revenium-python-sdk 0.15.0__tar.gz → 0.16.1__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/PKG-INFO +78 -9
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/README.md +77 -8
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/pyproject.toml +1 -1
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/cache_tokens.py +103 -15
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/enforcement.py +144 -46
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/error_text.py +26 -9
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/failed_calls.py +9 -6
- revenium_python_sdk-0.16.1/revenium_middleware/_core/inference_geo.py +39 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/metering.py +4 -1
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/metering_submission.py +8 -1
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/tool_event_transport.py +11 -1
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/wire_guards.py +15 -5
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/resources/ai.py +30 -2
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/ai_create_completion_params.py +16 -1
- revenium_python_sdk-0.16.1/revenium_middleware/litellm/_failed_calls.py +40 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/middleware.py +4 -20
- revenium_python_sdk-0.16.1/revenium_middleware/litellm/proxy/asgi.py +17 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/proxy/guardrail.py +55 -7
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/proxy/middleware.py +4 -3
- revenium_python_sdk-0.16.1/revenium_middleware/litellm/proxy/shutdown.py +104 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/middleware.py +141 -14
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_python_sdk.egg-info/PKG-INFO +78 -9
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_python_sdk.egg-info/SOURCES.txt +4 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/tests/test_metering.py +70 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/LICENSE +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/call_context.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/call_ownership.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/closable_scope.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/context.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/decorators.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/delivery_circuit.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/exceptions.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/json_safe.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/langchain_settings.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/load_diagnostics.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/log_sanitize.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/metering_buffer.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/metering_pool.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/metering_status.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/outcomes.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/package_versions.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/parse_failures.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/patch_registry.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/perplexity_cost.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/perplexity_usage_metadata.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/prompt_extraction.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/responses_status.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/secret_shapes.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/selective_metering.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/shutdown_signals.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/sse_body_tap.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/subscriber.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/usage_iterations.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/vendor_model_aliases.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/LICENSE +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_base_client.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_client.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_compat.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_constants.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_exceptions.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_files.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_models.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_qs.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_resource.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_response.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_streaming.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_types.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_logs.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_proxy.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_reflection.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_resources_proxy.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_streams.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_sync.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_transform.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_typing.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_utils/_utils.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/_version.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/context.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/decorator.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/py.typed +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/resources/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/resources/apis.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/resources/events.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/ai_create_audio_params.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/ai_create_image_params.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/ai_create_video_params.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/api_meter_request_params.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/api_meter_response_params.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/event_create_params.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_metering/types/metering_response_resource.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/agentic_outcomes.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/bedrock_adapter.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/bedrock_model_names.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/bedrock_transport.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/inference_profiles.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/middleware.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/prompt_extractor.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/provider.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/raw_response.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/stream_create.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/anthropic/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/_call.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/_metering.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/_queue.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/middleware.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/fal/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/common/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/common/exceptions.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/common/protocols.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/common/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/common/types.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/common/utils.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/google_ai/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/google_ai/middleware.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/google_ai/provider.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/prompt_extractor.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/vertex_ai/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/vertex_ai/middleware.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/google/vertex_ai/provider.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/_metadata.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/_retry.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/anthropic_driver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/budget_errors.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/litellm_driver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/ollama_driver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/openai_driver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/openai_embedding_driver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/griptape/universal_driver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/job_context.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/job_history.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/job_type_economics.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/context.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/decorators.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/hooks.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/integrations/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/integrations/crewai.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/client/validation.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/proxy/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/proxy/_health_check.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/litellm/proxy/_metering_owner.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/ollama/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/ollama/middleware.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/ollama/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/audio.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/azure_config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/azure_model_resolver.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/config.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/exceptions.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/_abandoned_runs.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/_model_call.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/_perplexity_cost_capture.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/_prompt_capture.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/_text_completion_usage.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/_utils.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/langchain/unified_handler.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/prompt_extractor.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/provider.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/raw_response.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/openai/xai_usage.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/_metering_scope.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/middleware.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/patching.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/perplexity_sdk.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/provider.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/raw_response.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/streaming.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/perplexity/trace_fields.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/webhooks/__init__.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/webhooks/_verify.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_python_sdk.egg-info/dependency_links.txt +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_python_sdk.egg-info/requires.txt +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_python_sdk.egg-info/top_level.txt +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/setup.cfg +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/tests/test_metering_dispatch_timestamps.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/tests/test_patch_coverage.py +0 -0
- {revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/tests/test_release_verify_ticket.py +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: revenium-python-sdk
|
|
3
|
-
Version: 0.
|
|
3
|
+
Version: 0.16.1
|
|
4
4
|
Summary: The official Revenium Python SDK — unified AI metering middleware for OpenAI, Anthropic, Google, Ollama, LiteLLM, Perplexity, and fal.ai.
|
|
5
5
|
Author-email: Revenium <support@revenium.io>
|
|
6
6
|
License: MIT
|
|
@@ -196,10 +196,10 @@ The calls below are billed by the provider (some, such as batch and fine-tuning
|
|
|
196
196
|
<!-- not-metered:start -->
|
|
197
197
|
