Multimodal content

Images, audio, video and documents in a request can be recorded on the span - by reference, or uploaded to a store with the span pointing at them. Off by default.

Variable	Default
GENAI_OTEL_MEDIA_CAPTURE_MODE	off	reference_only: type, MIME type and size; nothing kept or uploaded. full: the bytes are redacted (if a redactor is registered), uploaded, and the span carries the URI
GENAI_OTEL_MEDIA_STORE	none	filesystem, http, s3, minio
GENAI_OTEL_MEDIA_STORE_ENDPOINT		Directory (filesystem, default ./genai-otel-media), gateway URL (http), or S3 endpoint (s3 / minio; unset for AWS)
GENAI_OTEL_MEDIA_STORE_BUCKET	genai-otel-media	Bucket, or the subdirectory of the filesystem store
GENAI_OTEL_MEDIA_STORE_PREFIX	traces/{date}/{trace_id}/	Key prefix
GENAI_OTEL_MEDIA_STORE_ACCESS_KEY, ..._SECRET_KEY		S3 credentials; otherwise the AWS SDK's own chain
GENAI_OTEL_MEDIA_MAX_BYTES	10 MiB	Larger parts are not uploaded (stripped_reason = size_exceeded); inline media over it is not even decoded
GENAI_OTEL_MEDIA_ALLOWED_MODALITIES	image,audio,video,document	Others are stripped (modality_not_allowed)
GENAI_OTEL_MEDIA_REDACTOR		See below

Providers: OpenAI (and every client the OpenAI instrumentor traces), Anthropic, Google Gemini, Groq - as in the reference.

On the span

Per message n: gen_ai.prompt.{n}.role. Per part m: gen_ai.prompt.{n}.content.{m}.type, then for media media_uri, media_mime_type, media_byte_size, media_source (inline_offloaded, external_url, reference_only), and gen_ai.media.stripped_reason when the bytes were dropped.

Text needs content capture as well. A text part's .text, and the canonical gen_ai.input.messages (written under the GenAI semconv opt-in), carry what was said, so they are written only with GENAI_ENABLE_CONTENT_CAPTURE on too - unlike the reference, which writes them whenever media capture is on (SDK_ADDITIONS.md section 5). Media capture is its own switch: turning it on is the decision to keep the media.

An upload can take a while; the span waits for it and still ends with the time the call ended. flushTelemetry() waits for spans still uploading.

Redaction

The reference loads a redactor from a Python dotted path, with built-in ones that strip EXIF data, blur faces and scan PDFs. Here a redactor is a function the application registers:

import { setMediaRedactor } from '@traceverse/genai-otel-instrument-node';

setMediaRedactor(async (modality, mimeType, bytes) => stripMetadata(bytes));

It receives the bytes and returns what to upload. If it throws, the bytes are dropped (redactor_error). If GENAI_OTEL_MEDIA_REDACTOR names a redactor and none is registered - a Python dotted path, say - the bytes are dropped (redactor_not_allowed): media is never uploaded unredacted when redaction was asked for. The reference's built-in redactors need imaging libraries and are not ported.

Stores

Store	Runtime
filesystem	Node	Files are written 0600 under a 0700 directory; a key that would leave the store is refused. URI: file://...
http	all	PUT <endpoint>/<key>. Credentials in the endpoint URL are sent as Basic authorization and never appear in the URI on the span
s3, minio	Node	Needs @aws-sdk/client-s3 installed by the application. Path-style addressing; the bucket is created if missing. URI: <endpoint>/<bucket>/<key>, or s3://<bucket>/<key> for AWS

Under a strict profile, or with external egress off, the network stores refuse a plain http:// endpoint, as in the reference.

What was run

The detector, the offload rules and the canonical form are checked against vectors made by running the reference's genai_otel.media (tests/fixtures/media/python-vectors.json). The span attributes are tested through the real OpenAI and Anthropic clients against the mock provider, with the filesystem store. The HTTP and S3 stores are tested against a mock server. Not run against a real MinIO or S3.