revenium-python-sdk 0.14.0__tar.gz → 0.16.0__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-0.14.0 → revenium_python_sdk-0.16.0}/PKG-INFO +97 -17
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/README.md +95 -15
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/pyproject.toml +5 -2
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/__init__.py +4 -1
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/cache_tokens.py +137 -31
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/enforcement.py +631 -193
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/error_text.py +26 -9
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/failed_calls.py +9 -6
- revenium_python_sdk-0.16.0/revenium_middleware/_core/inference_geo.py +39 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/metering.py +4 -1
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/metering_submission.py +8 -1
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/tool_event_transport.py +11 -1
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/wire_guards.py +15 -5
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/resources/ai.py +30 -2
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/ai_create_completion_params.py +16 -1
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/job_type_economics.py +20 -0
- revenium_python_sdk-0.16.0/revenium_middleware/litellm/_failed_calls.py +40 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/middleware.py +46 -21
- revenium_python_sdk-0.16.0/revenium_middleware/litellm/proxy/_health_check.py +50 -0
- revenium_python_sdk-0.16.0/revenium_middleware/litellm/proxy/asgi.py +17 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/proxy/guardrail.py +60 -8
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/proxy/middleware.py +11 -3
- revenium_python_sdk-0.16.0/revenium_middleware/litellm/proxy/shutdown.py +104 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/middleware.py +141 -14
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_python_sdk.egg-info/PKG-INFO +97 -17
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_python_sdk.egg-info/SOURCES.txt +5 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_python_sdk.egg-info/requires.txt +1 -1
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/tests/test_metering.py +70 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/LICENSE +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/call_context.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/call_ownership.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/closable_scope.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/context.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/decorators.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/delivery_circuit.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/exceptions.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/json_safe.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/langchain_settings.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/load_diagnostics.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/log_sanitize.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/metering_buffer.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/metering_pool.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/metering_status.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/outcomes.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/package_versions.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/parse_failures.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/patch_registry.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/perplexity_cost.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/perplexity_usage_metadata.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/prompt_extraction.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/responses_status.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/secret_shapes.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/selective_metering.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/shutdown_signals.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/sse_body_tap.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/subscriber.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/usage_iterations.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_core/vendor_model_aliases.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/LICENSE +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_base_client.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_client.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_compat.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_constants.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_exceptions.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_files.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_models.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_qs.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_resource.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_response.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_streaming.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_types.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_logs.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_proxy.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_reflection.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_resources_proxy.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_streams.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_sync.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_transform.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_typing.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_utils/_utils.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/_version.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/context.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/decorator.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/py.typed +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/resources/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/resources/apis.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/resources/events.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/ai_create_audio_params.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/ai_create_image_params.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/ai_create_video_params.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/api_meter_request_params.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/api_meter_response_params.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/event_create_params.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/_metering/types/metering_response_resource.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/agentic_outcomes.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/bedrock_adapter.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/bedrock_model_names.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/bedrock_transport.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/inference_profiles.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/middleware.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/prompt_extractor.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/provider.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/raw_response.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/stream_create.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/anthropic/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/_call.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/_metering.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/_queue.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/middleware.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/fal/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/common/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/common/exceptions.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/common/protocols.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/common/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/common/types.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/common/utils.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/google_ai/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/google_ai/middleware.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/google_ai/provider.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/prompt_extractor.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/vertex_ai/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/vertex_ai/middleware.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/google/vertex_ai/provider.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/_metadata.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/_retry.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/anthropic_driver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/budget_errors.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/litellm_driver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/ollama_driver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/openai_driver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/openai_embedding_driver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/griptape/universal_driver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/job_context.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/job_history.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/context.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/decorators.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/hooks.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/integrations/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/integrations/crewai.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/client/validation.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/proxy/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/litellm/proxy/_metering_owner.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/ollama/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/ollama/middleware.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/ollama/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/audio.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/azure_config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/azure_model_resolver.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/config.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/exceptions.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/_abandoned_runs.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/_model_call.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/_perplexity_cost_capture.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/_prompt_capture.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/_text_completion_usage.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/_utils.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/langchain/unified_handler.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/prompt_extractor.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/provider.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/raw_response.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/openai/xai_usage.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/_metering_scope.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/middleware.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/patching.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/perplexity_sdk.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/provider.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/raw_response.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/streaming.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/perplexity/trace_fields.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/webhooks/__init__.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_middleware/webhooks/_verify.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_python_sdk.egg-info/dependency_links.txt +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/revenium_python_sdk.egg-info/top_level.txt +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/setup.cfg +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/tests/test_metering_dispatch_timestamps.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/tests/test_patch_coverage.py +0 -0
- {revenium_python_sdk-0.14.0 → revenium_python_sdk-0.16.0}/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: 0.16.0
|
|
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
|
|
@@ -80,7 +80,7 @@ Requires-Dist: litellm>=1.40.0; extra == "litellm"
|
|
|
80
80
|
Requires-Dist: python-dotenv>=0.19.0; extra == "litellm"
|
|
81
81
|
Provides-Extra: litellm-proxy
|
|
82
82
|
Requires-Dist: wrapt>=1.14.0; extra == "litellm-proxy"
|
|
83
|
-
Requires-Dist: litellm[proxy]>=1.