- **OpenAI** (`OpenAI` and `AsyncOpenAI` clients): `batches.create`, `beta.agents.automations.create`, `beta.agents.automations.invocations.create`, `beta.agents.sessions.create`, `beta.agents.sessions.events.create`, `beta.agents.sessions.events.stream`, `beta.agents.sessions.stream`, `beta.agents.sessions.update`, `beta.chatkit.sessions.create`, `beta.realtime.connect`, `beta.responses.compact`, `beta.responses.connect`, `beta.responses.create`, `beta.threads.create_and_run`, `beta.threads.create_and_run_poll`, `beta.threads.create_and_run_stream`, `beta.threads.runs.create`, `beta.threads.runs.create_and_poll`, `beta.threads.runs.create_and_stream`, `beta.threads.runs.stream`, `beta.threads.runs.submit_tool_outputs`, `beta.threads.runs.submit_tool_outputs_and_poll`, `beta.threads.runs.submit_tool_outputs_stream`, `completions.create`, `decisions.create`, `evals.runs.create`, `fine_tuning.alpha.graders.run`, `fine_tuning.jobs.create`, `images.create_variation`, `images.edit`, `images.generate`, `live.connect`, `live.create`, `live.forks.connect`, `live.sessions.accept`, `live.sessions.fork`, `live.sideband.connect`, `realtime.calls.accept`, `realtime.calls.create`, `realtime.connect`, `realtime.translations.connect`, `responses.compact`, `responses.connect`, `videos.create`, `videos.create_and_poll`, `videos.create_character`, `videos.edit`, `videos.extend`, `videos.remix`
|
|
198
198
|
- **Anthropic** (`Anthropic` and `AsyncAnthropic` clients): `beta.deployments.run`, `beta.dreams.create`, `beta.environments.work.poller`, `beta.environments.work.worker`, `beta.messages.batches.create`, `beta.sessions.create`, `beta.sessions.events.send`, `beta.sessions.events.stream`, `beta.sessions.events.tool_runner`, `beta.sessions.threads.events.stream`, `messages.batches.create`
|
|
199
|
-
- **Google Gemini** (`google.genai` `Client` and `Client.aio`): `batches.create`, `batches.create_embeddings`, `caches.create`, `caches.update`, `decisions.create`, `file_search_stores.import_file`, `file_search_stores.upload_to_file_search_store`, `interactions.create`, `live.connect`, `live.music.connect`, `models.edit_image`, `models.recontext_image`, `models.segment_image`, `models.upscale_image`, `triggers.run`, `tunings.tune`
|
|
199
|
+
- **Google Gemini** (`google.genai` `Client` and `Client.aio`): `batches.create`, `batches.create_embeddings`, `caches.create`, `caches.update`, `decisions.create`, `file_search_stores.import_file`, `file_search_stores.upload_to_file_search_store`, `interactions.connect`, `interactions.create`, `live.connect`, `live.music.connect`, `models.edit_image`, `models.recontext_image`, `models.segment_image`, `models.upscale_image`, `triggers.run`, `tunings.tune`
|
|
200
200
|
- **Google Vertex AI** (`vertexai` classes): `batch_prediction.BatchPredictionJob.submit`, `caching.CachedContent.create`, `caching.CachedContent.update`, `language_models.ChatModel.tune_model`, `language_models.ChatModel.tune_model_rlhf`, `language_models.ChatSession.send_message`, `language_models.ChatSession.send_message_async`, `language_models.ChatSession.send_message_streaming`, `language_models.ChatSession.send_message_streaming_async`, `language_models.CodeChatModel.tune_model`, `language_models.CodeChatSession.send_message`, `language_models.CodeChatSession.send_message_async`, `language_models.CodeChatSession.send_message_streaming`, `language_models.CodeChatSession.send_message_streaming_async`, `language_models.CodeGenerationModel.batch_predict`, `language_models.CodeGenerationModel.predict`, `language_models.CodeGenerationModel.predict_async`, `language_models.CodeGenerationModel.predict_streaming`, `language_models.CodeGenerationModel.predict_streaming_async`, `language_models.CodeGenerationModel.tune_model`, `language_models.TextEmbeddingModel.batch_predict`, `language_models.TextEmbeddingModel.deploy_tuned_model`, `language_models.TextEmbeddingModel.tune_model`, `language_models.TextGenerationModel.batch_predict`, `language_models.TextGenerationModel.predict`, `language_models.TextGenerationModel.predict_async`, `language_models.TextGenerationModel.predict_streaming`, `language_models.TextGenerationModel.predict_streaming_async`, `language_models.TextGenerationModel.tune_model`, `language_models.TextGenerationModel.tune_model_rlhf`, `preview.caching.CachedContent.create`, `preview.caching.CachedContent.update`, `preview.vision_models.ImageCaptioningModel.get_captions`, `preview.vision_models.ImageGenerationModel.upscale_image`, `preview.vision_models.ImageQnAModel.ask_question`, `preview.vision_models.ImageSegmentationModel.segment_image`, `preview.vision_models.ImageTextModel.ask_question`, `preview.vision_models.ImageTextModel.get_captions`, `preview.vision_models.MultiModalEmbeddingModel.get_embeddings`, `preview.vision_models.WatermarkVerificationModel.verify_image`, `tuning.sft.train`, `vision_models.ImageCaptioningModel.get_captions`, `vision_models.ImageGenerationModel.upscale_image`, `vision_models.ImageQnAModel.ask_question`, `vision_models.ImageTextModel.ask_question`, `vision_models.ImageTextModel.get_captions`, `vision_models.MultiModalEmbeddingModel.get_embeddings`
|
|
201
201
|
- **AWS Bedrock** (boto3 bedrock-runtime client; Claude models are metered when REVENIUM_BEDROCK_TRANSPORT is on): `apply_guardrail`, `converse`, `converse_stream`, `invoke_guardrail_checks`, `invoke_model`, `invoke_model_with_response_stream`, `start_async_invoke`
|
|
202
|
-
- **Perplexity** (`Perplexity` and `AsyncPerplexity` clients): `async_.chat.completions.create`, `batches.create`, `browser.sessions.create`, `contextualized_embeddings.create`, `decisions.create`, `embeddings.create`, `responses.create`, `search.create`
|
|
202
|
+
- **Perplexity** (`Perplexity` and `AsyncPerplexity` clients): `async_.chat.completions.create`, `batches.create`, `browser.sessions.create`, `context.create`, `contextualized_embeddings.create`, `decisions.create`, `embeddings.create`, `monitors.create`, `monitors.run`, `responses.create`, `search.create`
|
|
203
203
|
- **LiteLLM** (the `litellm` module): `aadapter_completion`, `aadapter_generate_content`, `acode_interpreter_tool`, `acompact_responses`, `acreate_batch`, `acreate_fine_tuning_job`, `acreate_run`, `acreate_sandbox`, `adapter_completion`, `adecisions`, `aextraction`, `agenerate_content`, `aimage_edit`, `aimage_generation`, `aimage_variation`, `aingest`, `anthropic_interface.messages.acreate`, `anthropic_interface.messages.create`, `anthropic_messages`, `anthropic_messages_handler`, `aocr`, `aquery`, `arealtime_calls`, `arerank`, `aresponses`, `aresponses_api_with_mcp`, `aresponses_with_retries`, `arun_code`, `arun_thread`, `arun_thread_stream`, `asearch`, `aspeech`, `asystemone`, `atranscription`, `avideo_create_character`, `avideo_edit`, `avideo_extension`, `avideo_generation`, `avideo_remix`, `compact_responses`, `config_completion`, `create_batch`, `create_fine_tuning_job`, `create_run`, `decisions`, `extraction`, `image_edit`, `image_generation`, `image_variation`, `ingest`, `nlp_cloud_chat_completion`, `ocr`, `query`, `replicate_chat_completion`, `rerank`, `responses`, `responses_with_retries`, `run_thread`, `run_thread_stream`, `search`, `speech`, `systemone`, `transcription`, `video_create_character`, `video_edit`, `video_extension`, `video_generation`, `video_remix`
|
|
204
204
|
- **Ollama** (`Client`, `AsyncClient` and the `ollama` module): `systemone`, `web_fetch`, `web_search`
|
|
205
205
|
- **fal.ai** (`SyncClient`, `AsyncClient` and the `fal_client` module): `realtime`, `realtime_async`, `ws_connect`, `ws_connect_async`
|
|
@@ -563,6 +563,8 @@ client.close()
|
|
|
563
563
|
|
|
564
564
|
In an `emit_completion` payload the agent is named by `agent`. `agentName` and `currency` are not sent (Revenium does not store them); `emit_completion` names them in its one-time warning about unsent keys.
|
|
565
565
|
|
|
566
|
+
`cacheCreationTokenCount` is the whole cache write. When the provider reports how that write splits by cache lifetime, send the split beside it as `cacheCreation5mTokenCount`, `cacheCreation1hTokenCount` and `cacheCreation12hTokenCount` so each part is priced at its own rate; leave a bucket out when the provider does not report it rather than sending 0. `get_client().ai.create_completion` takes the same buckets as `cache_creation5m_token_count`, `cache_creation1h_token_count` and `cache_creation12h_token_count`.
|
|
567
|
+
|
|
566
568
|
`emit_completion` and `emit_tool_event` queue the record and return at once, like every automatic record: the SDK's background threads send it with the same timeouts and retries (see [Metering Delivery and Timeouts](#metering-delivery-and-timeouts)), and a record that fails with a timeout, a refused connection or a 5xx goes to the store-and-forward buffer and is sent when Revenium is reachable again. Delivery errors are logged, not raised; `emit_completion` still raises `ValueError` before anything is queued when the payload is missing a required key. Both calls return the payload you passed.
|
|
567
569
|
|
|
568
570
|
Records go to the address set in code with `initialize_metering(base_url=...)` or a `metering_key(..., base_url=...)` block when one is set, else `REVENIUM_METERING_BASE_URL`; pass `meter_base_url=` to `AgenticOutcomeSettings` to pin another address. They are always sent under the settings' `api_key`, never a `metering_key()` key. `http_client=` serves jobs and outcomes only.
|
|
@@ -1180,7 +1182,7 @@ response = litellm.completion(
|
|
|
1180
1182
|
)
|
|
1181
1183
|
```
|
|
1182
1184
|
|
|
1183
|
-
Client mode meters `completion`, `acompletion`, `text_completion`, `atext_completion`, `completion_with_retries` and `acompletion_with_retries`, streamed or not, and `embedding` / `aembedding` as embedding calls. When you stream without `stream_options={"include_usage": True}`, the middleware asks LiteLLM for the usage chunk itself and keeps it out of your stream. LiteLLM adds that chunk for every provider, including ones it does not send `stream_options` to. `litellm.responses` and `litellm.aresponses` are not metered, except on providers LiteLLM serves through `completion`. A completion that fails, or a stream that breaks off, sends one record with stop reason `ERROR`, zero tokens for a call that never started, and the error's class and HTTP status as its error reason (plus the provider's redacted error message when both `REVENIUM_CAPTURE_ERROR_MESSAGES` and `REVENIUM_CAPTURE_PROMPTS` are `true`; a provider's error can quote the request, so it is never sent with prompt capture off); your code still gets the original exception. The HTTP status is also sent as the record's error code; a call that got no HTTP error answer (a connection error, a timeout, or a stream that broke off) has none. LiteLLM gives a connection error a 500 and a timeout a 408 that no server answered, so those get no error code and no HTTP status in their error reason, including the `InternalServerError` LiteLLM raises for a refused connection on its `openai/` and `anthropic/` providers. LiteLLM also raises a few real provider answers, such as a Vertex AI 502, as connection errors, and those get no error code either. Failure records sent by the LiteLLM gateway (proxy)
|
|
1185
|
+
Client mode meters `completion`, `acompletion`, `text_completion`, `atext_completion`, `completion_with_retries` and `acompletion_with_retries`, streamed or not, and `embedding` / `aembedding` as embedding calls. When you stream without `stream_options={"include_usage": True}`, the middleware asks LiteLLM for the usage chunk itself and keeps it out of your stream. LiteLLM adds that chunk for every provider, including ones it does not send `stream_options` to. `litellm.responses` and `litellm.aresponses` are not metered, except on providers LiteLLM serves through `completion`. A completion that fails, or a stream that breaks off, sends one record with stop reason `ERROR`, zero tokens for a call that never started, and the error's class and HTTP status as its error reason (plus the provider's redacted error message when both `REVENIUM_CAPTURE_ERROR_MESSAGES` and `REVENIUM_CAPTURE_PROMPTS` are `true`; a provider's error can quote the request, so it is never sent with prompt capture off); your code still gets the original exception. The HTTP status is also sent as the record's error code; a call that got no HTTP error answer (a connection error, a timeout, or a stream that broke off) has none. LiteLLM gives a connection error a 500 and a timeout a 408 that no server answered, so those get no error code and no HTTP status in their error reason, including the `InternalServerError` LiteLLM raises for a refused connection on its `openai/` and `anthropic/` providers. LiteLLM also raises a few real provider answers, such as a Vertex AI 502, as connection errors, and those get no error code either. Failure records sent by the LiteLLM gateway (proxy) are described under Proxy Mode below.
|
|
1184
1186
|
|
|
1185
1187
|
Inside a LiteLLM proxy running `ReveniumGuardrail` with `default_on: true` and `post_call` among its modes, client mode leaves each proxied request to the guardrail, so the request is recorded once, with the guardrail's subscriber. The row names the deployment's model (the `litellm_params.model` behind the `model_name` your callers use, without its provider prefix), so a call to an alias is priced as the model it reached. The guardrail also meters streamed `/v1/responses` calls (Codex CLI, the OpenAI Agents SDK) for every provider. Client mode still records the proxied calls the guardrail does not meter, which today are streamed `/v1/completions` calls, and every call your own code in the proxy's process makes to LiteLLM directly.
|
|
1186
1188
|
|
|
@@ -1223,6 +1225,26 @@ calls: a failed call is recorded only for a request the Revenium guardrail's
|
|
|
1223
1225
|
sends no record. Adding `pre_call` enforces no budget unless
|
|
1224
1226
|
`REVENIUM_CIRCUIT_BREAKER_ENABLED=true`.
