diff --git a/litellm/llms/parallel_ai/decisions/transformation.py b/litellm/llms/parallel_ai/decisions/transformation.py index 3fb2fc89374..80823f026b2 100644 --- a/litellm/llms/parallel_ai/decisions/transformation.py +++ b/litellm/llms/parallel_ai/decisions/transformation.py @@ -1,4 +1,4 @@ -"""Parallel's web-grounded Decisions use System One questions with mode-based routing. +"""Parallel's decision model uses System One questions and the native fast mode. Both LiteLLM Decisions routes share this adapter. Citations describe the whole request, so they stay on the response rather than being assigned to answers. @@ -53,12 +53,14 @@ class ParallelAIDecisionsConfig(BaseDecisionsConfig): def transform_decisions_request( self, model: str, request: DecisionsIRRequest, custom_llm_provider: str ) -> Mapping[str, object] | UnsupportedDecisionsRequest: + if model != "decision": + return UnsupportedDecisionsRequest(reason="Parallel Decisions supports the decision model") if len(request.questions) > 100: return UnsupportedDecisionsRequest(reason="Parallel Decisions supports at most 100 questions") body: Final = ir_to_systemone_request(model, request) if isinstance(body, UnsupportedDecisionsRequest): return body - return {"mode": model, **{key: value for key, value in body.items() if key != "model"}} + return {"mode": "fast", **{key: value for key, value in body.items() if key != "model"}} def parse_response(self, payload: object, request: DecisionsIRRequest) -> DecisionsIRResponse: response: Final = _ParallelResponse.model_validate(payload) @@ -66,7 +68,7 @@ class ParallelAIDecisionsConfig(BaseDecisionsConfig): return super().parse_response( { **response.model_dump(), - "model": response.mode, + "model": "decision", "answers": {key: answer for key, answer in response.answers.items() if answer.get("type") != "refusal"}, }, request, diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 7e159f92341..97185550bf9 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -46248,7 +46248,7 @@ "supports_response_schema": true, "supports_tool_choice": false }, - "parallel_ai/fast": { + "parallel_ai/decision": { "input_cost_per_query": 0.001, "input_cost_per_token": 0.0, "output_cost_per_token": 0.0, diff --git a/litellm/proxy/public_endpoints/provider_create_fields.json b/litellm/proxy/public_endpoints/provider_create_fields.json index 918e6e79ea0..0dbccff2c68 100644 --- a/litellm/proxy/public_endpoints/provider_create_fields.json +++ b/litellm/proxy/public_endpoints/provider_create_fields.json @@ -2639,7 +2639,7 @@ "default_value": null } ], - "default_model_placeholder": "parallel_ai/fast" + "default_model_placeholder": "parallel_ai/decision" }, { "provider": "NEBIUS", diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 7e159f92341..97185550bf9 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -46248,7 +46248,7 @@ "supports_response_schema": true, "supports_tool_choice": false }, - "parallel_ai/fast": { + "parallel_ai/decision": { "input_cost_per_query": 0.001, "input_cost_per_token": 0.0, "output_cost_per_token": 0.0, diff --git a/tests/unit/llms/parallel_ai/decisions/test_transformation.py b/tests/unit/llms/parallel_ai/decisions/test_transformation.py index 96b29cfadec..d3a4d9b3cc3 100644 --- a/tests/unit/llms/parallel_ai/decisions/test_transformation.py +++ b/tests/unit/llms/parallel_ai/decisions/test_transformation.py @@ -62,7 +62,7 @@ async def test_sdk_sends_parallel_wire_format_and_preserves_evidence( monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) litellm.in_memory_llm_clients_cache.flush_cache() route: Final = respx_mock.post("https://api.parallel.ai/v1/decisions").respond(json=PAYLOAD) - connection: Final = {"model": "parallel_ai/fast", "api_key": "test-key", "timeout": 