|
|
83
|
+
Requires-Dist: litellm[proxy]>=1.94.0; extra == "litellm-proxy"
|
|
84
84
|
Provides-Extra: dev
|
|
85
85
|
Requires-Dist: pytest>=7.0.0; extra == "dev"
|
|
86
86
|
Requires-Dist: pytest-asyncio; extra == "dev"
|
|
@@ -196,11 +196,11 @@ 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`, `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`
|
|
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`, `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`, `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`
|
|
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
|
+
- **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`
|
|
206
206
|
- **LangChain** (embeddings models from vendors other than OpenAI, Azure OpenAI and Ollama, such as `MistralAIEmbeddings`): `Embeddings.aembed_documents`, `Embeddings.aembed_query`, `Embeddings.embed_documents`, `Embeddings.embed_query`
|
|
@@ -382,6 +382,13 @@ explicit override. Economics metric directions are
|
|
|
382
382
|
Use `CUSTOMER_DECLARED` or `MEASURED` for a baseline override. Use `MEASURED`,
|
|
383
383
|
`SELF_REPORTED`, or `DERIVED` for a period fact override.
|
|
384
384
|
|
|
385
|
+
A baseline can also record the work's throughput and headcount before AI:
|
|
386
|
+
`throughput_value` with a `throughput_period` of `HOUR`, `DAY`, `WEEK`, or
|
|
387
|
+
`MONTH`, an optional `throughput_scope` of `TEAM` (the default) or
|
|
388
|
+
`PER_PERSON`, and `headcount`. Send `throughput_period` whenever
|
|
389
|
+
`throughput_value` is set; the server rejects a period or scope without a
|
|
390
|
+
value.
|
|
391
|
+
|
|
385
392
|
Facts are append-only and keyed on the period, dimension and metric key
|
|
386
393
|
together. Re-appending that tuple supersedes the active fact, and the server
|
|
387
394
|
requires `reason=` on the entry when it does.
|
|
@@ -556,6 +563,8 @@ client.close()
|
|
|
556
563
|
|
|
557
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.
|
|
558
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
|
+
|
|
559
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.
|
|
560
569
|
|
|
561
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.
|
|
@@ -1173,7 +1182,7 @@ response = litellm.completion(
|
|
|
1173
1182
|
)
|
|
1174
1183
|
```
|
|
1175
1184
|
|
|
1176
|
-
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.
|
|
1177
1186
|
|
|
1178
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.
|
|
1179
1188
|
|
|
@@ -1188,6 +1197,10 @@ except calls a spending cap refuses.
|
|
|
1188
1197
|
pip install "revenium-python-sdk[litellm-proxy]" # requires Python 3.10+
|
|
1189
1198
|
```
|
|
1190
1199
|
|
|
1200
|
+
The extra requires **LiteLLM 1.94.0 or newer**. On LiteLLM 1.93.x the proxy cannot
|
|
1201
|
+
load the guardrail from the config below: it only looks for a module file beside
|
|
1202
|
+
`config.yaml` and stops with `Could not find module file`.
|
|
1203
|
+
|
|
1191
1204
|
```yaml
|
|
1192
1205
|
guardrails:
|
|
1193
1206
|
- guardrail_name: "revenium"
|
|
@@ -1212,6 +1225,26 @@ calls: a failed call is recorded only for a request the Revenium guardrail's
|
|
|
1212
1225
|
sends no record. Adding `pre_call` enforces no budget unless
|
|
1213
1226
|
`REVENIUM_CIRCUIT_BREAKER_ENABLED=true`.
|
|
1214
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
|
+
|
|
1215
1248
|
**Budget enforcement** reuses the SDK's own circuit breaker, so a proxy enforces
|
|
1216
1249
|
exactly what every other Revenium integration enforces — including department
|
|
1217
1250
|
budgets. An organization cap is checked against the organization the call names,
|
|
@@ -1235,7 +1268,7 @@ the call proceeds, and it proceeds without waiting. The pre-call check reads the
|
|
|
1235
1268
|
rules the SDK already has cached and refreshes them on a background thread, so a
|
|
1236
1269
|
Revenium endpoint that hangs, refuses connections, returns errors or rejects your
|
|
1237
1270
|
key (403) neither blocks nor slows the proxy. A failed refresh is retried once
|
|
1238
|
-
per `REVENIUM_CB_POLL_INTERVAL_SECONDS
|
|
1271
|
+
per `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s), not once per request. Right after the
|
|
1239
1272
|
proxy starts, before the first rules fetch completes, no rules are cached and
|
|
1240
1273
|
calls go through; see [Fail-Open vs Fail-Closed](#fail-open-vs-fail-closed).