|
|
1225
1227
|
|
|
1228
|
+
**A failed call** sends one record with stop reason `ERROR`, zero tokens (or the
|
|
1229
|
+
tokens the error reports), and the error's class as its error reason, followed by
|
|
1230
|
+
its HTTP status when the call got an HTTP error answer: `RateLimitError (HTTP 429)`
|
|
1231
|
+
for a provider's rate limit, `HTTPException (HTTP 400)` for a request another
|
|
1232
|
+
guardrail refused. That status is also the record's error code. A timeout, a
|
|
1233
|
+
refused connection and a failure with no HTTP status (`RuntimeError`, say) carry
|
|
1234
|
+
no error code. The message is added only as for the other integrations: with
|
|
1235
|
+
`REVENIUM_CAPTURE_ERROR_MESSAGES=true`, redacted, at most 255 characters, and for
|
|
1236
|
+
a provider's error or a guardrail's refusal, which can quote the request, only with
|
|
1237
|
+
`REVENIUM_CAPTURE_PROMPTS=true` as well. The duration runs from when LiteLLM
|
|
1238
|
+
started on the request (LiteLLM 1.102.0 and later; on older releases, from when
|
|
1239
|
+
it first handed the call to a provider) to the failure, and is 0 when LiteLLM
|
|
1240
|
+
recorded neither: before 1.102.0 for a call refused before it reached a provider,
|
|
1241
|
+
and on the moderation and audio routes. The time to first token is 0 and the
|
|
1242
|
+
completion start time is the moment of failure, as on the other integrations'
|
|
1243
|
+
failed records. A content block, whether a guardrail's or a provider's
|
|
1244
|
+
content-policy refusal, is recorded as `ERROR` with its class as the reason, not
|
|
1245
|
+
as `CONTENT_FILTER`. The deprecated callback's failure record carries the same
|
|
1246
|
+
error reason and error code.
|
|
1247
|
+
|
|
1226
1248
|
**Budget enforcement** reuses the SDK's own circuit breaker, so a proxy enforces
|
|
1227
1249
|
exactly what every other Revenium integration enforces — including department
|
|
1228
1250
|
budgets. An organization cap is checked against the organization the call names,
|
|
@@ -1495,6 +1517,44 @@ return of double counting. Your other routes never carry an identifier and are n
|
|
|
1495
1517
|
counted or warned about, so a proxy that also serves chat completions or embeddings
|
|
1496
1518
|
stays quiet.
|
|
1497
1519
|
|
|
1520
|
+
##### Restarts and shutdown
|
|
1521
|
+
|
|
1522
|
+
A proxy stopped by SIGTERM (a rolling restart, a scale-down) loses the usage
|
|
1523
|
+
records the SDK still holds for its last calls unless you serve it as shown here.
|
|
1524
|
+
uvicorn, which the `litellm` command starts, ends the process by re-raising SIGTERM
|
|
1525
|
+
once its own graceful shutdown is over, before the SDK's exit flush can run (see
|
|
1526
|
+
[Shutdown](#shutdown)), and LiteLLM calls no shutdown method on a guardrail or
|
|
1527
|
+
callback (checked on LiteLLM 1.94.0, 1.105.7 and 1.109.1), so the `guardrails`
|
|
1528
|
+
block cannot add one.
|
|
1529
|
+
|
|
1530
|
+
Serve the proxy with uvicorn, through the SDK's wrapper of LiteLLM's app, instead of
|
|
1531
|
+
the `litellm` command:
|
|
1532
|
+
|
|
1533
|
+
```bash
|
|
1534
|
+
CONFIG_FILE_PATH=config.yaml uvicorn revenium_middleware.litellm.proxy.asgi:app --host 0.0.0.0 --port 4000
|
|
1535
|
+
```
|
|
1536
|
+
|
|
1537
|
+
It is LiteLLM's own app with one addition: once LiteLLM has finished its shutdown,
|
|
1538
|
+
spend-log flush included, the SDK sends the records it still holds, queued or held
|
|
1539
|
+
back while Revenium was not answering, before uvicorn ends the process. Requests
|
|
1540
|
+
pass through untouched. The flush takes at most `REVENIUM_SHUTDOWN_TIMEOUT_SECONDS`
|
|
1541
|
+
(5 seconds by default), so a Revenium that is slow or down delays the stop by that
|
|
1542
|
+
much at most; records still unsent then are dropped and counted in one warning, as
|
|
1543
|
+
at any other exit. An `on_metering_error` callback still running when that time is up
|
|
1544
|
+
does not hold the stop either: the proxy stops without waiting for it and logs a
|
|
1545
|
+
warning that says so. Leave room for it in the time your platform gives the proxy to
|
|
1546
|
+
stop (Kubernetes `terminationGracePeriodSeconds`, 30 seconds by default), on top of
|
|
1547
|
+
LiteLLM's own wait for in-flight requests. The flush ends metering in that process,
|
|
1548
|
+
and each uvicorn worker process flushes its own records.
|
|
1549
|
+
|
|
1550
|
+
What the `litellm` command did before it started uvicorn is then yours to do. Its
|
|
1551
|
+
options become uvicorn options or environment variables (`--num_workers` is
|
|
1552
|
+
uvicorn's `--workers`), and a proxy with a database needs its schema set up first:
|
|
1553
|
+
`litellm --config config.yaml --skip_server_startup` does that and exits, on
|
|
1554
|
+
LiteLLM releases that have that option. An ASGI app you build around LiteLLM's
|
|
1555
|
+
yourself gets the same flush from `DrainMeteringOnShutdown(app)` in
|
|
1556
|
+
`revenium_middleware.litellm.proxy.shutdown`.
|
|
1557
|
+
|
|
1498
1558
|
##### Migrating from the callback
|
|
1499
1559
|
|
|
1500
1560
|
`revenium_middleware.litellm.proxy.middleware.MiddlewareHandler` — the
|
|
@@ -1923,7 +1983,8 @@ Add business context to a call the SDK records by passing a `usage_metadata` dic
|
|
|
1923
1983
|
These fields are shortened with a WARNING when longer than Revenium stores, so the call is still
|
|
1924
1984
|
recorded: `trace_id`, `task_type`, `agent`, `subscription_id`, `parent_transaction_id`, `environment`,
|
|
1925
1985
|
`region`, `transaction_name`, `operation_subtype`, `subscriber.id` and `subscriber.email` at 255
|
|
1926
|
-
characters, and `product_name` at 224. Other text fields are sent as given.
|
|
1986
|
+
characters, and `product_name` at 224. Other text fields are sent as given. A spending cap filtered
|
|
1987
|
+
on agent, task type or product judges the call by the shortened value, the one its record carries.
|
|
1927
1988
|
|
|
1928
1989
|
| Field | Description | Use Case |
|
|
1929
1990
|
|-------|-------------|----------|
|
|
@@ -1941,7 +2002,7 @@ characters, and `product_name` at 224. Other text fields are sent as given.
|
|
|
1941
2002
|
| `effort` | Reasoning effort level requested of the model (free-form string, max 16 chars, `^[A-Za-z0-9_-]+$`) | Report what share of AI spend is high-effort reasoning, and compare cost per effort level |
|
|
1942
2003
|
| `entrypoint` | How the coding assistant that made the call was launched, such as `cli`, `sdk-py` or `claude-vscode` (completions only, 1 to 64 chars, `^[A-Za-z0-9._-]+$`) | Split coding-assistant spend between the terminal, the SDKs and the editor or chat integrations |
|
|
1943
2004
|
| `ccr_session_id` | The Claude Code cloud session the call ran in; `ccrSessionId` also accepted (completions only, 1 to 128 chars, `^[A-Za-z0-9_-]+$`) | Group the spend of one Claude Code cloud session |
|
|
1944
|
-
| `inference_geo` | Where the provider reports the inference ran, such as `us`; `inferenceGeo` also accepted (completions only, 1 to 16 chars, `^[A-Za-z0-9_-]+$`) | Passed to Revenium with the call
|
|
2005
|
+
| `inference_geo` | Where the provider reports the inference ran, such as `us`; `inferenceGeo` also accepted (completions only, 1 to 16 chars, `^[A-Za-z0-9_-]+$`) | Passed to Revenium with the call, except on a Claude call recorded under a provider other than `ANTHROPIC`, since only Anthropic's own API applies a US-inference rate. It is not stored as a dimension to group spend by |
|
|
1945
2006
|
|
|
1946
2007
|
**Example:**
|
|
1947
2008
|
|
|
@@ -1999,7 +2060,13 @@ where the provider says the inference ran. Both are values you supply per call;
|
|
|
1999
2060
|
the SDK does not read them from the environment or from the provider response.
|
|
2000
2061
|
`inference_geo` is sent exactly as given, with no lowercasing or allow-list;
|
|
2001
2062
|
the backend normalizes it. It is not stored as a dimension to group spend by,
|
|
2002
|
-
so it cannot break spend down by geography.
|
|
2063
|
+
so it cannot break spend down by geography. Only Anthropic's own API applies a
|
|
2064
|
+
US-inference rate, so on a Claude call `inference_geo` is sent only when the call
|
|
2065
|
+
is recorded with provider `ANTHROPIC`. On a Claude call recorded under any other provider, such as
|
|
2066
|
+
`Bedrock`, `Foundry`, `Google` or `LITELLM`, it is left off. Claude through the
|
|
2067
|
+
older `AnthropicVertex` client is recorded as `ANTHROPIC` and still sends it.
|
|
2068
|
+
Calls to other vendors' models, and a direct `client.ai.create_completion`
|
|
2069
|
+
call, send it as given. A value that is empty, too long (128 characters
|
|
2003
2070
|
for `ccr_session_id`, 16 for `inference_geo`) or holds anything but letters,
|
|
2004
2071
|
digits, hyphens and underscores is left off the record with a WARNING, and the
|
|
2005
2072
|
call is still recorded. Both are sent on completions only; image, video and
|
|
@@ -2730,7 +2797,9 @@ or unsent at the end are dropped and logged once, as a warning with a count.
|
|
|
2730
2797
|
|
|
2731
2798
|
A process ended by SIGTERM's default action skips `atexit`, and with it that
|
|
2732
2799
|
final flush. uvicorn, and so the LiteLLM proxy, ends that way once its own
|
|
2733
|
-
graceful shutdown has finished
|
|
2800
|
+
graceful shutdown has finished; serve the proxy as described in
|
|
2801
|
+
[Restarts and shutdown](#restarts-and-shutdown) to flush within that shutdown
|
|
2802
|
+
instead. For a script or worker with no SIGTERM handling
|
|
2734
2803
|
of its own, set `REVENIUM_INSTALL_SIGNAL_HANDLERS=1`: SIGTERM then flushes
|
|
2735
2804
|
first and ends the process the same way. A SIGTERM handler your code installed
|
|
2736
2805
|
before importing the SDK is called unchanged, and one installed afterwards
|
|
@@ -2782,7 +2851,7 @@ timeout raises `ValueError`.
