Configure the Python agent for AI applications (0.1.14 and higher)

Determines if the OTLP metric exporter reports cumulative totals, deltas, or low-memory-friendly temporality for emitted metrics.

Enriches the Python logger to include trace and span correlation fields. Defaults to true.

Accepted values: true, false

Determines if input, output, and system messages are included in spans and logs (as events).

Designates the telemetry to include messages on. Options include spans as attributes, events as bodies, or both.

Accepted values:

• NO_CONTENT (default)

• SPAN_AND_EVENT (for instrumentation-side evaluations)

• SPAN_ONLY (for platform-side evaluations)

• EVENT_ONLY

This setting uses the values defined by the OpenTelemetry semantic conventions. This list of values is provided as a frame of reference. For the latest values,

see OpenTelemetry OpenAI Agents Instrumentation in the opentelemetry-python-contrib GitHub repository.

Controls what telemetry data is generated and emitted during GenAI operations, such as LLM calls and agent invocations. Defaults to span.

Accepted values:

• span

• span_metric (for platform-side evaluations)

• span_metric_event (for instrumentation-side evaluations)

• span_metric_event,splunk

  Note: To use the span_metric_event,splunk value, you must first install the required package by running pip install splunk-otel-genai-emitters-splunk.

Enables opt-in debug logging for GenAI telemetry operations. Helps troubleshoot instrumentation issues by logging internal events without dumping full message content.

Accepted values: true, false

Determines the metric types that are run by the evaluator for AgentInvocation and LLMInvocation. If this variable isn't set, all registered evaluators are enabled with their default metric sets and default thresholds.

Examples of accepted values:

• Deepeval: The default value. Runs the default DeepEval bundle, which measures all of the metrics for every metric type (bias, toxicity, relevance, hallucination, and sentiment).

• Deepeval(LLMInvocation(bias,toxicity),AgentInvocation(hallucination)): Runs DeepEval evaluations only for bias and toxicity for LLM invocations, and only for hallucination for agent invocations.

• Deepeval(LLMInvocation(bias(threshold=0.75)),AgentInvocation(bias(threshold=0.5))): Overrides the bias threshold. Sets a bias level of over 0.75 as positive for LLM invocations. Sets a bias level of over 0.5 as positive for agent invocations. Does not run the remaining metrics.

• none: Deactivates the evaluator.

Condenses evaluations into a single event. Defaults to false.

Accepted values: true, false

Determines the sample rate of traces for evaluations. Sampling decisions are made probabilistically based on rate, where rate is the probability (between 0.0 and 1.0) that a span will be sampled for evaluation.

If this setting isn't configured, it defaults to 1 and all spans will be sampled for evaluation.

Accepted values: Values between 0.0 and 1.0.

Determines whether the instrumentation framework runs evaluations in a separate process. Use true to run evaluations in a child process with the OpenTelemetry SDK deactivated and prevent evaluator LLM calls from polluting application telemetry.

Defaults to false, meaning that the instrumentation frameworks run evaluations in the same process as your application. LLM calls from evaluators such as DeepEval are instrumented alongside application telemetry.

Accepted values: true, false

• Required for OpenAI instrumentation when evaluations are enabled

• Optional for all other instrumentation frameworks

Customizes which emitters handle evaluation results.

Accepted values: replace-category:SplunkEvaluationResults

Enables capture of gen_ai.tool.definitions on LLM invocation spans.

Requires OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT to also be enabled. When both variables are set to truthy values, the JSON-serialized tool/function schemas passed to the model are recorded as a span attribute.

Defaults to false because the payload can be large.

Accepted values: Any truthy or falsy value

Activates or deactivates automatic propagation of GenAI context (conversation_id and association properties) to child spans.

Defaults to true. Set to false to prevent context attributes from being automatically copied to nested GenAI invocations. When deactivated, only values explicitly set on each invocation object are emitted.

Comma-separated list of GenAI context attribute keys to include as metric dimensions. Default value is empty; context attributes aren't included on metrics.

gen_ai.conversation.id
user.id,customer.id
all

When set to a truthy value, the root GenAI span created by instrumentation frameworks uses the Workflow type instead of the default AgentInvocation type. Set to truthy value if you need an explicit workflow root.

Defaults to false, meaning the root span uses the AgentInvocation type. This type aligns with OpenTelemetry GenAI semantic conventions.

Accepted values: Any truthy or falsy value

Controls whether GenAI content log events (such as gen_ai.client.inference.operation.details) are emitted. Accepted values: true or false.

When unset, this setting defaults to the OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT mode:

• true for EVENT_ONLY and SPAN_AND_EVENT

• false for NO_CONTENT and SPAN_ONLY

Polling interval in seconds for the evaluation worker loop. Defaults to 5.0 seconds.

Maximum size of the evaluation queue. When set to a positive integer, the queue becomes bounded and will apply backpressure when full, which means that new items are dropped with a warning.

When unset or set to 0, the queue is unbounded. This is the default behavior.

Recommended values: 100 to 1000, depending on memory constraints and evaluation throughput requirements.

Enables concurrent evaluation processing. When set to a truthy value, evaluations are processed concurrently using multiple worker threads and asynchronous LLM calls.

When unset or set to a falsy value, evaluations are processed sequentially. Concurrent mode significantly improves throughput for LLM-as-a-judge evaluations.

Determines the trace-id ratio sampling rate for evaluations.

Accepted values: Values between 0.0 and 1.0.

Determines if logs are created when measurements are skipped.

Accepted values: true, false

Determines if a log is created for each evaluation result.

Accepted values: true, false

Determines if DeepEval can write temporary artifacts to the filesystem.

Use READ_ONLY in locked down environments. Defaults to READ_WRITE.

Accepted values: READ_WRITE, READ_ONLY

The custom LLM endpoint URL. Required if you want to use a custom LLM provider for DeepEval evaluations instead of OpenAI.

This setting creates a LiteLLMModel configured for the custom endpoint. It supports both static API keys and OAuth2 token-based authentication.

A JSON-formatted string containing key-value pairs that will be added as HTTP headers to all LLM API requests. Use this setting if your API gateway requires custom headers for authentication or tracking.

Example value: '{"system-code": "APP-123", "x-tenant-id": "tenant-abc"}'.

LiteLLM does not natively support setting extra_headers with environment variables and requires programmatically passing this parameter. This setting is provided for DeepEval users who need to configure custom headers without code changes. For more details on headers in LiteLLM, see Optional Fields in the LiteLLM documentation.

The OAuth2 token endpoint. Used for providers that require OAuth2 token-based authentication. This setting enables OAuth2 mode for DeepEval.

Example value: https://identity.example.com/oauth2/default/v1/token

The OAuth2 client ID.

Used for providers that require OAuth2 token-based authentication. Requires DEEPEVAL_LLM_TOKEN_URL to be set.

The OAuth2 client secret. Used for providers that require OAuth2 token-based authentication. Requires DEEPEVAL_LLM_TOKEN_URL to be set.