|
|
1241
1274
|
Metering is likewise non-disruptive — nothing in the post-call path can turn a
|
|
@@ -1433,7 +1466,7 @@ Blank or unset means on; `true`, `1`, `yes`, `on` mean on; any other value turns
|
|
|
1433
1466
|
New calls go straight back to two records when you do; records already merged stay
|
|
1434
1467
|
merged. Four things have to be true for it to work:
|
|
1435
1468
|
|
|
1436
|
-
* **LiteLLM 1.
|
|
1469
|
+
* **LiteLLM 1.94.0 or newer.** That is the floor of the `litellm-proxy` extra, which
|
|
1437
1470
|
is what installs the guardrail. On an older LiteLLM the guardrail logs one warning
|
|
1438
1471
|
at startup and behaves exactly as it does when opted out: no header is added,
|
|
1439
1472
|
the provider's response id is reported, and you get two records rather than none.
|
|
@@ -1484,6 +1517,44 @@ return of double counting. Your other routes never carry an identifier and are n
|
|
|
1484
1517
|
counted or warned about, so a proxy that also serves chat completions or embeddings
|
|
1485
1518
|
stays quiet.
|
|
1486
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
|
+
|
|
1487
1558
|
##### Migrating from the callback
|
|
1488
1559
|
|
|
1489
1560
|
`revenium_middleware.litellm.proxy.middleware.MiddlewareHandler` — the
|
|
@@ -1912,7 +1983,8 @@ Add business context to a call the SDK records by passing a `usage_metadata` dic
|
|
|
1912
1983
|
These fields are shortened with a WARNING when longer than Revenium stores, so the call is still
|
|
1913
1984
|
recorded: `trace_id`, `task_type`, `agent`, `subscription_id`, `parent_transaction_id`, `environment`,
|
|
1914
1985
|
`region`, `transaction_name`, `operation_subtype`, `subscriber.id` and `subscriber.email` at 255
|
|
1915
|
-
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.
|
|
1916
1988
|
|
|
1917
1989
|
| Field | Description | Use Case |
|
|
1918
1990
|
|-------|-------------|----------|
|
|
@@ -1930,7 +2002,7 @@ characters, and `product_name` at 224. Other text fields are sent as given.
|
|
|
1930
2002
|
| `effort` | Reasoning effort level requested of the model (free-form string, max 16 chars, `^[A-Za-z0-9_-]+$`) | Report what share of AI spend is high-effort reasoning, and compare cost per effort level |
|
|
1931
2003
|
| `entrypoint` | How the coding assistant that made the call was launched, such as `cli`, `sdk-py` or `claude-vscode` (completions only, 1 to 64 chars, `^[A-Za-z0-9._-]+$`) | Split coding-assistant spend between the terminal, the SDKs and the editor or chat integrations |
|
|
1932
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_-]+$`) | Group the spend of one Claude Code cloud session |
|
|
1933
|
-
| `inference_geo` | Where the provider reports the inference ran, such as `us`; `inferenceGeo` also accepted (completions only, 1 to 16 chars, `^[A-Za-z0-9_-]+$`) | 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 |
|
|
1934
2006
|
|
|
1935
2007
|
**Example:**
|
|
1936
2008
|
|
|
@@ -1988,7 +2060,13 @@ where the provider says the inference ran. Both are values you supply per call;
|
|
|
1988
2060
|
the SDK does not read them from the environment or from the provider response.
|
|
1989
2061
|
`inference_geo` is sent exactly as given, with no lowercasing or allow-list;
|
|
1990
2062
|
the backend normalizes it. It is not stored as a dimension to group spend by,
|
|
1991
|
-
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
|
|
1992
2070
|
for `ccr_session_id`, 16 for `inference_geo`) or holds anything but letters,
|
|
1993
2071
|
digits, hyphens and underscores is left off the record with a WARNING, and the
|
|
1994
2072
|
call is still recorded. Both are sent on completions only; image, video and
|
|
@@ -2390,7 +2468,7 @@ REVENIUM_ENFORCEMENT_BASE_URL=https://api.revenium.ai/profitstream # optional:
|
|
|
2390
2468
|
| `REVENIUM_BYPASS` | `false` | When `true`, every `check_enforcement` call short-circuits to a no-op. Useful for incident response. |
|
|
2391
2469
|
| `REVENIUM_TEAM_ID` | — | Hashed team ID. Path component on rule fetches; required when the breaker is enabled. |
|
|
2392
2470
|
| `REVENIUM_ENFORCEMENT_BASE_URL` | `https://api.revenium.ai/profitstream` | Base URL for the enforcement API, taken as given. When unset it is the address outcomes and jobs use (`REVENIUM_PROFITSTREAM_BASE_URL`, else the Revenium address `REVENIUM_METERING_BASE_URL` or `initialize_metering(base_url=...)` sits under, without its `/meter` segment) plus `/profitstream`. A rules fetch that answers 404 is retried once at the same address with `/profitstream` removed, or added if it was missing, and the address that answers with rules is used from then on, so a self-hosted server that serves the API at the root, or a value set without `/profitstream`, still reaches the rules. |
