance(raw, Mapping) and (part := _inline_image_part(raw)) is not None), + llm_config.vision.max_images, + ) + ) + async def _llm_classifier_outcome( self, prompt: str, @@ -1865,9 +1926,18 @@ class ComplexityRouter(CustomLogger): } 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 + {"type": "text", "text": user_payload}, + *image_parts, + ] + if image_parts + else user_payload + ) messages_for_call: Final[list[AllMessageValues]] = [ # mutable-ok: SDK request payload list is built once {"role": "system", "content": classifier_system_prompt}, - {"role": "user", "content": user_payload}, + {"role": "user", "content": user_content}, ] response_format: Final = classifier_response_format classifier_call_params: Mapping[str, str] = EMPTY_MAPPING @@ -2558,31 +2628,53 @@ class ComplexityRouter(CustomLogger): return pinned_model return self.get_model_for_tier(escalated_tier) - def _model_accepts_image_input(self, model_name: str) -> bool: - """Whether a routed model or pool entry can serve an image request. + def _vision_verdicts(self, model_name: str) -> tuple[bool | None, ...]: + """Declared vision support per deployment serving the name: True, False, or None when + nothing declares either way. Resolved through the deployments that would actually serve the name; a name with no deployment on the router is served by the SDK directly and is checked against the model - cost map itself. Only an explicit supports_vision false excludes, a deployment-level - model_info override first and the map otherwise, so unmapped custom names stay routable. + cost map itself. A deployment-level model_info override wins over the map. + + One verdict set, two readings, because the two callers fail in opposite directions. + Routing a user's image asks whether anything RULES IT OUT, so an undeclared model stays + eligible and unmapped custom names keep routing. Handing an image to the classifier asks + whether something RULES IT IN: an undeclared model that turns out to be text-only rejects + every image request, and that rejection is swallowed by the classifier's own fallback, so + the router quietly serves all image traffic from the fallback tier and pays for the failed + call each time. An undeclared model instead keeps today's text-only payload, which is a + visible no-op the operator fixes by declaring supports_vision on the deployment. + """ + from litellm.utils import is_vision_explicitly_disabled, supports_vision + + def model_verdict(model: str) -> bool | None: + if supports_vision(model): + return True + return False if is_vision_explicitly_disabled(model) else None + + def deployment_verdict(deployment: Mapping[str, Any]) -> bool | None: + declared: Final = (deployment.get("model_info") or EMPTY_MAPPING).get("supports_vision") + if declared is not None: + return declared is True + return model_verdict((deployment.get("litellm_params") or EMPTY_MAPPING).get("model") or model_name) + + deployments: Final = self.litellm_router_instance.get_model_list(model_name=model_name) + if not deployments: + return (model_verdict(model_name),) + return tuple(deployment_verdict(deployment) for deployment in deployments) + + def _model_accepts_image_input(self, model_name: str) -> bool: + """Whether a routed model or pool entry can serve an image request. A multi-deployment group must accept on EVERY deployment: the router picks a deployment inside the group after this gate runs, so a mixed group marked eligible could still hand the image to its text-only member and fail with the exact 400 the gate exists to prevent. """ - from litellm.utils import is_vision_explicitly_disabled + return all(verdict is not False for verdict in self._vision_verdicts(model_name)) - def deployment_accepts(deployment: Mapping[str, Any]) -> bool: - declared: Final = (deployment.get("model_info") or EMPTY_MAPPING).get("supports_vision") - if declared is not None: - return declared is True - litellm_model: Final = (deployment.get("litellm_params") or EMPTY_MAPPING).get("model") or model_name - return not is_vision_explicitly_disabled(litellm_model) - - deployments: Final = self.litellm_router_instance.get_model_list(model_name=model_name) - if not deployments: - return not is_vision_explicitly_disabled(model_name) - return all(deployment_accepts(deployment) for deployment in deployments) + def _model_declares_vision_support(self, model_name: str) -> bool: + """Whether every deployment serving the name is declared vision-capable.""" + return all(verdict is True for verdict in self._vision_verdicts(model_name)) def _modality_eligible_models(self) -> frozenset[str]: """Every configured pool entry, plus default_model, that can serve an image request.""" @@ -3374,8 +3466,9 @@ class ComplexityRouter(CustomLogger): has_original_messages: Final = messages is not None and len(messages) > 0 user_message, system_prompt = _extract_current_ask_and_system_prompt(resolved_messages, self._reminder_markers) + classifier_images: Final = self._classifier_image_parts(resolved_messages) - if user_message is None: + if user_message is None and not classifier_images: verbose_router_logger.debug("ComplexityRouter: No user message found, routing to default model") default_model_first: Final = not self.config.plugins and self.config.default_model if default_model_first: @@ -3402,6 +3495,7 @@ class ComplexityRouter(CustomLogger): ), ) + ask: Final = user_message or "" newest_ask: Final = _newest_turn_ask(resolved_messages, self._reminder_markers) escalation_keyword: Final = self._matched_escalation_keyword(newest_ask) if newest_ask is not None else None # Resolved here rather than beside the classifier because the keyword-override path below @@ -3439,7 +3533,7 @@ class ComplexityRouter(CustomLogger): ), ) - override: Final = await self._resolve_keyword_tier_override(user_message, request_kwargs) + override: Final = await self._resolve_keyword_tier_override(ask, request_kwargs) if override is not None: keyword_bumped_tier: Final = ( self._escalate_tier(override.tier) if escalation_keyword is not None else override.tier @@ -3486,9 +3580,7 @@ class ComplexityRouter(CustomLogger): outcome: Final = ( ClassificationOutcome(tier=housekeeping_tier, score=None, signals=("housekeeping",), cause="housekeeping") if housekeeping_tier is not None - else await self.aclassify( - user_message, system_prompt, request_kwargs, resolved_messages, raw_messages=messages - ) + else await self.aclassify(ask, system_prompt, request_kwargs, resolved_messages, raw_messages=messages) ) tier, score, signals = outcome.tier, outcome.score, outcome.signals classified_tier: Final = tier @@ -3558,7 +3650,7 @@ class ComplexityRouter(CustomLogger): # under is not a floor. routed_model = self._soft_floor_pick( tier, - user_message, + ask, request_kwargs, hard_floor=tier if context_original_tier is not None else plan_floor, hard_ceiling=housekeeping_ceiling, diff --git a/litellm/router_strategy/complexity_router/config.py b/litellm/router_strategy/complexity_router/config.py index 5848973a9c7..c483a0b7073 100644 --- a/litellm/router_strategy/complexity_router/config.py +++ b/litellm/router_strategy/complexity_router/config.py @@ -442,12 +442,47 @@ DEFAULT_TIER_MODELS: Final[dict[str, str]] = { } +class ClassifierVisionConfig(BaseModel): + """Whether the LLM classifier sees the images on the request it is classifying. + + Off by default because images cost far more than the text ask they arrive with, and the + classifier runs on every request. A turn whose complexity lives in the image ("what is wrong in + this stack trace screenshot") is invisible to a text-only classifier, which is what this buys. + """ + + enabled: bool = Field( + default=False, + description=( + "Forward image content to the classifier. Requires a classifier model declared " + "supports_vision, on the deployment's model_info or in the model cost map; images stay " + "stripped otherwise, so a classifier that cannot read them is never sent one. Declare " + "model_info.supports_vision on the deployment to enable a model the cost map does not " + "describe. Only inline data: URIs are forwarded. A request whose images are http(s) " + "URLs still classifies on its text alone, because some providers fetch such a URL from " + "the proxy rather than the provider, which would let a caller aim a proxy-side request " + "at an address of their choosing." + ), + ) + max_images: int = Field( + default=1, + ge=1, + description=( + "How many images from the newest user turn to forward, in wire order. Bounds the added " + "cost of a turn that attaches many images. Images on earlier turns are never forwarded." + ), + ) + + class ClassifierLLMConfig(BaseModel): """Configuration for the LLM-based complexity classifier.""" model: str = Field( description="Model name (from the router's model_list) to call for classification", ) + vision: ClassifierVisionConfig = Field( + default_factory=ClassifierVisionConfig, + description="Whether the classifier sees images on the request, and how many", + ) reasoning_effort: REASONING_EFFORT | None = Field( default=None, description=( diff --git a/tests/test_litellm/router_strategy/test_complexity_router.py b/tests/test_litellm/router_strategy/test_complexity_router.py index 8de138688f1..52e58304476 100644 --- a/tests/test_litellm/router_strategy/test_complexity_router.py +++ b/tests/test_litellm/router_strategy/test_complexity_router.py @@ -12429,3 +12429,277 @@ class TestTierHealthFailover: for _ in range(20) ] assert {r.model for r in results} == {"live-c"} + + +ANTHROPIC_IMG_PART = {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "aGk="}} +RESPONSES_IMG_PART = {"type": "input_image", "image_url": "data:image/png;base64,aGk="} + + +class TestClassifierVision: + """classifier_llm_config.vision: what the LLM classifier is shown for an image-bearing turn.""" + + TIERS = {"SIMPLE": "t-simple", "MEDIUM": "t-medium", "COMPLEX": "t-complex", "REASONING": "t-reasoning"} + + @staticmethod + def _router(mock_router_instance, *, vision, classifier_declares_vision=True, classifier_type="llm", **extra): + def get_model_list(model_name=None): + if model_name != "clf": + return [{"model_name": model_name, "litellm_params": {"model": "openai/gpt-4o"}}] + declared = classifier_declares_vision + return [ + { + "model_name": "clf", + "litellm_params": {"model": "openai/unmapped-classifier"}, + "model_info": {} if declared is None else {"supports_vision": declared}, + } + ] + + mock_router_instance.get_model_list = get_model_list + classifier_llm_config = {"model": "clf", "circuit_breaker_enabled": False} + return ComplexityRouter( + model_name="vision-classifier-router", + litellm_router_instance=mock_router_instance, + complexity_router_config={ + "classifier_type": classifier_type, + "classifier_llm_config": ( + classifier_llm_config if vision is None else {**classifier_llm_config, "vision": vision} + ), + "tiers": dict(TestClassifierVision.TIERS), + **extra, + }, + ) + + @staticmethod + def _classifier_user_content(mock_router_instance): + return mock_router_instance.acompletion.call_args.kwargs["messages"][-1]["content"] + + @staticmethod + def _turn(*parts): + return [{"role": "user", "content": list(parts)}] + + @pytest.fixture(autouse=True) + def _classifier_answers_complex(self, mock_router_instance): + mock_router_instance.acompletion = AsyncMock(return_value=_llm_response('{"tier": "COMPLEX"}')) + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "vision, classifier_declares_vision", + [ + (None, True), + ({"enabled": False}, True), + ({"enabled": True}, False), + ({"enabled": True}, None), + ], + ids=["vision_unset", "vision_disabled", "classifier_declared_text_only", "classifier_undeclared"], + ) + async def test_payload_stays_text_only(self, mock_router_instance, vision, classifier_declares_vision): + """Off, or a classifier not declared vision-capable, keeps the plain-string payload. + + The undeclared case is the polarity. A text-only classifier handed an image rejects the + call, the rejection is swallowed by the classifier's own fallback, and every image request + then serves from the fallback tier while still paying for the failed call. Staying text-only + is instead a visible no-op the operator fixes by declaring supports_vision. + """ + router = self._router( + mock_router_instance, vision=vision, classifier_declares_vision=classifier_declares_vision + ) + await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "what is this"}, IMG_PART) + ) + content = self._classifier_user_content(mock_router_instance) + assert isinstance(content, str) + assert "what is this" in content + + @pytest.mark.asyncio + async def test_deployment_model_info_enables_a_classifier_the_cost_map_does_not_describe( + self, mock_router_instance + ): + """The escape hatch for an unmapped classifier name, and the reason undeclared can stay off. + + `_router` gives every deployment an `openai/unmapped-*` litellm_params model, so nothing in + the cost map declares it and the verdict comes only from model_info. + """ + router = self._router(mock_router_instance, vision={"enabled": True}, classifier_declares_vision=True) + await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "what is this"}, IMG_PART) + ) + assert [b["type"] for b in self._classifier_user_content(mock_router_instance)] == ["text", "image_url"] + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "part", + [IMG_PART, ANTHROPIC_IMG_PART, RESPONSES_IMG_PART], + ids=["chat_completions", "anthropic_messages", "responses"], + ) + async def test_image_reaches_the_classifier_in_chat_completions_dialect(self, mock_router_instance, part): + """Every surface's dialect arrives as a chat-completions image_url on the classifier call. + + /v1/messages hands the hook an Anthropic image block untranslated, so forwarding verbatim + would send the classifier a content part its own request dialect has no meaning for. + """ + router = self._router(mock_router_instance, vision={"enabled": True}) + await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "what is this"}, part) + ) + content = self._classifier_user_content(mock_router_instance) + assert [block["type"] for block in content] == ["text", "image_url"] + assert content[1]["image_url"] == {"url": "data:image/png;base64,aGk="} + assert "what is this" in content[0]["text"] + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "part", + [ + {"type": "image_url", "image_url": {"url": "http://169.254.169.254/latest/meta-data/"}}, + {"type": "image_url", "image_url": {"url": "https://example.internal/secret.png"}}, + {"type": "input_image", "image_url": "https://example.internal/secret.png"}, + {"type": "image", "source": {"type": "url", "url": "https://example.internal/secret.png"}}, + ], + ids=["metadata_service", "chat_completions", "responses", "anthropic"], + ) + async def test_remote_url_images_are_never_forwarded(self, mock_router_instance, part): + """A caller-supplied URL must not reach an internal call the caller did not ask for. + + Provider adapters do not uniformly delegate fetching: gigachat downloads any non-data URL + from the proxy host, so forwarding one would turn a router-scoped key into a proxy-side GET + at an address of the caller's choosing. + """ + router = self._router(mock_router_instance, vision={"enabled": True}) + await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "what is this"}, part) + ) + assert isinstance(self._classifier_user_content(mock_router_instance), str) + + @pytest.mark.asyncio + async def test_remote_url_image_only_turn_does_not_reach_the_classifier(self, mock_router_instance): + """With nothing forwardable left, the turn stays unclassifiable rather than sending the URL.""" + router = self._router(mock_router_instance, vision={"enabled": True}) + response = await router.async_pre_routing_hook( + model="m", + request_kwargs={}, + messages=self._turn({"type": "image_url", "image_url": {"url": "https://example.internal/x.png"}}), + ) + assert response.routing_decision["cause"] == "default_fallback" + mock_router_instance.acompletion.assert_not_awaited() + + @pytest.mark.asyncio + async def test_image_only_turn_is_classified_instead_of_falling_back(self, mock_router_instance): + """A turn carrying only an image reaches the classifier rather than the default model. + + It flattens to empty text, so before this it never reached the classifier at all and was + routed as default_fallback on text the request never contained. + """ + router = self._router(mock_router_instance, vision={"enabled": True}) + response = await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn(IMG_PART) + ) + assert response.routing_decision["cause"] == "llm_classifier" + assert response.model == "t-complex" + assert [block["type"] for block in self._classifier_user_content(mock_router_instance)] == [ + "text", + "image_url", + ] + + @pytest.mark.asyncio + async def test_image_only_turn_still_falls_back_when_vision_is_off(self, mock_router_instance): + router = self._router(mock_router_instance, vision={"enabled": False}) + response = await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn(IMG_PART) + ) + assert response.routing_decision["cause"] == "default_fallback" + mock_router_instance.acompletion.assert_not_awaited() + + @pytest.mark.asyncio + @pytest.mark.parametrize("max_images, expected", [(1, 1), (2, 2), (5, 3)]) + async def test_max_images_caps_what_is_forwarded(self, mock_router_instance, max_images, expected): + router = self._router(mock_router_instance, vision={"enabled": True, "max_images": max_images}) + images = [dict(IMG_PART, image_url={"url": f"data:image/png;base64,{n}"}) for n in ("a", "b", "c")] + await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "look"}, *images) + ) + content = self._classifier_user_content(mock_router_instance) + forwarded = [block for block in content if block["type"] == "image_url"] + assert len(forwarded) == expected + assert [block["image_url"]["url"] for block in forwarded] == [ + f"data:image/png;base64,{n}" for n in ("a", "b", "c")[:expected] + ] + + @pytest.mark.asyncio + async def test_earlier_turn_images_are_not_forwarded(self, mock_router_instance): + """Only the newest user turn's images ride along, so history cannot inflate every call. + + The two turns carry different images on purpose: identical ones would pass this assertion + whichever turn the helper read. + """ + older = dict(IMG_PART, image_url={"url": "data:image/png;base64,OLDER"}) + newer = dict(IMG_PART, image_url={"url": "data:image/png;base64,NEWER"}) + router = self._router(mock_router_instance, vision={"enabled": True, "max_images": 5}) + await router.async_pre_routing_hook( + model="m", + request_kwargs={}, + messages=[ + {"role": "user", "content": [{"type": "text", "text": "first"}, older]}, + {"role": "assistant", "content": "ok"}, + {"role": "user", "content": [{"type": "text", "text": "second"}, newer]}, + ], + ) + content = self._classifier_user_content(mock_router_instance) + forwarded = [block for block in content if block["type"] == "image_url"] + assert [block["image_url"]["url"] for block in forwarded] == ["data:image/png;base64,NEWER"] + + @pytest.mark.asyncio + async def test_logged_request_body_matches_what_was_sent(self, mock_router_instance): + """proxy_server_request is the logged copy of the classifier call and must not drift.""" + router = self._router(mock_router_instance, vision={"enabled": True}) + await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "what is this"}, IMG_PART) + ) + call_kwargs = mock_router_instance.acompletion.call_args.kwargs + assert call_kwargs["proxy_server_request"]["body"]["messages"] == call_kwargs["messages"] + + SHORT_CIRCUIT_ARMS = [ + ("heuristic_first", {"heuristic_first_max_tier": "SIMPLE"}, "heuristic_first_short_circuit"), + ("hybrid", {"hybrid_boundary_margin": 0.05}, "hybrid_short_circuit"), + ] + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "classifier_type, extra, short_circuit_cause", SHORT_CIRCUIT_ARMS, ids=["heuristic_first", "hybrid"] + ) + async def test_local_scorer_cannot_short_circuit_a_turn_it_cannot_see( + self, mock_router_instance, classifier_type, extra, short_circuit_cause + ): + """The scorer reads text alone, so its confidence is not a verdict on an image turn. + + Both arms are tuned so the scorer WOULD short-circuit on this exact text, which is what + makes the image the only variable; a margin loose enough to leave the score undecided + would pass whether or not the guard exists. + """ + router = self._router( + mock_router_instance, vision={"enabled": True}, classifier_type=classifier_type, **extra + ) + response = await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=self._turn({"type": "text", "text": "what is this"}, IMG_PART) + ) + assert response.routing_decision["cause"] == "llm_classifier" + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "classifier_type, extra, short_circuit_cause", SHORT_CIRCUIT_ARMS, ids=["heuristic_first", "hybrid"] + ) + async def test_local_scorer_still_short_circuits_without_images( + self, mock_router_instance, classifier_type, extra, short_circuit_cause + ): + """The negative class: same router, same text, no image, and the scorer still decides.""" + router = self._router( + mock_router_instance, vision={"enabled": True}, classifier_type=classifier_type, **extra + ) + response = await router.async_pre_routing_hook( + model="m", request_kwargs={}, messages=[{"role": "user", "content": "what is this"}] + ) + assert response.routing_decision["cause"] == short_circuit_cause + mock_router_instance.acompletion.assert_not_awaited() + + def test_max_images_must_be_positive(self): + with pytest.raises(ValidationError): + ClassifierLLMConfig(model="clf", vision={"enabled": True, "max_images": 0}) diff --git a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts index edb32c6717f..1b5bb9e72eb 100644 --- a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts +++ b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts @@ -592,6 +592,12 @@ describe("classifier prompt and fallback", () => { timeout_ms: 1, }); }); + + it.each([{}, { system_prompt: "x" }])("normalizeClassifierLlmConfig carries vision through %o", (extra) => { + const base = { model: "m", timeout_ms: 1, ...extra }; + const vision = { enabled: true, max_images: 2 }; + expect(normalizeClassifierLlmConfig({ ...base, vision })).toEqual({ ...base, vision }); + }); }); describe("tier labels", () => { 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 1a395c38779..d7974484970 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 @@ -1,4 +1,7 @@ import { KeywordTierRule } from "./KeywordTierRules"; + +type ClassifierLLMConfigWire = ClassifierLLMConfig & { vision?: { enabled?: boolean; max_images?: number } }; + import type { ModelGroup } from "../llm_calls/fetch_models"; import { type CustomTierSet, @@ -61,7 +64,8 @@ export const normalizeClassifierLlmConfig = ({ reasoning_effort, classification_rubric, system_prompt, -}: ClassifierLLMConfig): ClassifierLLMConfig => + vision, +}: ClassifierLLMConfigWire): ClassifierLLMConfigWire => system_prompt?.trim() ? { model, @@ -69,6 +73,7 @@ export const normalizeClassifierLlmConfig = ({ ...(circuit_breaker_enabled !== undefined && { circuit_breaker_enabled }), ...(circuit_breaker_cooldown_seconds !== undefined && { circuit_breaker_cooldown_seconds }), ...(reasoning_effort && { reasoning_effort }), + ...(vision && { vision }), system_prompt, } : { @@ -78,6 +83,7 @@ export const normalizeClassifierLlmConfig = ({ ...(circuit_breaker_cooldown_seconds !== undefined && { circuit_breaker_cooldown_seconds }), ...(reasoning_effort && { reasoning_effort }), ...(classification_rubric && { classification_rubric }), + ...(vision && { vision }), }; interface ScorerKnobInputs { @@ -324,7 +330,7 @@ export const getSemanticConfigError = ({ }; interface CustomTierWireFieldInputs { - classifierLlmConfig: ClassifierLLMConfig | undefined; + classifierLlmConfig: ClassifierLLMConfigWire | undefined; planModeMinTierId: string | undefined; classificationPrompt: string | undefined; classificationExamples: string | undefined; @@ -356,6 +362,7 @@ export const customTierWireFields = ( circuit_breaker_cooldown_seconds: classifierLlmConfig.circuit_breaker_cooldown_seconds, }), ...(classifierLlmConfig.reasoning_effort && { reasoning_effort: classifierLlmConfig.reasoning_effort }), + ...(classifierLlmConfig.vision && { vision: classifierLlmConfig.vision }), }, }), session_affinity: false, diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 6f60905d4c1..45e1b1cab0b 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -25305,6 +25305,30 @@ export interface components { * @default 3000 */ timeout_ms: number; + /** @description Whether the classifier sees images on the request, and how many */ + vision?: components["schemas"]["ClassifierVisionConfig"]; + }; + /** + * ClassifierVisionConfig + * @description Whether the LLM classifier sees the images on the request it is classifying. + * + * Off by default because images cost far more than the text ask they arrive with, and the + * classifier runs on every request. A turn whose complexity lives in the image ("what is wrong in + * this stack trace screenshot") is invisible to a text-only classifier, which is what this buys. + */ + ClassifierVisionConfig: { + /** + * Enabled + * @description Forward image content to the classifier. Requires a classifier model declared supports_vision, on the deployment's model_info or in the model cost map; images stay stripped otherwise, so a classifier that cannot read them is never sent one. Declare model_info.supports_vision on the deployment to enable a model the cost map does not describe. Only inline data: URIs are forwarded. A request whose images are http(s) URLs still classifies on its text alone, because some providers fetch such a URL from the proxy rather than the provider, which would let a caller aim a proxy-side request at an address of their choosing. + * @default false + */ + enabled: boolean; + /** + * Max Images + * @description How many images from the newest user turn to forward, in wire order. Bounds the added cost of a turn that attaches many images. Images on earlier turns are never forwarded. + * @default 1 + */ + max_images: number; }; /** * CloudZeroExportRequest