Prism OpenTelemetry for Python

OpenTelemetry spans for Prism generations, using the GenAI semantic conventions and the OpenInference keys that Phoenix and Arize read. The Python port of particle-academy/prism-opentelemetry.

Zero runtime dependencies. Python 3.10+.

pip install prism-ai-opentelemetry

from prism_opentelemetry import GenerationContext, PendingTool, TelemetrySubscriber, Usage

telemetry = TelemetrySubscriber(my_tracer)

context = GenerationContext(
    trace_id="run-42", operation="text", provider="anthropic", model="claude-opus-5"
)

telemetry.on_generation_started(context, prompt)
telemetry.on_tool_invoked("run-42", PendingTool(name="search", call_id="call_1", step_index=0))
telemetry.on_step_completed(
    "run-42", 0, "claude-opus-5", "anthropic", Usage(prompt_tokens=812, completion_tokens=64)
)
telemetry.on_generation_completed("run-42", finish_reason="stop", output=response_text)

my_tracer is anything with start_span(name, start_time_nanos=None, parent=None) returning a span with set_attribute, set_status, record_exception and end. Wrap your OpenTelemetry SDK tracer in that method; the package imports no OpenTelemetry code.

Spans

• One root span per generation, with the operation, provider, model and, when given, session.id and user.id.
• A child span per step, with token usage and, when given, cost.
• A child span per tool call (execute_tool <name>), attached to the step it belongs to.
• A failed generation ends its root span with an error status.

Children are parented off the root stored in SpanStore under the trace id, never off the ambient context. Prism's tool loop is recursive, and ambient context does not survive it.

Content is off by default

Prompts, completions and tool arguments are user content, and a span export leaves the application. They are recorded only with capture_content=True.

• Captured content is cut at max_content_length characters (65,536 by default) and marked …[truncated].
• Even with capture on, media bytes (images, audio, video, documents) are replaced by their size unless you also pass capture_media=True. without_media_bytes() is the function that does it.
• The model, provider, token counts and rate limits are not content and are always recorded.
• record_exceptions is on by default and puts the exception's message on the root span. Turn it off if your errors can carry user content.

Rate limits

Pass the provider's quota buckets as RateLimit(name, limit, remaining, resets_at) to on_generation_completed or on the error given to on_generation_failed. They are written as prism.rate_limit.<name>.* attributes, on success as well as on failure.

Parity

prism-parity's opentelemetry-media-content corpus pins the media-byte rule, and all three languages agree on it.

Its opentelemetry-span-attributes corpus does not agree yet. Rate-limit attributes are identical in PHP, TypeScript and Python. These still differ:

• the operation name and span name, which the PHP reference maps onto the GenAI vocabulary and the ports pass through;
• the span kind for image and stream operations;
• the status of a successful span, unset in PHP and ok in the ports;
• gen_ai.response.finish_reasons;
• captured content, which Python encodes with spaces after separators and escaped non-ASCII;
• where truncation cuts a long value.

Spans from a PHP application and a Python agent in the same trace show those differences.

License

MIT. See LICENSE.