|
|
2393
|
-
| `REVENIUM_CB_POLL_INTERVAL_SECONDS` | `
|
|
2471
|
+
| `REVENIUM_CB_POLL_INTERVAL_SECONDS` | `30` | Seconds between background rule refreshes, and the wait before a failed refresh is retried. Whole seconds from a minimum of `5` to a maximum of `3600`: a lower value is raised to `5` with one warning per process start, and a higher one is capped at `3600`. Anything other than plain digits after trimming spaces (a sign, a decimal point, other text), an empty value and `0` use the default. Read once when polling starts, so a change takes effect when the process restarts. While calls are arriving, rules older than 120 seconds are refreshed on the next call whatever the interval, so a value above `120` only spaces out refreshes on an idle process and the retry after a failed refresh. |
|
|
2394
2472
|
| `REVENIUM_CB_FAIL_MODE` | `open` | `open` (default) lets calls through when no cache exists; `closed` raises `BudgetExceededError` until rules are loaded. |
|
|
2395
2473
|
| `REVENIUM_CACHE_DIR` | unset | When set, the rule cache is mirrored to `<dir>/revenium_enforcement_rules.<scope>.json` so a restarted process doesn't fail-closed on the very first call. `<scope>` is a short hash of the enforcement base URL and `REVENIUM_TEAM_ID`, so services for different teams can share one folder and each loads only its own rules. Nothing is saved without `REVENIUM_TEAM_ID`. |
|
|
2396
2474
|
|
|
@@ -2423,9 +2501,9 @@ LangChain embeddings models run no callbacks: `OpenAIEmbeddings` and `AzureOpenA
|
|
|
2423
2501
|
When the circuit breaker is disabled, it is a no-op. When enabled:
|
|
2424
2502
|
|
|
2425
2503
|
1. A daemon thread (`revenium-enforcement-poll`) starts on first use.
|
|
2426
|
-
2. It polls `GET {REVENIUM_ENFORCEMENT_BASE_URL}/v2/api/ai/enforcement-rules/{REVENIUM_TEAM_ID}` every `REVENIUM_CB_POLL_INTERVAL_SECONDS` with the `x-api-key` header.
|
|
2504
|
+
2. It polls `GET {REVENIUM_ENFORCEMENT_BASE_URL}/v2/api/ai/enforcement-rules/{REVENIUM_TEAM_ID}` every `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s) with the `x-api-key` header.
|
|
2427
2505
|
3. Rules are cached in-process. The pre-call check only reads that cache: it never makes a network request and never sleeps, so it is safe on an event loop (LiteLLM proxy, `AsyncOpenAI`). When the cache is more than 120 s old the check wakes the poller to refresh it early and carries on with the rules it has. Only one refresh runs at a time per process.
|
|
2428
|
-
4. A refresh that fails (timeout, refused connection, 5xx, or any 4xx such as a 403 for a wrong key or team id) keeps the cached rules, logs a warning, and is not retried for one `REVENIUM_CB_POLL_INTERVAL_SECONDS
|
|
2506
|
+
4. A refresh that fails (timeout, refused connection, 5xx, or any 4xx such as a 403 for a wrong key or team id) keeps the cached rules, logs a warning, and is not retried for one `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s), however much traffic arrives meanwhile. A 4xx answer the SDK does not retry (any but 408 and 429) is logged at WARNING with the URL and the status, so a wrong address (404) or a wrong key or team id (401, 403) is visible without debug logging.
|
|
2429
2507
|
5. `204 No Content` is treated as "no rules configured" — the cache is cleared.
|
|
2430
2508
|
|
|
2431
2509
|
A rule with filters (a cap scoped to an agent, model, provider, subscriber, credential, task type, product or organization) blocks or warns only the calls that match all of them; a filter the SDK cannot evaluate, such as the model of a LiteLLM proxy request, keeps the rule applying.
|
|
@@ -2533,7 +2611,7 @@ for budget in budgets or []:
|
|
|
2533
2611
|
|
|
2534
2612
|
Each row has `source` `ANOMALY` and `action` `NOTIFY`, groups by `SUBSCRIBER` with a per-person `groupBreakdown`, and carries `readingAt` when the current window has a reading. The call returns `[]` when the team has no advisory budgets and `None` when the enforcement API cannot be reached.
|
|
2535
2613
|
|
|
2536
|
-
The background poller is unaffected: it keeps reading the **whole team's** rules every `REVENIUM_CB_POLL_INTERVAL_SECONDS
|
|
2614
|
+
The background poller is unaffected: it keeps reading the **whole team's** rules every `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s). That is deliberate. The server computes the department-budget maps team-wide and attaches them to the team-wide read, so a poll narrowed to a single rule would stop receiving them and department budgets would quietly stop blocking anyone. Narrowing is an explicit, opt-in inspection call and never the refresh.