5, "num_retries": 0} + connection: Final = {"model": "parallel_ai/decision", "api_key": "test-key", "timeout": 5, "num_retries": 0} fields: Final = ( {"state": "Example", "questions": QUESTIONS} if surface == "systemone" @@ -81,26 +81,27 @@ async def test_sdk_sends_parallel_wire_format_and_preserves_evidence( assert body["state"] == "Example" assert sent.headers["x-api-key"] == "test-key" assert "authorization" not in sent.headers + assert response.model == "decision" assert response.model_dump()["citations"] == PAYLOAD["citations"] assert response.model_dump()["decision_id"] == PAYLOAD["decision_id"] await GLOBAL_LOGGING_WORKER.clear_queue() def test_provider_routes_systemone_questions_to_parallel_without_a_model_field() -> None: - config: Final = ProviderConfigManager.get_provider_decisions_config("fast", LlmProviders.PARALLEL_AI) + config: Final = ProviderConfigManager.get_provider_decisions_config("decision", LlmProviders.PARALLEL_AI) assert config is not None - assert config.transform_decisions_request("fast", REQUEST, "parallel_ai") == { + assert config.transform_decisions_request("decision", REQUEST, "parallel_ai") == { "mode": "fast", "state": {"product": "Example"}, "questions": QUESTIONS, } - assert config.get_complete_url("https://gateway.example/v1/", "fast") == "https://gateway.example/v1/decisions" + assert config.get_complete_url("https://gateway.example/v1/", "decision") == "https://gateway.example/v1/decisions" def test_auth_replaces_conflicting_headers_without_forwarding_bearer_credentials() -> None: headers: Final = ParallelAIDecisionsConfig().validate_environment( {"Authorization": "Bearer old", "X-API-Key": "old", "CONTENT-TYPE": "text/plain", "X-Trace": "test"}, - "fast", + "decision", "test-key", {}, ) @@ -116,7 +117,7 @@ def test_openai_predicate_is_translated_to_a_named_noul_question() -> None: } ) ) - assert ParallelAIDecisionsConfig().transform_decisions_request("fast", request, "parallel_ai") == { + assert ParallelAIDecisionsConfig().transform_decisions_request("decision", request, "parallel_ai") == { "mode": "fast", "state": "Example product", "questions": {"available": {"type": "noul", "instructions": "Available?"}}, @@ -130,7 +131,7 @@ def test_all_answers_usage_and_evidence_survive_response_translation(surface: st formatted: Final = ( ir_to_systemone_response(response, REQUEST) if surface == "systemone" - else ir_to_openai_response(response, REQUEST, "parallel_ai/fast") + else ir_to_openai_response(response, REQUEST, "parallel_ai/decision") ) result: Final = formatted.model_dump(mode="json") assert result["citations"] == PAYLOAD["citations"] @@ -157,7 +158,15 @@ def test_rejects_more_than_one_hundred_questions() -> None: ) ) assert isinstance( - ParallelAIDecisionsConfig().transform_decisions_request("fast", request, "parallel_ai"), + ParallelAIDecisionsConfig().transform_decisions_request("decision", request, "parallel_ai"), + UnsupportedDecisionsRequest, + ) + + +@pytest.mark.parametrize("model", ["fast", "unknown"]) +def test_rejects_unregistered_models_instead_of_forwarding_them_as_modes(model: str) -> None: + assert isinstance( + ParallelAIDecisionsConfig().transform_decisions_request(model, REQUEST, "parallel_ai"), UnsupportedDecisionsRequest, ) @@ -173,7 +182,7 @@ def test_optional_confidence_is_preserved_without_inventing_a_value(confidence: formatted: Final = ( ir_to_systemone_response(response, REQUEST) if surface == "systemone" - else ir_to_openai_response(response, REQUEST, "parallel_ai/fast") + else ir_to_openai_response(response, REQUEST, "parallel_ai/decision") ) answers: Final = formatted.model_dump(mode="json")["answers"] values: Final = answers.values() if isinstance(answers, dict) else