|
|
|
2782
2851
|
| `REVENIUM_METERING_BASE_URL` | `https://api.revenium.ai` | Revenium API endpoint |
|
|
2783
2852
|
| `REVENIUM_LOG_LEVEL` | `INFO` | Log level: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL` |
|
|
2784
2853
|
| `REVENIUM_CAPTURE_PROMPTS` | `false` | Enable prompt capture |
|
|
2785
|
-
| `REVENIUM_CAPTURE_ERROR_MESSAGES` | `false` | Send the exception message, redacted and capped at 2048 characters, with failed `@meter_tool` events and auto-failed `JobContext` outcomes, and cut to 255 characters (keeping its "(HTTP <status>)" ending) on the record of a failed OpenAI, Anthropic, LiteLLM, Google, Perplexity or Ollama call; off sends the exception class only (with the HTTP status for a failed call). A provider HTTP error's message, which can quote the request, is sent only when `REVENIUM_CAPTURE_PROMPTS` is also on |
|
|
2854
|
+
| `REVENIUM_CAPTURE_ERROR_MESSAGES` | `false` | Send the exception message, redacted and capped at 2048 characters, with failed `@meter_tool` events and auto-failed `JobContext` outcomes, and cut to 255 characters (keeping its "(HTTP <status>)" ending) on the record of a failed OpenAI, Anthropic, LiteLLM, Google, Perplexity or Ollama call; off sends the exception class only (with the HTTP status for a failed call). A provider HTTP error's message, or on a failed LiteLLM proxy call's record the `HTTPException` a guardrail refused it with, which can quote the request, is sent only when `REVENIUM_CAPTURE_PROMPTS` is also on |
|
|
2786
2855
|
| `REVENIUM_SELECTIVE_METERING` | `false` | Only meter `@revenium_meter` decorated functions |
|
|
2787
2856
|
| `REVENIUM_TEAM_ID` | - | Team ID for cost lookups and outcome reporting (JobContext team resolution) |
|
|
2788
2857
|
| `REVENIUM_ENVIRONMENT` | - | Deployment environment (auto-detects from `ENVIRONMENT`, `DEPLOYMENT_ENV`) |
|
|
@@ -80,10 +80,10 @@ The calls below are billed by the provider (some, such as batch and fine-tuning
|
|
|
80
80
|
<!-- not-metered:start -->
|
|
81
81
|
- **OpenAI** (`OpenAI` and `AsyncOpenAI` clients): `batches.create`, `beta.agents.automations.create`, `beta.agents.automations.invocations.create`, `beta.agents.sessions.create`, `beta.agents.sessions.events.create`, `beta.agents.sessions.events.stream`, `beta.agents.sessions.stream`, `beta.agents.sessions.update`, `beta.chatkit.sessions.create`, `beta.realtime.connect`, `beta.responses.compact`, `beta.responses.connect`, `beta.responses.create`, `beta.threads.create_and_run`, `beta.threads.create_and_run_poll`, `beta.threads.create_and_run_stream`, `beta.threads.runs.create`, `beta.threads.runs.create_and_poll`, `beta.threads.runs.create_and_stream`, `beta.threads.runs.stream`, `beta.threads.runs.submit_tool_outputs`, `beta.threads.runs.submit_tool_outputs_and_poll`, `beta.threads.runs.submit_tool_outputs_stream`, `completions.create`, `decisions.create`, `evals.runs.create`, `fine_tuning.alpha.graders.run`, `fine_tuning.jobs.create`, `images.create_variation`, `images.edit`, `images.generate`, `live.connect`, `live.create`, `live.forks.connect`, `live.sessions.accept`, `live.sessions.fork`, `live.sideband.connect`, `realtime.calls.accept`, `realtime.calls.create`, `realtime.connect`, `realtime.translations.connect`, `responses.compact`, `responses.connect`, `videos.create`, `videos.create_and_poll`, `videos.create_character`, `videos.edit`, `videos.extend`, `videos.remix`
|
|
82
82
|
- **Anthropic** (`Anthropic` and `AsyncAnthropic` clients): `beta.deployments.run`, `beta.dreams.create`, `beta.environments.work.poller`, `beta.environments.work.worker`, `beta.messages.batches.create`, `beta.sessions.create`, `beta.sessions.events.send`, `beta.sessions.events.stream`, `beta.sessions.events.tool_runner`, `beta.sessions.threads.events.stream`, `messages.batches.create`
|
|
83
|
-
- **Google Gemini** (`google.genai` `Client` and `Client.aio`): `batches.create`, `batches.create_embeddings`, `caches.create`, `caches.update`, `decisions.create`, `file_search_stores.import_file`, `file_search_stores.upload_to_file_search_store`, `interactions.create`, `live.connect`, `live.music.connect`, `models.edit_image`, `models.recontext_image`, `models.segment_image`, `models.upscale_image`, `triggers.run`, `tunings.tune`
|
|
83
|
+
- **Google Gemini** (`google.genai` `Client` and `Client.aio`): `batches.create`, `batches.create_embeddings`, `caches.create`, `caches.update`, `decisions.create`, `file_search_stores.import_file`, `file_search_stores.upload_to_file_search_store`, `interactions.connect`, `interactions.create`, `live.connect`, `live.music.connect`, `models.edit_image`, `models.recontext_image`, `models.segment_image`, `models.upscale_image`, `triggers.run`, `tunings.tune`
|
|
84
84
|
- **Google Vertex AI** (`vertexai` classes): `batch_prediction.BatchPredictionJob.submit`, `caching.CachedContent.create`, `caching.CachedContent.update`, `language_models.ChatModel.tune_model`, `language_models.ChatModel.tune_model_rlhf`, `language_models.ChatSession.send_message`, `language_models.ChatSession.send_message_async`, `language_models.ChatSession.send_message_streaming`, `language_models.ChatSession.send_message_streaming_async`, `language_models.CodeChatModel.tune_model`, `language_models.CodeChatSession.send_message`, `language_models.CodeChatSession.send_message_async`, `language_models.CodeChatSession.send_message_streaming`, `language_models.CodeChatSession.send_message_streaming_async`, `language_models.CodeGenerationModel.batch_predict`, `language_models.CodeGenerationModel.predict`, `language_models.CodeGenerationModel.predict_async`, `language_models.CodeGenerationModel.predict_streaming`, `language_models.CodeGenerationModel.predict_streaming_async`, `language_models.CodeGenerationModel.tune_model`, `language_models.TextEmbeddingModel.batch_predict`, `language_models.TextEmbeddingModel.deploy_tuned_model`, `language_models.TextEmbeddingModel.tune_model`, `language_models.TextGenerationModel.batch_predict`, `language_models.TextGenerationModel.predict`, `language_models.TextGenerationModel.predict_async`, `language_models.TextGenerationModel.predict_streaming`, `language_models.TextGenerationModel.predict_streaming_async`, `language_models.TextGenerationModel.tune_model`, `language_models.TextGenerationModel.tune_model_rlhf`, `preview.caching.CachedContent.create`, `preview.caching.CachedContent.update`, `preview.vision_models.ImageCaptioningModel.get_captions`, `preview.vision_models.ImageGenerationModel.upscale_image`, `preview.vision_models.ImageQnAModel.ask_question`, `preview.vision_models.ImageSegmentationModel.segment_image`, `preview.vision_models.ImageTextModel.ask_question`, `preview.vision_models.ImageTextModel.get_captions`, `preview.vision_models.MultiModalEmbeddingModel.get_embeddings`, `preview.vision_models.WatermarkVerificationModel.verify_image`, `tuning.sft.train`, `vision_models.ImageCaptioningModel.get_captions`, `vision_models.ImageGenerationModel.upscale_image`, `vision_models.ImageQnAModel.ask_question`, `vision_models.ImageTextModel.ask_question`, `vision_models.ImageTextModel.get_captions`, `vision_models.MultiModalEmbeddingModel.get_embeddings`
|
|
85
85
|
- **AWS Bedrock** (boto3 bedrock-runtime client; Claude models are metered when REVENIUM_BEDROCK_TRANSPORT is on): `apply_guardrail`, `converse`, `converse_stream`, `invoke_guardrail_checks`, `invoke_model`, `invoke_model_with_response_stream`, `start_async_invoke`
|
|
86
|
-
- **Perplexity** (`Perplexity` and `AsyncPerplexity` clients): `async_.chat.completions.create`, `batches.create`, `browser.sessions.create`, `contextualized_embeddings.create`, `decisions.create`, `embeddings.create`, `responses.create`, `search.create`
|
|
86
|
+
- **Perplexity** (`Perplexity` and `AsyncPerplexity` clients): `async_.chat.completions.create`, `batches.create`, `browser.sessions.create`, `context.create`, `contextualized_embeddings.create`, `decisions.create`, `embeddings.create`, `monitors.create`, `monitors.run`, `responses.create`, `search.create`
|
|
87
87
|
- **LiteLLM** (the `litellm` module): `aadapter_completion`, `aadapter_generate_content`, `acode_interpreter_tool`, `acompact_responses`, `acreate_batch`, `acreate_fine_tuning_job`, `acreate_run`, `acreate_sandbox`, `adapter_completion`, `adecisions`, `aextraction`, `agenerate_content`, `aimage_edit`, `aimage_generation`, `aimage_variation`, `aingest`, `anthropic_interface.messages.acreate`, `anthropic_interface.messages.create`, `anthropic_messages`, `anthropic_messages_handler`, `aocr`, `aquery`, `arealtime_calls`, `arerank`, `aresponses`, `aresponses_api_with_mcp`, `aresponses_with_retries`, `arun_code`, `arun_thread`, `arun_thread_stream`, `asearch`, `aspeech`, `asystemone`, `atranscription`, `avideo_create_character`, `avideo_edit`, `avideo_extension`, `avideo_generation`, `avideo_remix`, `compact_responses`, `config_completion`, `create_batch`, `create_fine_tuning_job`, `create_run`, `decisions`, `extraction`, `image_edit`, `image_generation`, `image_variation`, `ingest`, `nlp_cloud_chat_completion`, `ocr`, `query`, `replicate_chat_completion`, `rerank`, `responses`, `responses_with_retries`, `run_thread`, `run_thread_stream`, `search`, `speech`, `systemone`, `transcription`, `video_create_character`, `video_edit`, `video_extension`, `video_generation`, `video_remix`
|
|
88
88
|
- **Ollama** (`Client`, `AsyncClient` and the `ollama` module): `systemone`, `web_fetch`, `web_search`
|
|
89
89
|
- **fal.ai** (`SyncClient`, `AsyncClient` and the `fal_client` module): `realtime`, `realtime_async`, `ws_connect`, `ws_connect_async`
|
|
@@ -447,6 +447,8 @@ client.close()
|
|
|
447
447
|
|
|
448
448
|
In an `emit_completion` payload the agent is named by `agent`. `agentName` and `currency` are not sent (Revenium does not store them); `emit_completion` names them in its one-time warning about unsent keys.