|
|
2537
2615
|
|
|
2538
2616
|
### End-to-End Example
|
|
2539
2617
|
|
|
@@ -2719,7 +2797,9 @@ or unsent at the end are dropped and logged once, as a warning with a count.
|
|
|
2719
2797
|
|
|
2720
2798
|
A process ended by SIGTERM's default action skips `atexit`, and with it that
|
|
2721
2799
|
final flush. uvicorn, and so the LiteLLM proxy, ends that way once its own
|
|
2722
|
-
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
|
|
2723
2803
|
of its own, set `REVENIUM_INSTALL_SIGNAL_HANDLERS=1`: SIGTERM then flushes
|
|
2724
2804
|
first and ends the process the same way. A SIGTERM handler your code installed
|
|
2725
2805
|
before importing the SDK is called unchanged, and one installed afterwards
|
|
@@ -2771,7 +2851,7 @@ timeout raises `ValueError`.
|
|
|
2771
2851
|
| `REVENIUM_METERING_BASE_URL` | `https://api.revenium.ai` | Revenium API endpoint |
|
|
2772
2852
|
| `REVENIUM_LOG_LEVEL` | `INFO` | Log level: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL` |
|
|
2773
2853
|
| `REVENIUM_CAPTURE_PROMPTS` | `false` | Enable prompt capture |
|
|
2774
|
-
| `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 |
|
|
2775
2855
|
| `REVENIUM_SELECTIVE_METERING` | `false` | Only meter `@revenium_meter` decorated functions |
|
|
2776
2856
|
| `REVENIUM_TEAM_ID` | - | Team ID for cost lookups and outcome reporting (JobContext team resolution) |
|
|
2777
2857
|
| `REVENIUM_ENVIRONMENT` | - | Deployment environment (auto-detects from `ENVIRONMENT`, `DEPLOYMENT_ENV`) |
|
|
@@ -80,11 +80,11 @@ 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`, `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`
|
|
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`, `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`, `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`
|
|
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
|
+
- **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`
|
|
90
90
|
- **LangChain** (embeddings models from vendors other than OpenAI, Azure OpenAI and Ollama, such as `MistralAIEmbeddings`): `Embeddings.aembed_documents`, `Embeddings.aembed_query`, `Embeddings.embed_documents`, `Embeddings.embed_query`
|
|
@@ -266,6 +266,13 @@ explicit override. Economics metric directions are
|
|
|
266
266
|
Use `CUSTOMER_DECLARED` or `MEASURED` for a baseline override. Use `MEASURED`,
|
|
267
267
|
`SELF_REPORTED`, or `DERIVED` for a period fact override.
|
|
268
268
|
|
|
269
|
+
A baseline can also record the work's throughput and headcount before AI:
|
|
270
|
+
`throughput_value` with a `throughput_period` of `HOUR`, `DAY`, `WEEK`, or
|
|
271
|
+
`MONTH`, an optional `throughput_scope` of `TEAM` (the default) or
|
|
272
|
+
`PER_PERSON`, and `headcount`. Send `throughput_period` whenever
|
|
273
|
+
`throughput_value` is set; the server rejects a period or scope without a
|
|
274
|
+
value.
|
|
275
|
+
|
|
269
276
|
Facts are append-only and keyed on the period, dimension and metric key
|
|
270
277
|
together. Re-appending that tuple supersedes the active fact, and the server
|
|
271
278
|
requires `reason=` on the entry when it does.
|
|
@@ -440,6 +447,8 @@ client.close()
|
|
|
440
447
|
|
|
441
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.
|
|
442
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
|
+
|
|
443
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.
|
|
444
453
|
|
|
445
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.
|
|
@@ -1057,7 +1066,7 @@ response = litellm.completion(
|
|
|
1057
1066
|
)
|
|
1058
1067
|
```
|
|
1059
1068
|
|
|
1060
|
-
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.
|
|
1061
1070
|
|
|
1062
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.
|
|
1063
1072
|
|
|
@@ -1072,6 +1081,10 @@ except calls a spending cap refuses.
|
|
|
1072
1081
|
pip install "revenium-python-sdk[litellm-proxy]" # requires Python 3.10+
|
|
1073
1082
|
```
|
|
1074
1083
|
|
|
1084
|
+
The extra requires **LiteLLM 1.94.0 or newer**. On LiteLLM 1.93.x the proxy cannot
|
|
1085
|
+
load the guardrail from the config below: it only looks for a module file beside
|
|
1086
|
+
`config.yaml` and stops with `Could not find module file`.
|
|
1087
|
+
|
|
1075
1088
|
```yaml
|
|
1076
1089
|
guardrails:
|
|
1077
1090
|
- guardrail_name: "revenium"
|
|
@@ -1096,6 +1109,26 @@ calls: a failed call is recorded only for a request the Revenium guardrail's
|
|
|
1096
1109
|
sends no record. Adding `pre_call` enforces no budget unless
|
|
1097
1110
|
`REVENIUM_CIRCUIT_BREAKER_ENABLED=true`.
