Summary
Currently, the embeddings tracing span emits full embedding vectors and serialized response payloads as span attributes (e.g., output.value, embedding.embeddings.{index}.embedding.vector). While this conforms to the OpenInference/OpenTelemetry specification, it can result in very large span payloads for high-dimension or batched embeddings, risking dropped traces and high telemetry overhead.
Redaction support should be added to allow users to optionally suppress or truncate these large attributes (e.g., replace full vectors with compact metadata such as count and dimension).
Context
Acceptance Criteria
/cc @bzp2010
Summary
Currently, the embeddings tracing span emits full embedding vectors and serialized response payloads as span attributes (e.g.,
output.value,embedding.embeddings.{index}.embedding.vector). While this conforms to the OpenInference/OpenTelemetry specification, it can result in very large span payloads for high-dimension or batched embeddings, risking dropped traces and high telemetry overhead.Redaction support should be added to allow users to optionally suppress or truncate these large attributes (e.g., replace full vectors with compact metadata such as count and dimension).
Context
Acceptance Criteria
output.value,embedding.embeddings.{index}.embedding.vector)/cc @bzp2010