|
|
449
449
|
|
|
450
|
+
`cacheCreationTokenCount` is the whole cache write. When the provider reports how that write splits by cache lifetime, send the split beside it as `cacheCreation5mTokenCount`, `cacheCreation1hTokenCount` and `cacheCreation12hTokenCount` so each part is priced at its own rate; leave a bucket out when the provider does not report it rather than sending 0. `get_client().ai.create_completion` takes the same buckets as `cache_creation5m_token_count`, `cache_creation1h_token_count` and `cache_creation12h_token_count`.
|
|
451
|
+
|
|
450
452
|
`emit_completion` and `emit_tool_event` queue the record and return at once, like every automatic record: the SDK's background threads send it with the same timeouts and retries (see [Metering Delivery and Timeouts](#metering-delivery-and-timeouts)), and a record that fails with a timeout, a refused connection or a 5xx goes to the store-and-forward buffer and is sent when Revenium is reachable again. Delivery errors are logged, not raised; `emit_completion` still raises `ValueError` before anything is queued when the payload is missing a required key. Both calls return the payload you passed.
|
|
451
453
|
|
|
452
454
|
Records go to the address set in code with `initialize_metering(base_url=...)` or a `metering_key(..., base_url=...)` block when one is set, else `REVENIUM_METERING_BASE_URL`; pass `meter_base_url=` to `AgenticOutcomeSettings` to pin another address. They are always sent under the settings' `api_key`, never a `metering_key()` key. `http_client=` serves jobs and outcomes only.
|
|
@@ -1064,7 +1066,7 @@ response = litellm.completion(
|
|
|
1064
1066
|
)
|
|
1065
1067
|
```
|
|
1066
1068
|
|
|
1067
|
-
Client mode meters `completion`, `acompletion`, `text_completion`, `atext_completion`, `completion_with_retries` and `acompletion_with_retries`, streamed or not, and `embedding` / `aembedding` as embedding calls. When you stream without `stream_options={"include_usage": True}`, the middleware asks LiteLLM for the usage chunk itself and keeps it out of your stream. LiteLLM adds that chunk for every provider, including ones it does not send `stream_options` to. `litellm.responses` and `litellm.aresponses` are not metered, except on providers LiteLLM serves through `completion`. A completion that fails, or a stream that breaks off, sends one record with stop reason `ERROR`, zero tokens for a call that never started, and the error's class and HTTP status as its error reason (plus the provider's redacted error message when both `REVENIUM_CAPTURE_ERROR_MESSAGES` and `REVENIUM_CAPTURE_PROMPTS` are `true`; a provider's error can quote the request, so it is never sent with prompt capture off); your code still gets the original exception. The HTTP status is also sent as the record's error code; a call that got no HTTP error answer (a connection error, a timeout, or a stream that broke off) has none. LiteLLM gives a connection error a 500 and a timeout a 408 that no server answered, so those get no error code and no HTTP status in their error reason, including the `InternalServerError` LiteLLM raises for a refused connection on its `openai/` and `anthropic/` providers. LiteLLM also raises a few real provider answers, such as a Vertex AI 502, as connection errors, and those get no error code either. Failure records sent by the LiteLLM gateway (proxy)
|
|
1069
|
+
Client mode meters `completion`, `acompletion`, `text_completion`, `atext_completion`, `completion_with_retries` and `acompletion_with_retries`, streamed or not, and `embedding` / `aembedding` as embedding calls. When you stream without `stream_options={"include_usage": True}`, the middleware asks LiteLLM for the usage chunk itself and keeps it out of your stream. LiteLLM adds that chunk for every provider, including ones it does not send `stream_options` to. `litellm.responses` and `litellm.aresponses` are not metered, except on providers LiteLLM serves through `completion`. A completion that fails, or a stream that breaks off, sends one record with stop reason `ERROR`, zero tokens for a call that never started, and the error's class and HTTP status as its error reason (plus the provider's redacted error message when both `REVENIUM_CAPTURE_ERROR_MESSAGES` and `REVENIUM_CAPTURE_PROMPTS` are `true`; a provider's error can quote the request, so it is never sent with prompt capture off); your code still gets the original exception. The HTTP status is also sent as the record's error code; a call that got no HTTP error answer (a connection error, a timeout, or a stream that broke off) has none. LiteLLM gives a connection error a 500 and a timeout a 408 that no server answered, so those get no error code and no HTTP status in their error reason, including the `InternalServerError` LiteLLM raises for a refused connection on its `openai/` and `anthropic/` providers. LiteLLM also raises a few real provider answers, such as a Vertex AI 502, as connection errors, and those get no error code either. Failure records sent by the LiteLLM gateway (proxy) are described under Proxy Mode below.
|
|
1068
1070
|
|
|
1069
1071
|
Inside a LiteLLM proxy running `ReveniumGuardrail` with `default_on: true` and `post_call` among its modes, client mode leaves each proxied request to the guardrail, so the request is recorded once, with the guardrail's subscriber. The row names the deployment's model (the `litellm_params.model` behind the `model_name` your callers use, without its provider prefix), so a call to an alias is priced as the model it reached. The guardrail also meters streamed `/v1/responses` calls (Codex CLI, the OpenAI Agents SDK) for every provider. Client mode still records the proxied calls the guardrail does not meter, which today are streamed `/v1/completions` calls, and every call your own code in the proxy's process makes to LiteLLM directly.
|
|
1070
1072
|
|
|
@@ -1107,6 +1109,26 @@ calls: a failed call is recorded only for a request the Revenium guardrail's
|
|
|
1107
1109
|
sends no record. Adding `pre_call` enforces no budget unless
|
|
1108
1110
|
`REVENIUM_CIRCUIT_BREAKER_ENABLED=true`.
|
|
1109
1111
|
|
|
1112
|
+
**A failed call** sends one record with stop reason `ERROR`, zero tokens (or the
|
|
1113
|
+
tokens the error reports), and the error's class as its error reason, followed by
|
|
1114
|
+
its HTTP status when the call got an HTTP error answer: `RateLimitError (HTTP 429)`
|
|
1115
|
+
for a provider's rate limit, `HTTPException (HTTP 400)` for a request another
|
|
1116
|
+
guardrail refused. That status is also the record's error code. A timeout, a
|
|
1117
|
+
refused connection and a failure with no HTTP status (`RuntimeError`, say) carry
|
|
1118
|
+
no error code. The message is added only as for the other integrations: with
|
|
1119
|
+
`REVENIUM_CAPTURE_ERROR_MESSAGES=true`, redacted, at most 255 characters, and for
|
|
1120
|
+
a provider's error or a guardrail's refusal, which can quote the request, only with
|
|
1121
|
+
`REVENIUM_CAPTURE_PROMPTS=true` as well. The duration runs from when LiteLLM
|
|
1122
|
+
started on the request (LiteLLM 1.102.0 and later; on older releases, from when
|
|
1123
|
+
it first handed the call to a provider) to the failure, and is 0 when LiteLLM
|
|
1124
|
+
recorded neither: before 1.102.0 for a call refused before it reached a provider,
|
|
1125
|
+
and on the moderation and audio routes. The time to first token is 0 and the
|
|
1126
|
+
completion start time is the moment of failure, as on the other integrations'
|
|
1127
|
+
failed records. A content block, whether a guardrail's or a provider's
|
|
1128
|
+
content-policy refusal, is recorded as `ERROR` with its class as the reason, not
|
|
1129
|
+
as `CONTENT_FILTER`. The deprecated callback's failure record carries the same
|
|
1130
|
+
error reason and error code.
|
|
1131
|
+
|
|
1110
1132
|
**Budget enforcement** reuses the SDK's own circuit breaker, so a proxy enforces
|
|
1111
1133
|
exactly what every other Revenium integration enforces — including department
|
|
1112
1134
|
budgets. An organization cap is checked against the organization the call names,
|
|
@@ -1379,6 +1401,44 @@ return of double counting. Your other routes never carry an identifier and are n
|
|
|
1379
1401
|
counted or warned about, so a proxy that also serves chat completions or embeddings
|
|
1380
1402
|
stays quiet.
|
|
1381
1403
|
|
|
1404
|
+
##### Restarts and shutdown
|
|
1405
|
+
|
|
1406
|
+
A proxy stopped by SIGTERM (a rolling restart, a scale-down) loses the usage
|
|
1407
|
+
records the SDK still holds for its last calls unless you serve it as shown here.