|
|
1098
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
|
+
|
|
1099
1132
|
**Budget enforcement** reuses the SDK's own circuit breaker, so a proxy enforces
|
|
1100
1133
|
exactly what every other Revenium integration enforces — including department
|
|
1101
1134
|
budgets. An organization cap is checked against the organization the call names,
|
|
@@ -1119,7 +1152,7 @@ the call proceeds, and it proceeds without waiting. The pre-call check reads the
|
|
|
1119
1152
|
rules the SDK already has cached and refreshes them on a background thread, so a
|
|
1120
1153
|
Revenium endpoint that hangs, refuses connections, returns errors or rejects your
|
|
1121
1154
|
key (403) neither blocks nor slows the proxy. A failed refresh is retried once
|
|
1122
|
-
per `REVENIUM_CB_POLL_INTERVAL_SECONDS
|
|
1155
|
+
per `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s), not once per request. Right after the
|
|
1123
1156
|
proxy starts, before the first rules fetch completes, no rules are cached and
|
|
1124
1157
|
calls go through; see [Fail-Open vs Fail-Closed](#fail-open-vs-fail-closed).
|
|
1125
1158
|
Metering is likewise non-disruptive — nothing in the post-call path can turn a
|
|
@@ -1317,7 +1350,7 @@ Blank or unset means on; `true`, `1`, `yes`, `on` mean on; any other value turns
|
|
|
1317
1350
|
New calls go straight back to two records when you do; records already merged stay
|
|
1318
1351
|
merged. Four things have to be true for it to work:
|
|
1319
1352
|
|
|
1320
|
-
* **LiteLLM 1.
|
|
1353
|
+
* **LiteLLM 1.94.0 or newer.** That is the floor of the `litellm-proxy` extra, which
|
|
1321
1354
|
is what installs the guardrail. On an older LiteLLM the guardrail logs one warning
|
|
1322
1355
|
at startup and behaves exactly as it does when opted out: no header is added,
|
|
1323
1356
|
the provider's response id is reported, and you get two records rather than none.
|
|
@@ -1368,6 +1401,44 @@ return of double counting. Your other routes never carry an identifier and are n
|
|
|
1368
1401
|
counted or warned about, so a proxy that also serves chat completions or embeddings
|
|
1369
1402
|
stays quiet.
|
|
1370
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
|
+
|
|
1371
1442
|
##### Migrating from the callback
|
|
1372
1443
|
|
|
1373
1444
|
`revenium_middleware.litellm.proxy.middleware.MiddlewareHandler` — the
|
|
@@ -1796,7 +1867,8 @@ Add business context to a call the SDK records by passing a `usage_metadata` dic
|
|
|
1796
1867
|
These fields are shortened with a WARNING when longer than Revenium stores, so the call is still
|
|
1797
1868
|
recorded: `trace_id`, `task_type`, `agent`, `subscription_id`, `parent_transaction_id`, `environment`,
|
|
1798
1869
|
`region`, `transaction_name`, `operation_subtype`, `subscriber.id` and `subscriber.email` at 255
|
|
1799
|
-
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.
|
|
1800
1872
|
|
|
1801
1873
|
| Field | Description | Use Case |
|
|
1802
1874
|
|-------|-------------|----------|
|
|
@@ -1814,7 +1886,7 @@ characters, and `product_name` at 224. Other text fields are sent as given.
|
|
|
1814
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 |
|
|
1815
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._-]+$`) | Split coding-assistant spend between the terminal, the SDKs and the editor or chat integrations |
|
|
1816
1888
|
| `ccr_session_id` | The Claude Code cloud session the call ran in; `ccrSessionId` also accepted (completions only, 1 to 128 chars, `^[A-Za-z0-9_-]+$`) | Group the spend of one Claude Code cloud session |
|
|
1817
|
-
| `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 |
|
|
1818
1890
|
|
|
1819
1891
|
**Example:**
|
|
1820
1892
|
|
|
@@ -1872,7 +1944,13 @@ where the provider says the inference ran. Both are values you supply per call;
|
|
|
1872
1944
|
the SDK does not read them from the environment or from the provider response.