answers @@ -183,8 +192,8 @@ def test_optional_confidence_is_preserved_without_inventing_a_value(confidence: def test_provider_extras_cannot_replace_normalized_response_fields() -> None: response: Final = ParallelAIDecisionsConfig().parse_response(PAYLOAD, REQUEST) with_collision: Final = replace(response, extra={**response.extra, "model": "wrong", "answers": [], "usage": {}}) - formatted: Final = ir_to_openai_response(with_collision, REQUEST, "parallel_ai/fast") - assert formatted.model == "fast" + formatted: Final = ir_to_openai_response(with_collision, REQUEST, "parallel_ai/decision") + assert formatted.model == "decision" assert len(formatted.answers) == 4 assert formatted.usage.input_tokens == 4212 assert formatted.model_dump()["citations"] == PAYLOAD["citations"] @@ -193,9 +202,9 @@ def test_provider_extras_cannot_replace_normalized_response_fields() -> None: @pytest.mark.parametrize("tokens", [0, 100, 32000]) @pytest.mark.parametrize("call_type", ["decisions", "adecisions", "systemone", "asystemone"]) def test_request_pricing_is_independent_of_token_count(tokens: int, call_type: str) -> None: - rate: Final = litellm.model_cost["parallel_ai/fast"]["input_cost_per_query"] + rate: Final = litellm.model_cost["parallel_ai/decision"]["input_cost_per_query"] cost: Final = litellm.cost_per_token( - model="parallel_ai/fast", + model="parallel_ai/decision", prompt_tokens=tokens, completion_tokens=0, call_type=call_type, @@ -209,5 +218,5 @@ def test_request_pricing_is_independent_of_token_count(tokens: int, call_type: s def test_malformed_provider_responses_raise_a_provider_error(payload: Mapping[str, object]) -> None: with pytest.raises(DecisionsResponseValidationError): ParallelAIDecisionsConfig().transform_decisions_response( - "fast", "parallel_ai", httpx.Response(200, json=payload), REQUEST + "decision", "parallel_ai", httpx.Response(200, json=payload), REQUEST ) diff --git a/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx b/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx index b16d22e6fc3..caeb0645fc9 100644 --- a/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx +++ b/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx @@ -552,12 +552,12 @@ describe("provider_info_helpers", () => { it("lists Parallel decision models with their provider and model placeholder", () => { const modelMap = { - "parallel_ai/fast": { litellm_provider: "parallel_ai", mode: "evaluation" }, + "parallel_ai/decision": { litellm_provider: "parallel_ai", mode: "evaluation" }, }; - expect(getProviderModels("ParallelAI", modelMap)).toEqual(["parallel_ai/fast"]); + expect(getProviderModels("ParallelAI", modelMap)).toEqual(["parallel_ai/decision"]); expect(getProviderLogoAndName("parallel_ai").displayName).toBe("Parallel AI"); - expect(getPlaceholder("ParallelAI")).toBe("parallel_ai/fast"); + expect(getPlaceholder("ParallelAI")).toBe("parallel_ai/decision"); }); it("lists Nace decision models and resolves the provider name and model placeholder", () => { diff --git a/ui/litellm-dashboard/src/components/provider_info_helpers.tsx b/ui/litellm-dashboard/src/components/provider_info_helpers.tsx index 543b931988f..beb9641e222 100644 --- a/ui/litellm-dashboard/src/components/provider_info_helpers.tsx +++ b/ui/litellm-dashboard/src/components/provider_info_helpers.tsx @@ -502,7 +502,7 @@ const providerPlaceholderMap: Partial> = { [Providers.Tencent]: "tencent/deepseek-v4-pro", [Providers.TypeSafe]: "typesafe/jev-latest", [Providers.Nace]: "nace/drex-latest", - [Providers.ParallelAI]: "parallel_ai/fast", + [Providers.ParallelAI]: "parallel_ai/decision", [Providers.StreamLake]: "streamlake/", [Providers.Vertex_AI]: "gemini-pro", [Providers.VolcEngine]: "volcengine/",