|
|
1408
|
+
uvicorn, which the `litellm` command starts, ends the process by re-raising SIGTERM
|
|
1409
|
+
once its own graceful shutdown is over, before the SDK's exit flush can run (see
|
|
1410
|
+
[Shutdown](#shutdown)), and LiteLLM calls no shutdown method on a guardrail or
|
|
1411
|
+
callback (checked on LiteLLM 1.94.0, 1.105.7 and 1.109.1), so the `guardrails`
|
|
1412
|
+
block cannot add one.
|
|
1413
|
+
|
|
1414
|
+
Serve the proxy with uvicorn, through the SDK's wrapper of LiteLLM's app, instead of
|
|
1415
|
+
the `litellm` command:
|
|
1416
|
+
|
|
1417
|
+
```bash
|
|
1418
|
+
CONFIG_FILE_PATH=config.yaml uvicorn revenium_middleware.litellm.proxy.asgi:app --host 0.0.0.0 --port 4000
|
|
1419
|
+
```
|
|
1420
|
+
|
|
1421
|
+
It is LiteLLM's own app with one addition: once LiteLLM has finished its shutdown,
|
|
1422
|
+
spend-log flush included, the SDK sends the records it still holds, queued or held
|
|
1423
|
+
back while Revenium was not answering, before uvicorn ends the process. Requests
|
|
1424
|
+
pass through untouched. The flush takes at most `REVENIUM_SHUTDOWN_TIMEOUT_SECONDS`
|
|
1425
|
+
(5 seconds by default), so a Revenium that is slow or down delays the stop by that
|
|
1426
|
+
much at most; records still unsent then are dropped and counted in one warning, as
|
|
1427
|
+
at any other exit. An `on_metering_error` callback still running when that time is up
|
|
1428
|
+
does not hold the stop either: the proxy stops without waiting for it and logs a
|
|
1429
|
+
warning that says so. Leave room for it in the time your platform gives the proxy to
|
|
1430
|
+
stop (Kubernetes `terminationGracePeriodSeconds`, 30 seconds by default), on top of
|
|
1431
|
+
LiteLLM's own wait for in-flight requests. The flush ends metering in that process,
|
|
1432
|
+
and each uvicorn worker process flushes its own records.
|
|
1433
|
+
|
|
1434
|
+
What the `litellm` command did before it started uvicorn is then yours to do. Its
|
|
1435
|
+
options become uvicorn options or environment variables (`--num_workers` is
|
|
1436
|
+
uvicorn's `--workers`), and a proxy with a database needs its schema set up first:
|
|
1437
|
+
`litellm --config config.yaml --skip_server_startup` does that and exits, on
|
|
1438
|
+
LiteLLM releases that have that option. An ASGI app you build around LiteLLM's
|
|
1439
|
+
yourself gets the same flush from `DrainMeteringOnShutdown(app)` in
|
|
1440
|
+
`revenium_middleware.litellm.proxy.shutdown`.
|
|
1441
|
+
|
|
1382
1442
|
##### Migrating from the callback
|
|
1383
1443
|
|
|
1384
1444
|
`revenium_middleware.litellm.proxy.middleware.MiddlewareHandler` — the
|
|
@@ -1807,7 +1867,8 @@ Add business context to a call the SDK records by passing a `usage_metadata` dic
|
|
|
1807
1867
|
These fields are shortened with a WARNING when longer than Revenium stores, so the call is still
|
|
1808
1868
|
recorded: `trace_id`, `task_type`, `agent`, `subscription_id`, `parent_transaction_id`, `environment`,
|
|
1809
1869
|
`region`, `transaction_name`, `operation_subtype`, `subscriber.id` and `subscriber.email` at 255
|
|
1810
|
-
characters, and `product_name` at 224. Other text fields are sent as given.
|
|
1870
|
+
characters, and `product_name` at 224. Other text fields are sent as given. A spending cap filtered
|
|
1871
|
+
on agent, task type or product judges the call by the shortened value, the one its record carries.
|
|
1811
1872
|
|
|
1812
1873
|
| Field | Description | Use Case |
|
|
1813
1874
|
|-------|-------------|----------|
|
|
@@ -1825,7 +1886,7 @@ characters, and `product_name` at 224. Other text fields are sent as given.
|
|
|
1825
1886
|
| `effort` | Reasoning effort level requested of the model (free-form string, max 16 chars, `^[A-Za-z0-9_-]+$`) | Report what share of AI spend is high-effort reasoning, and compare cost per effort level |
|
|
1826
1887
|
| `entrypoint` | How the coding assistant that made the call was launched, such as `cli`, `sdk-py` or `claude-vscode` (completions only, 1 to 64 chars, `^[A-Za-z0-9._-]+$`) | Split coding-assistant spend between the terminal, the SDKs and the editor or chat integrations |
|
|
1827
1888
|
| `ccr_session_id` | The Claude Code cloud session the call ran in; `ccrSessionId` also accepted (completions only, 1 to 128 chars, `^[A-Za-z0-9_-]+$`) | Group the spend of one Claude Code cloud session |
|
|
1828
|
-
| `inference_geo` | Where the provider reports the inference ran, such as `us`; `inferenceGeo` also accepted (completions only, 1 to 16 chars, `^[A-Za-z0-9_-]+$`) | Passed to Revenium with the call
|
|
1889
|
+
| `inference_geo` | Where the provider reports the inference ran, such as `us`; `inferenceGeo` also accepted (completions only, 1 to 16 chars, `^[A-Za-z0-9_-]+$`) | Passed to Revenium with the call, except on a Claude call recorded under a provider other than `ANTHROPIC`, since only Anthropic's own API applies a US-inference rate. It is not stored as a dimension to group spend by |
|
|
1829
1890
|
|
|
1830
1891
|
**Example:**
|
|
1831
1892
|
|
|
@@ -1883,7 +1944,13 @@ where the provider says the inference ran. Both are values you supply per call;
|
|
|
1883
1944
|
the SDK does not read them from the environment or from the provider response.
|
|
1884
1945
|
`inference_geo` is sent exactly as given, with no lowercasing or allow-list;
|
|
1885
1946
|
the backend normalizes it. It is not stored as a dimension to group spend by,
|
|
1886
|
-
so it cannot break spend down by geography.
|
|
1947
|
+
so it cannot break spend down by geography. Only Anthropic's own API applies a
|
|
1948
|
+
US-inference rate, so on a Claude call `inference_geo` is sent only when the call
|
|
1949
|
+
is recorded with provider `ANTHROPIC`. On a Claude call recorded under any other provider, such as
|
|
1950
|
+
`Bedrock`, `Foundry`, `Google` or `LITELLM`, it is left off. Claude through the
|
|
1951
|
+
older `AnthropicVertex` client is recorded as `ANTHROPIC` and still sends it.
|
|
1952
|
+
Calls to other vendors' models, and a direct `client.ai.create_completion`
|
|
1953
|
+
call, send it as given. A value that is empty, too long (128 characters
|
|
1887
1954
|
for `ccr_session_id`, 16 for `inference_geo`) or holds anything but letters,
|
|
1888
1955
|
digits, hyphens and underscores is left off the record with a WARNING, and the
|
|
1889
1956
|
call is still recorded. Both are sent on completions only; image, video and
|
|
@@ -2614,7 +2681,9 @@ or unsent at the end are dropped and logged once, as a warning with a count.
|
|
|
2614
2681
|
|
|
2615
2682
|
A process ended by SIGTERM's default action skips `atexit`, and with it that
|
|
2616
2683
|
final flush. uvicorn, and so the LiteLLM proxy, ends that way once its own
|
|
2617
|
-
graceful shutdown has finished
|
|
2684
|
+
graceful shutdown has finished; serve the proxy as described in
|
|
2685
|
+
[Restarts and shutdown](#restarts-and-shutdown) to flush within that shutdown
|
|
2686
|
+
instead. For a script or worker with no SIGTERM handling
|
|
2618
2687
|
of its own, set `REVENIUM_INSTALL_SIGNAL_HANDLERS=1`: SIGTERM then flushes
|
|
2619
2688
|
first and ends the process the same way. A SIGTERM handler your code installed
|
|
2620
2689
|
before importing the SDK is called unchanged, and one installed afterwards
|
|
@@ -2666,7 +2735,7 @@ timeout raises `ValueError`.
|
|
|
2666
2735
|
| `REVENIUM_METERING_BASE_URL` | `https://api.revenium.ai` | Revenium API endpoint |
|
|
2667
2736
|
| `REVENIUM_LOG_LEVEL` | `INFO` | Log level: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL` |
|
|
2668
2737
|
| `REVENIUM_CAPTURE_PROMPTS` | `false` | Enable prompt capture |
|
|
2669
|
-
| `REVENIUM_CAPTURE_ERROR_MESSAGES` | `false` | Send the exception message, redacted and capped at 2048 characters, with failed `@meter_tool` events and auto-failed `JobContext` outcomes, and cut to 255 characters (keeping its "(HTTP <status>)" ending) on the record of a failed OpenAI, Anthropic, LiteLLM, Google, Perplexity or Ollama call; off sends the exception class only (with the HTTP status for a failed call). A provider HTTP error's message, which can quote the request, is sent only when `REVENIUM_CAPTURE_PROMPTS` is also on |
|
|
2738
|
+
| `REVENIUM_CAPTURE_ERROR_MESSAGES` | `false` | Send the exception message, redacted and capped at 2048 characters, with failed `@meter_tool` events and auto-failed `JobContext` outcomes, and cut to 255 characters (keeping its "(HTTP <status>)" ending) on the record of a failed OpenAI, Anthropic, LiteLLM, Google, Perplexity or Ollama call; off sends the exception class only (with the HTTP status for a failed call). A provider HTTP error's message, or on a failed LiteLLM proxy call's record the `HTTPException` a guardrail refused it with, which can quote the request, is sent only when `REVENIUM_CAPTURE_PROMPTS` is also on |
|
|
2670
2739
|
| `REVENIUM_SELECTIVE_METERING` | `false` | Only meter `@revenium_meter` decorated functions |
|
|
2671
2740
|
| `REVENIUM_TEAM_ID` | - | Team ID for cost lookups and outcome reporting (JobContext team resolution) |
|
|
2672
2741
|
| `REVENIUM_ENVIRONMENT` | - | Deployment environment (auto-detects from `ENVIRONMENT`, `DEPLOYMENT_ENV`) |
|
|
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
|
|
|
4
4
|
|
|
5
5
|
[project]
|
|
6
6
|
name = "revenium-python-sdk"
|
|
7
|
-
version = "0.