|
|
1873
1945
|
`inference_geo` is sent exactly as given, with no lowercasing or allow-list;
|
|
1874
1946
|
the backend normalizes it. It is not stored as a dimension to group spend by,
|
|
1875
|
-
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
|
|
1876
1954
|
for `ccr_session_id`, 16 for `inference_geo`) or holds anything but letters,
|
|
1877
1955
|
digits, hyphens and underscores is left off the record with a WARNING, and the
|
|
1878
1956
|
call is still recorded. Both are sent on completions only; image, video and
|
|
@@ -2274,7 +2352,7 @@ REVENIUM_ENFORCEMENT_BASE_URL=https://api.revenium.ai/profitstream # optional:
|
|
|
2274
2352
|
| `REVENIUM_BYPASS` | `false` | When `true`, every `check_enforcement` call short-circuits to a no-op. Useful for incident response. |
|
|
2275
2353
|
| `REVENIUM_TEAM_ID` | — | Hashed team ID. Path component on rule fetches; required when the breaker is enabled. |
|
|
2276
2354
|
| `REVENIUM_ENFORCEMENT_BASE_URL` | `https://api.revenium.ai/profitstream` | Base URL for the enforcement API, taken as given. When unset it is the address outcomes and jobs use (`REVENIUM_PROFITSTREAM_BASE_URL`, else the Revenium address `REVENIUM_METERING_BASE_URL` or `initialize_metering(base_url=...)` sits under, without its `/meter` segment) plus `/profitstream`. A rules fetch that answers 404 is retried once at the same address with `/profitstream` removed, or added if it was missing, and the address that answers with rules is used from then on, so a self-hosted server that serves the API at the root, or a value set without `/profitstream`, still reaches the rules. |
|
|
2277
|
-
| `REVENIUM_CB_POLL_INTERVAL_SECONDS` | `
|
|
2355
|
+
| `REVENIUM_CB_POLL_INTERVAL_SECONDS` | `30` | Seconds between background rule refreshes, and the wait before a failed refresh is retried. Whole seconds from a minimum of `5` to a maximum of `3600`: a lower value is raised to `5` with one warning per process start, and a higher one is capped at `3600`. Anything other than plain digits after trimming spaces (a sign, a decimal point, other text), an empty value and `0` use the default. Read once when polling starts, so a change takes effect when the process restarts. While calls are arriving, rules older than 120 seconds are refreshed on the next call whatever the interval, so a value above `120` only spaces out refreshes on an idle process and the retry after a failed refresh. |
|
|
2278
2356
|
| `REVENIUM_CB_FAIL_MODE` | `open` | `open` (default) lets calls through when no cache exists; `closed` raises `BudgetExceededError` until rules are loaded. |
|
|
2279
2357
|
| `REVENIUM_CACHE_DIR` | unset | When set, the rule cache is mirrored to `<dir>/revenium_enforcement_rules.<scope>.json` so a restarted process doesn't fail-closed on the very first call. `<scope>` is a short hash of the enforcement base URL and `REVENIUM_TEAM_ID`, so services for different teams can share one folder and each loads only its own rules. Nothing is saved without `REVENIUM_TEAM_ID`. |
|
|
2280
2358
|
|
|
@@ -2307,9 +2385,9 @@ LangChain embeddings models run no callbacks: `OpenAIEmbeddings` and `AzureOpenA
|
|
|
2307
2385
|
When the circuit breaker is disabled, it is a no-op. When enabled:
|
|
2308
2386
|
|
|
2309
2387
|
1. A daemon thread (`revenium-enforcement-poll`) starts on first use.
|
|
2310
|
-
2. It polls `GET {REVENIUM_ENFORCEMENT_BASE_URL}/v2/api/ai/enforcement-rules/{REVENIUM_TEAM_ID}` every `REVENIUM_CB_POLL_INTERVAL_SECONDS` with the `x-api-key` header.
|
|
2388
|
+
2. It polls `GET {REVENIUM_ENFORCEMENT_BASE_URL}/v2/api/ai/enforcement-rules/{REVENIUM_TEAM_ID}` every `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s) with the `x-api-key` header.
|
|
2311
2389
|
3. Rules are cached in-process. The pre-call check only reads that cache: it never makes a network request and never sleeps, so it is safe on an event loop (LiteLLM proxy, `AsyncOpenAI`). When the cache is more than 120 s old the check wakes the poller to refresh it early and carries on with the rules it has. Only one refresh runs at a time per process.
|
|
2312
|
-
4. A refresh that fails (timeout, refused connection, 5xx, or any 4xx such as a 403 for a wrong key or team id) keeps the cached rules, logs a warning, and is not retried for one `REVENIUM_CB_POLL_INTERVAL_SECONDS
|
|
2390
|
+
4. A refresh that fails (timeout, refused connection, 5xx, or any 4xx such as a 403 for a wrong key or team id) keeps the cached rules, logs a warning, and is not retried for one `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s), however much traffic arrives meanwhile. A 4xx answer the SDK does not retry (any but 408 and 429) is logged at WARNING with the URL and the status, so a wrong address (404) or a wrong key or team id (401, 403) is visible without debug logging.
|
|
2313
2391
|
5. `204 No Content` is treated as "no rules configured" — the cache is cleared.
|
|
2314
2392
|
|
|
2315
2393
|
A rule with filters (a cap scoped to an agent, model, provider, subscriber, credential, task type, product or organization) blocks or warns only the calls that match all of them; a filter the SDK cannot evaluate, such as the model of a LiteLLM proxy request, keeps the rule applying.
