SpecificationData Masking

General

Attributes

Span Conventions

Messages and Content

Feedback

Configuration

Data Masking

Prompts, completions, retrieved documents, and media frequently contain personal or sensitive information. The trace configuration settings let an application keep that content out of its traces. This page describes what each setting hides.

Redacted Content

When content is hidden by a setting, its value is replaced with the placeholder "__REDACTED__", or the attribute is omitted. The constant placeholder lets consumers tell that content was intentionally hidden rather than missing or empty.

Inputs and Outputs

  • OPENINFERENCE_HIDE_INPUTS hides input.value, input.images, input.documents, all input messages, and the tool definitions advertised to the LLM (llm.tools.*).
  • OPENINFERENCE_HIDE_OUTPUTS hides output.value, output.images, output.documents, and all output messages.

Messages

  • Input messages are hidden if either OPENINFERENCE_HIDE_INPUTS or OPENINFERENCE_HIDE_INPUT_MESSAGES is true.
  • Output messages are hidden if either OPENINFERENCE_HIDE_OUTPUTS or OPENINFERENCE_HIDE_OUTPUT_MESSAGES is true.

When messages are not already hidden, finer settings hide parts of them:

  • OPENINFERENCE_HIDE_INPUT_TEXT replaces text content in input messages with "__REDACTED__", including message document document.content.
  • OPENINFERENCE_HIDE_OUTPUT_TEXT does the same for output messages.
  • OPENINFERENCE_HIDE_INPUT_IMAGES replaces image URLs in input messages with "__REDACTED__".

LLM Request Details

  • OPENINFERENCE_HIDE_LLM_INVOCATION_PARAMETERS hides llm.invocation_parameters, independent of input and output hiding.
  • OPENINFERENCE_HIDE_LLM_TOOLS hides the tool definitions advertised to the LLM (llm.tools.*). They are also hidden when OPENINFERENCE_HIDE_INPUTS is true.
  • OPENINFERENCE_HIDE_PROMPTS hides the prompts of a completions API call (llm.prompts.*).
  • OPENINFERENCE_HIDE_CHOICES hides the choices of a completions API call (llm.choices.*).

Embeddings

  • OPENINFERENCE_HIDE_EMBEDDINGS_VECTORS replaces each embedding.embeddings.N.embedding.vector value with "__REDACTED__", so vector data is not included in traces. OPENINFERENCE_HIDE_EMBEDDING_VECTORS is a deprecated name for the same setting.
  • OPENINFERENCE_HIDE_EMBEDDINGS_TEXT replaces each embedding.embeddings.N.embedding.text value with "__REDACTED__", so the input text is not included in traces.

Retrieval Documents

OPENINFERENCE_HIDE_RETRIEVAL_DOCUMENTS replaces retrieval.documents.*.document.content and retrieval.documents.*.document.metadata values with "__REDACTED__". OPENINFERENCE_HIDE_OUTPUTS implies it. Only the Python SDK supports this setting today.

Images and Documents

The image settings apply to the span-kind-independent input.images and output.images attributes the same way they apply to message-content images:

  • OPENINFERENCE_HIDE_INPUT_IMAGES removes all input.images.* attributes.
  • OPENINFERENCE_HIDE_INPUTS removes input.images.* and OPENINFERENCE_HIDE_OUTPUTS removes output.images.*, just as they hide input.value and output.value.

Base64 Size Limit

OPENINFERENCE_BASE64_IMAGE_MAX_LENGTH (default 32000) limits the length of a base64 data URI recorded in an image image.url or a document document.url, in messages and in <input|output>.images and <input|output>.documents. The setting keeps its image-specific name but sets the limit for both.

  • A data URI longer than the limit is replaced with "__REDACTED__".
  • If a blob uploader is configured, an over-limit image or document is uploaded instead, and the attribute records the destination URI.
  • Hide settings take precedence: hidden content is never uploaded.
  • Audio and video data URIs are not size-gated, and the shared masking does not yet hide audio or video.

SDK Support

Support for the image and document settings differs by SDK: Python enforces the size limit and can externalize to a blob uploader; JavaScript enforces the size limit and redacts oversized images to "__REDACTED__"; Java and Go apply the hide flags only.

Feedback

Annotation and evaluation explanations and metadata can contain sensitive data and MUST follow the same masking and access policies as other span content.