|
|
7
|
+
version = "0.16.1"
|
|
8
8
|
description = "The official Revenium Python SDK — unified AI metering middleware for OpenAI, Anthropic, Google, Ollama, LiteLLM, Perplexity, and fal.ai."
|
|
9
9
|
readme = "README.md"
|
|
10
10
|
requires-python = ">=3.8"
|
{revenium_python_sdk-0.15.0 → revenium_python_sdk-0.16.1}/revenium_middleware/_core/cache_tokens.py
RENAMED
|
@@ -102,7 +102,9 @@ def without_openai_cache_writes(prompt_tokens: int, usage: Any) -> int:
|
|
|
102
102
|
"""``prompt_tokens`` less the OpenAI cache write it contains, never below 0.
|
|
103
103
|
|
|
104
104
|
Cache reads stay in: the platform removes them itself for OpenAI-shaped
|
|
105
|
-
rows (see ``billable_input_tokens``).
|
|
105
|
+
rows (see ``billable_input_tokens``). The one exception is Bedrock's
|
|
106
|
+
OpenAI-compatible runtime routes, whose read ``billable_input_tokens``
|
|
107
|
+
removes (``_BEDROCK_OPENAI_SHAPED_ROUTES``, BACK-6110).
|
|
106
108
|
"""
|
|
107
109
|
written = openai_cache_write_tokens(usage)
|
|
108
110
|
if not prompt_tokens or not written:
|
|
@@ -141,8 +143,12 @@ def billable_input_tokens(
|
|
|
141
143
|
and through ``prompt_folds_claude_cache`` for the OpenAI client's records
|
|
142
144
|
of a gateway serving Claude. On Bedrock the cache is also removed for every
|
|
143
145
|
call whose usage LiteLLM built with a conversion that counts it inside the
|
|
144
|
-
prompt, whatever the model is called (``prompt_tokens_fold_cache``).
|
|
145
|
-
|
|
146
|
+
prompt, whatever the model is called (``prompt_tokens_fold_cache``). That
|
|
147
|
+
includes the one OpenAI-shaped case removed here: GPT and Grok on Bedrock's
|
|
148
|
+
OpenAI-compatible runtime routes, priced from a ``bedrock_converse`` row the
|
|
149
|
+
platform does not net (``_BEDROCK_OPENAI_SHAPED_ROUTES``, BACK-6110). If the
|
|
150
|
+
platform starts netting those rows, that case must go in the same release.
|
|
151
|
+
Every other model's prompt count is reported exactly as LiteLLM built it.
|
|
146
152
|
|
|
147
153
|
A count that comes out negative contradicts the cache counts reported
|
|
148
154
|
beside it, so it is reported unchanged and said out loud rather than
|
|
@@ -167,10 +173,11 @@ ANTHROPIC_MESSAGES_CALL_TYPE = "anthropic_messages"
|
|
|
167
173
|
# LiteLLM ``custom_llm_provider`` values whose usage conversion folds the cache
|
|
168
174
|
# buckets into ``prompt_tokens`` *and* whose rows the platform stores as sent.
|
|
169
175
|
# Anthropic-shaped usage is the one usage shape Revenium never normalizes
|
|
170
|
-
# server-side (see billable_input_tokens), so it is the
|
|
171
|
-
#
|
|
176
|
+
# server-side (see billable_input_tokens), so it is the overlap this row
|
|
177
|
+
# removes: here for direct Anthropic, and in ``_CLAUDE_HOSTING_PROVIDERS``
|
|
172
178
|
# for Claude on Bedrock and Vertex AI, plus the Bedrock conversions that build
|
|
173
|
-
# the same shape whatever the model is named.
|
|
179
|
+
# the same shape whatever the model is named. The only OpenAI-shaped overlap
|
|
180
|
+
# removed is the read on ``_BEDROCK_OPENAI_SHAPED_ROUTES``.
|
|
174
181
|
_CACHE_FOLDING_PROVIDERS = frozenset({"anthropic", "anthropic_text"})
|
|
175
182
|
|
|
176
183
|
# Providers that fold the same way when the model they serve is Claude, or, on
|
|
@@ -179,7 +186,12 @@ _CACHE_FOLDING_PROVIDERS = frozenset({"anthropic", "anthropic_text"})
|
|
|
179
186
|
# conversions all build ``prompt_tokens`` as uncached input plus cache read plus
|
|
180
187
|
# cache write. Gemini on Vertex AI is reported gross by Google and netted by the
|
|
181
188
|
# platform.
|
|
182
|
-
|
|
189
|
+
#
|
|
190
|
+
# ``bedrock_mantle`` is LiteLLM's second name for Bedrock's Mantle endpoint
|
|
191
|
+
# (``bedrock/mantle/<model>`` reports ``bedrock``). Only Claude is netted there:
|
|
192
|
+
# the platform nets the OpenAI-shaped Mantle rows itself, so netting them here
|
|
193
|
+
# as well would take the read out twice (BACK-6110, BACK-6099).
|
|
194
|
+
_CLAUDE_HOSTING_PROVIDERS = frozenset({"bedrock", "bedrock_mantle", "vertex_ai"})
|
|
183
195
|
|
|
184
196
|
# Only the top-level fields LiteLLM fills with the deployment it called.
|
|
185
197
|
_MODEL_FIELDS = ("model", "litellm_model_name")
|
|
@@ -284,8 +296,8 @@ def prompt_tokens_fold_cache(*sources):
|
|
|
284
296
|
arithmetic that sets their billed input count (see BACK-3190 for the same
|
|
285
297
|
rule on headers).
|
|
286
298
|
Measured against a live litellm 1.102.0 proxy: the logging event's
|
|
287
|
-
``kwargs`` and its ``litellm_params`` both carry it (also on
|
|
288
|
-
|
|
299
|
+
``kwargs`` and its ``litellm_params`` both carry it (also on 1.93.0),
|
|
300
|
+
while on the non-streamed ``/v1/messages`` route the post-call hook
|
|
289
301
|
is handed a raw provider body with no ``_hidden_params`` and a request dict
|
|
290
302
|
that names no provider -- there it is ``data["litellm_logging_obj"]``'s
|
|
291
303
|
``model_call_details`` that carries both the provider and the call type.
|
|
@@ -297,12 +309,14 @@ def prompt_tokens_fold_cache(*sources):
|
|
|
297
309
|
Anthropic (BACK-4162), anything else is not. On ``bedrock`` Claude is folded
|
|
298
310
|
by name, and any other model is folded when LiteLLM's own routing for it
|
|
299
311
|
builds usage with a conversion that counts the cache inside the prompt
|
|
300
|
-
(BACK-5780): an application inference profile ARN, Nova
|
|
312
|
+
(BACK-5780): an application inference profile ARN, Nova, and the
|
|
313
|
+
OpenAI-compatible runtime routes unless the id is an in-Region ``openai.*``
|
|
314
|
+
one (BACK-6110). On ``bedrock_mantle`` only Claude is folded. The model is the
|
|
301
315
|
first top-level ``model`` or ``litellm_model_name`` on any source, both
|
|
302
316
|
filled by LiteLLM with the deployment it called. The caller's alias, in the
|
|
303
317
|
request body, its metadata or a gateway's non-streamed ``response.model``,
|
|
304
318
|
is never handed in by a gateway hook, so it cannot turn a Gemini
|
|
305
|
-
or Bedrock
|
|
319
|
+
or Bedrock Mantle GPT deployment's count into a netted one.
|
|
306
320
|
|
|
307
321
|
With nothing to read, the prompt count stays as LiteLLM built it: reporting
|
|
308
322
|
a gross count costs the cache overlap twice for Anthropic only, while
|
|
@@ -329,6 +343,25 @@ _CONVERSE_READ_ONLY_FOLD_RELEASES = ((1, 66, 0), (1, 78, 0))
|
|
|
329
343
|
|
|
330
344
|
_BEDROCK_CONVERSE_ROUTES = frozenset({"converse", "converse_like"})
|
|
331
345
|
|
|
346
|
+
# LiteLLM's OpenAI-compatible routes to Bedrock's runtime endpoint: its default
|
|
347
|
+
# for GPT 5.6 and, from 1.110, Grok, and the ``chat_completions/`` and
|
|
348
|
+
# ``openai/`` prefixes. They report the cache read inside ``prompt_tokens`` the
|
|
349
|
+
# way OpenAI does, but the platform prices these models from the same catalog
|
|
350
|
+
# row as their Converse calls, which arrive net, so it cannot net one and not
|
|
351
|
+
# the other: the read comes out here (BACK-6110).
|
|
352
|
+
_BEDROCK_OPENAI_SHAPED_ROUTES = frozenset({"chat_completions", "openai"})
|
|
353
|
+
_OPENAI_SHAPED = "openai"
|
|
354
|
+
|
|
355
|
+
# An in-Region OpenAI id (``openai.gpt-5.6-luna``, no ``us.`` or ``global.``
|
|
356
|
+
# profile) has no catalog row of its own: the platform prices it from the
|
|
357
|
+
# ``bedrock_mantle/openai.*`` row and nets the read there (hypercurrent's
|
|
358
|
+
# ``CachePoolShape.isOpenAiOnMantle``), so it stays gross here.
|
|
359
|
+
_IN_REGION_OPENAI_ID_PREFIX = "openai."