|
|
@@ -2417,7 +2495,7 @@ for budget in budgets or []:
|
|
|
2417
2495
|
|
|
2418
2496
|
Each row has `source` `ANOMALY` and `action` `NOTIFY`, groups by `SUBSCRIBER` with a per-person `groupBreakdown`, and carries `readingAt` when the current window has a reading. The call returns `[]` when the team has no advisory budgets and `None` when the enforcement API cannot be reached.
|
|
2419
2497
|
|
|
2420
|
-
The background poller is unaffected: it keeps reading the **whole team's** rules every `REVENIUM_CB_POLL_INTERVAL_SECONDS
|
|
2498
|
+
The background poller is unaffected: it keeps reading the **whole team's** rules every `REVENIUM_CB_POLL_INTERVAL_SECONDS` (default 30 s, minimum 5 s). That is deliberate. The server computes the department-budget maps team-wide and attaches them to the team-wide read, so a poll narrowed to a single rule would stop receiving them and department budgets would quietly stop blocking anyone. Narrowing is an explicit, opt-in inspection call and never the refresh.
|
|
2421
2499
|
|
|
2422
2500
|
### End-to-End Example
|
|
2423
2501
|
|
|
@@ -2603,7 +2681,9 @@ or unsent at the end are dropped and logged once, as a warning with a count.
|
|
|
2603
2681
|
|
|
2604
2682
|
A process ended by SIGTERM's default action skips `atexit`, and with it that
|
|
2605
2683
|
final flush. uvicorn, and so the LiteLLM proxy, ends that way once its own
|
|
2606
|
-
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
|
|
2607
2687
|
of its own, set `REVENIUM_INSTALL_SIGNAL_HANDLERS=1`: SIGTERM then flushes
|
|
2608
2688
|
first and ends the process the same way. A SIGTERM handler your code installed
|
|
2609
2689
|
before importing the SDK is called unchanged, and one installed afterwards
|
|
@@ -2655,7 +2735,7 @@ timeout raises `ValueError`.
|
|
|
2655
2735
|
| `REVENIUM_METERING_BASE_URL` | `https://api.revenium.ai` | Revenium API endpoint |
|
|
2656
2736
|
| `REVENIUM_LOG_LEVEL` | `INFO` | Log level: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL` |
|
|
2657
2737
|
| `REVENIUM_CAPTURE_PROMPTS` | `false` | Enable prompt capture |
|
|
2658
|
-
| `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 |
|
|
2659
2739
|
| `REVENIUM_SELECTIVE_METERING` | `false` | Only meter `@revenium_meter` decorated functions |
|
|
2660
2740
|
| `REVENIUM_TEAM_ID` | - | Team ID for cost lookups and outcome reporting (JobContext team resolution) |
|
|
2661
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 = "0.16.0"
|
|
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"
|
|
@@ -110,9 +110,12 @@ litellm = [
|
|
|
110
110
|
# dispatch that meters streamed calls. litellm[proxy] itself requires Python
|
|
111
111
|
# 3.10+, which is why revenium_middleware.litellm.proxy.guardrail carries an
|
|
112
112
|
# import guard for it while the rest of the SDK stays on requires-python >=3.8.
|
|
113
|
+
# 1.94.0 is the first LiteLLM whose config loader (get_instance_fn) imports an
|
|
114
|
+
# installed package: 1.93.x only loads a module file sitting beside config.yaml,
|
|
115
|
+
# so the README's guardrail entry cannot load there.
|
|
113
116
|
litellm-proxy = [
|
|
114
117
|
"wrapt>=1.14.0",
|
|
115
|
-
"litellm[proxy]>=1.
|
|
118
|
+
"litellm[proxy]>=1.94.0",
|
|
116
119
|
]
|
|
117
120
|
|
|
118
121
|
dev = [
|
|
@@ -213,7 +213,8 @@ from .job_history import ( # noqa: E402
|
|
|
213
213
|
get_job_transactions, get_job_types, get_outcome_history, list_jobs,
|
|
214
214
|
)
|
|
215
215
|
from .job_type_economics import ( # noqa: E402
|
|
216
|
-
Baseline, BaselineProvenance,
|
|
216
|
+
Baseline, BaselineProvenance, BaselineThroughputPeriod,
|
|
217
|
+
BaselineThroughputScope, JobTypeEconomics, JobTypeMetricDirection,
|
|
217
218
|
JobTypeMonetizationBasis, JobTypeMonetizationCategory,
|
|
218
219
|
OutcomeMetricProvenance, PeriodFactEntry,
|
|
219
220
|
create_baseline, get_job_type_economics, list_baselines,
|
|
@@ -293,6 +294,8 @@ __all__ = [
|
|
|
293
294
|
"PeriodFactEntry",
|
|
294
295
|
"Baseline",
|
|
295
296
|
"BaselineProvenance",
|
|
297
|
+
"BaselineThroughputPeriod",
|
|
298
|
+
"BaselineThroughputScope",
|
|
296
299
|
"JobTypeEconomics",
|
|
297
300
|
"JobTypeMetricDirection",
|
|
298
301
|
"JobTypeMonetizationBasis",
|