|
|
360
|
+
|
|
361
|
+
|
|
362
|
+
def _platform_nets_the_read(model: str) -> bool:
|
|
363
|
+
return model.rsplit("/", 1)[-1].lower().startswith(_IN_REGION_OPENAI_ID_PREFIX)
|
|
364
|
+
|
|
332
365
|
|
|
333
366
|
def _installed_litellm_version() -> Optional[str]:
|
|
334
367
|
return installed_version("litellm")
|
|
@@ -360,7 +393,9 @@ def _bedrock_usage_family(model: str) -> Tuple[Any, Optional[BaseException]]:
|
|
|
360
393
|
|
|
361
394
|
Returns ``(family, error)``. ``family`` is ``"converse"`` or ``"anthropic"``,
|
|
362
395
|
both of which count the cache inside ``prompt_tokens`` and differ in when
|
|
363
|
-
LiteLLM started counting the write; ``
|
|
396
|
+
LiteLLM started counting the write; ``_OPENAI_SHAPED`` for a route that
|
|
397
|
+
counts the read inside it and whose write ``without_openai_cache_writes``
|
|
398
|
+
has already removed; ``None`` for any other route or
|
|
364
399
|
provider; ``_UNKNOWN`` when LiteLLM cannot route at all; or
|
|
365
400
|
``_INVOKE_PROVIDER_UNKNOWN`` for an invoke route whose provider it could not
|
|
366
401
|
name. ``error`` is the exception LiteLLM raised for the last two, else
|
|
@@ -375,6 +410,8 @@ def _bedrock_usage_family(model: str) -> Tuple[Any, Optional[BaseException]]:
|
|
|
375
410
|
return _UNKNOWN, error
|
|
376
411
|
if route in _BEDROCK_CONVERSE_ROUTES:
|
|
377
412
|
return "converse", None
|
|
413
|
+
if route in _BEDROCK_OPENAI_SHAPED_ROUTES:
|
|
414
|
+
return _OPENAI_SHAPED, None
|
|
378
415
|
if route != "invoke":
|
|
379
416
|
return None, None
|
|
380
417
|
try:
|
|
@@ -408,6 +445,8 @@ def _bedrock_usage_folds_cache(model: str) -> bool:
|
|
|
408
445
|
if family is _UNKNOWN or family is _INVOKE_PROVIDER_UNKNOWN:
|
|
409
446
|
_warn_bedrock_routing_unavailable_once(error)
|
|
410
447
|
return False
|
|
448
|
+
if family == _OPENAI_SHAPED:
|
|
449
|
+
return not _platform_nets_the_read(model)
|
|
411
450
|
return family is not None
|
|
412
451
|
|
|
413
452
|
|
|
@@ -429,17 +468,34 @@ def routes_via_bedrock_converse(*sources) -> bool:
|
|
|
429
468
|
|
|
430
469
|
|
|
431
470
|
def cache_creation_inside_litellm_prompt(cache_creation_tokens: int, *sources) -> int:
|
|
432
|
-
"""The part of a cache write
|
|
471
|
+
"""The part of a cache write still inside a LiteLLM prompt count that has
|
|
472
|
+
been through ``without_openai_cache_writes``.
|
|
433
473
|
|
|
434
474
|
What ``billable_input_tokens`` may take out of a folded LiteLLM prompt
|
|
435
475
|
count; the write itself is still reported in full beside it. ``sources``
|
|
436
476
|
are the ones handed to ``prompt_tokens_fold_cache`` for the same call.
|
|
477
|
+
|
|
478
|
+
Callers apply ``without_openai_cache_writes`` to ``prompt_tokens`` first.
|
|
479
|
+
On Bedrock's OpenAI-compatible routes (``_BEDROCK_OPENAI_SHAPED_ROUTES``)
|
|
480
|
+
that step has already removed the write, so the answer there is 0.
|
|
437
481
|
"""
|
|
482
|
+
if _write_left_the_prompt_as_an_openai_write(sources):
|
|
483
|
+
return 0
|
|
438
484
|
if litellm_folds_cache_write(converse=routes_via_bedrock_converse(*sources)):
|
|
439
485
|
return cache_creation_tokens
|
|
440
486
|
return 0
|
|
441
487
|
|
|
442
488
|
|
|
489
|
+
def _write_left_the_prompt_as_an_openai_write(sources) -> bool:
|
|
490
|
+
if _upstream_provider(sources) != "bedrock":
|
|
491
|
+
return False
|
|
492
|
+
model = _first_named_model(sources)
|
|
493
|
+
if not model or names_claude_model(model):
|
|
494
|
+
return False
|
|
495
|
+
family, _ = _bedrock_usage_family(str(model))
|
|
496
|
+
return family == _OPENAI_SHAPED
|
|
497
|
+
|
|
498
|
+
|
|
443
499
|
def total_priced_tokens(
|
|
444
500
|
input_tokens: int, output_tokens: int, cache_read_tokens: int, cache_creation_tokens: int
|
|
445
501
|
) -> int:
|
|
@@ -476,6 +532,7 @@ _CONVERSE_TTL_PARAMETERS = {
|
|
|
476
532
|
"5m": "cache_creation5m_token_count",
|
|
477
533
|
"1h": "cache_creation1h_token_count",
|
|
478
534
|
}
|
|
535
|
+
_TWELVE_HOUR_PARAMETER = "cache_creation12h_token_count"
|
|
479
536
|
|
|
480
537
|
|
|
481
538
|
def _reported_split(split: Any) -> Dict[str, int]:
|
|
@@ -509,10 +566,32 @@ def _converse_split(usage: Any) -> Dict[str, int]:
|
|
|
509
566
|
return counts
|
|
510
567
|
|
|
511
568
|
|
|
569
|
+
def _openai_responses_split(usage: Any) -> Dict[str, int]:
|
|
570
|
+
"""The 12-hour share of an OpenAI Responses write, kept only when it fits inside the write.
|
|
571
|
+
|
|
572
|
+
Only ``ttl_12h_tokens`` has a bucket: ``ttl_30m_tokens`` stays in the
|
|
573
|
+
write's unattributed remainder, which the platform prices at the
|
|
574
|
+
five-minute column, where it keeps OpenAI's 30-minute write price. So the
|
|
575
|
+
share need not add up to the write, but it must lie within the aggregate
|
|
576
|
+
``openai_cache_write_tokens`` reads from the same usage. A share with no
|
|
577
|
+
write to belong to, a negative one or one larger than the write is dropped
|
|
578
|
+
and the aggregate is sent alone, as ``_converse_split`` does: the platform
|
|
579
|
+
rejects such a bucket and bills the write flat, or at nothing for a model
|
|
580
|
+
priced at the input rate. Anthropic-shaped usage reports no OpenAI write
|
|
581
|
+
(see ``openai_cache_write_tokens``), so it never yields this bucket.
|
|
582
|
+
"""
|
|
583
|
+
details = _get(_get(usage, "input_tokens_details"), "cache_write_tokens_details")
|
|
584
|
+
tokens = _optional_token_count(details, "ttl_12h_tokens")
|
|
585
|
+
written = openai_cache_write_tokens(usage)
|
|
586
|
+
if not written or tokens is None or not 0 <= tokens <= written:
|
|
587
|
+
return {}
|
|
588
|
+
return {_TWELVE_HOUR_PARAMETER: tokens}
|
|
589
|
+
|
|
590
|
+
|
|
512
591
|
def extract_cache_creation_ttl_counts(usage: Any) -> Dict[str, int]:
|
|
513
592
|
"""Extract the per-TTL cache-creation breakdown as metering parameters.
|
|
514
593
|
|
|
515
|
-
|
|
594
|
+
Four usage shapes carry the split, and this reads whichever is present:
|
|
516
595
|
|
|
517
596
|
- Anthropic, natively and on Bedrock InvokeModel: ``usage.cache_creation``
|
|
518
597
|
(``ephemeral_5m_input_tokens`` / ``ephemeral_1h_input_tokens``).
|
|
@@ -520,6 +599,9 @@ def extract_cache_creation_ttl_counts(usage: Any) -> Dict[str, int]:
|
|
|
520
599
|
``usage.prompt_tokens_details.cache_creation_token_details``.
|
|
521
600
|
- Bedrock Converse: ``usage.cacheDetails``, one ``{ttl, inputTokens}``
|
|
522
601
|
entry per TTL.
|
|
602
|
+
- OpenAI Responses: ``usage.input_tokens_details.cache_write_tokens_details``
|
|
603
|
+
(``ttl_12h_tokens`` / ``ttl_30m_tokens``), of which only the 12-hour
|
|
604
|
+
share becomes a bucket.
|
|
523
605
|
|
|
524
606
|
The split is absent on older responses and whenever the extended cache-TTL
|
|
525
607
|
beta is not in play, so it is read defensively.
|
|
@@ -530,10 +612,16 @@ def extract_cache_creation_ttl_counts(usage: Any) -> Dict[str, int]:
|
|
|
530
612
|
zero: the aggregate ``cache_creation_token_count`` stays authoritative and
|
|
531
613
|
the backend prices the flat fallback from it when no split is present,
|
|
532
614
|
whereas an explicit zero would assert a split that never happened. A
|
|
533
|
-
|
|
615
|
+
split that does not fit the aggregate write the same usage reports is
|
|
616
|
+
dropped as a whole and the aggregate is sent alone: a Converse list that
|
|
617
|
+
does not add up to the write or names a TTL Revenium has no price for, and
|
|
618
|
+
an OpenAI 12-hour share that is negative, larger than the write, or
|
|
619
|
+
reported with no write (a reported 0 included). Any other bucket the
|
|
620
|
+
provider reports as 0 is forwarded as 0.
|
|
534
621
|
"""
|
|
535
622
|
return (
|
|
536
623
|
_reported_split(_get(usage, "cache_creation"))
|
|
537
624
|
or _reported_split(_get(_get(usage, "prompt_tokens_details"), "cache_creation_token_details"))
|
|
538
625
|
or _converse_split(usage)
|
|
626
|
+
or _openai_responses_split(usage)
|
|
539
627
|
)
|