Privacy Controls

Implement data filtering, PII masking, and selective tracing suppression in OpenInference

Privacy Controls

OpenInference provides comprehensive privacy controls to help you maintain security and compliance while still benefiting from observability. You can mask PII, filter sensitive data, and selectively suppress tracing.

Overview

Privacy controls in OpenInference allow you to:

• Mask sensitive data - Replace PII with "__REDACTED__" placeholder
• Filter specific content - Hide images, text, or embeddings selectively
• Suppress tracing - Completely disable tracing for specific code blocks
• Control payload size - Limit large data like base64-encoded images

Redacted Value Placeholder

When content is hidden, OpenInference uses the constant "__REDACTED__" as a placeholder. This allows trace consumers to identify that content was intentionally hidden rather than missing or empty.

The redacted value constant is available for import:

Python

from openinference.instrumentation import REDACTED_VALUE
print(REDACTED_VALUE)  # "__REDACTED__"

PII Masking

Hiding Text Content

Protect user messages and LLM responses while preserving trace structure:

Python

from openinference.instrumentation import TraceConfig

config = TraceConfig(
    hide_input_text=True,
    hide_output_text=True,
)

This configuration:

• Replaces message text with "__REDACTED__"
• Preserves message structure and metadata
• Maintains trace flow and timing information

Hiding Complete Messages

Remove entire message objects from traces:

Python

config = TraceConfig(
    hide_input_messages=True,
    hide_output_messages=True,
)

Hiding All Inputs/Outputs

Hide both values and messages completely:

Python

config = TraceConfig(
    hide_inputs=True,
    hide_outputs=True,
)

Note: Setting hide_inputs=True automatically hides input messages. Setting hide_input_messages=True hides only messages but not the input value.

Data Filtering

Hiding Images

Remove images from traces while keeping text:

Python

config = TraceConfig(
    hide_input_images=True,
)

This is useful when:

• Images may contain sensitive or identifying information
• You want to reduce storage costs
• Compliance requires removing visual data

Limiting Image Size

Truncate large base64-encoded images:

Python

config = TraceConfig(
    base64_image_max_length=8000,  # Limit to 8KB
)

Images exceeding this length are replaced with "__REDACTED__".

Hiding Embeddings

Protect embedding vectors and their associated text:

Python

config = TraceConfig(
    hide_embeddings_vectors=True,
    hide_embeddings_text=True,
)

Suppressing Tracing

Context Manager (Python)

Completely disable tracing for specific code blocks:

from openinference.instrumentation import suppress_tracing

# Normal tracing occurs here
response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello"}]
)

# No tracing occurs within this block
with suppress_tracing():
    sensitive_response = client.chat.completions.create(
        model="gpt-4",
        messages=[{"role": "user", "content": "SSN: 123-45-6789"}]
    )

# Normal tracing resumes here

Async Support (Python)

from openinference.instrumentation import suppress_tracing

async def process_sensitive_data():
    async with suppress_tracing():
        # No tracing occurs in this async block
        result = await async_llm_call()
    return result

JavaScript Suppression

Use OpenTelemetry’s built-in suppression:

import { suppressTracing } from "@opentelemetry/core";
import { context } from "@opentelemetry/api";

// Normal tracing
const response1 = await openai.chat.completions.create({
  model: "gpt-4",
  messages: [{ role: "user", content: "Hello" }],
});

// Suppress tracing for this call
await context.with(suppressTracing(context.active()), async () => {
  const sensitiveResponse = await openai.chat.completions.create({
    model: "gpt-4",
    messages: [{ role: "user", content: "SSN: 123-45-6789" }],
  });
});

Common Privacy Scenarios

GDPR Compliance

Minimize personal data collection:

config = TraceConfig(
    hide_input_text=True,
    hide_output_text=True,
    hide_embeddings_text=True,
    hide_input_images=True,
)

HIPAA Compliance

Protect healthcare information:

config = TraceConfig(
    hide_inputs=True,
    hide_outputs=True,
    hide_embeddings_vectors=True,
    hide_embeddings_text=True,
)

Financial Services

Hide transaction details and account numbers:

config = TraceConfig(
    hide_input_text=True,
    hide_output_text=True,
    hide_llm_invocation_parameters=True,
)

Development vs. Production

Use environment-specific configurations:

import os
from openinference.instrumentation import TraceConfig

if os.getenv("ENVIRONMENT") == "production":
    config = TraceConfig(
        hide_input_text=True,
        hide_output_text=True,
    )
else:
    # Development: full observability
    config = TraceConfig()

Cost Optimization

Reduce storage costs while maintaining observability:

config = TraceConfig(
    hide_input_images=True,
    base64_image_max_length=4000,
    hide_embeddings_vectors=True,
)

Advanced Patterns

Selective Message Filtering

Hide text but keep images for debugging:

config = TraceConfig(
    hide_input_text=True,
    hide_output_text=True,
    hide_input_images=False,  # Keep images
)

Embedding Privacy

Hide vectors but keep text for searchability:

config = TraceConfig(
    hide_embeddings_vectors=True,
    hide_embeddings_text=False,
)

LLM-Specific Controls

Hide only invocation parameters:

config = TraceConfig(
    hide_llm_invocation_parameters=True,  # Hide model config
    hide_inputs=False,  # Show inputs
    hide_outputs=False,  # Show outputs
)

Completions API Privacy

For legacy completions API:

config = TraceConfig(
    hide_prompts=True,  # Hide prompt strings
    hide_choices=True,  # Hide completion choices
)

Best Practices

1. Start restrictive - Begin with more privacy controls and relax as needed
2. Test in development - Verify traces contain expected information before deploying
3. Document your choices - Keep a record of which privacy controls are enabled and why
4. Use environment variables - Make privacy settings configurable per environment
5. Review regularly - Audit your privacy settings as requirements change
6. Combine methods - Use both TraceConfig and suppress_tracing for comprehensive control
7. Monitor impact - Ensure privacy controls don’t hide critical debugging information

Validation

Verify your privacy controls are working:

from openinference.instrumentation import TraceConfig, REDACTED_VALUE
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor, ConsoleSpanExporter

# Set up tracing with privacy controls
tracer_provider = TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(ConsoleSpanExporter()))

config = TraceConfig(hide_input_text=True)

# Instrument and test
from openinference.instrumentation.openai import OpenAIInstrumentor
OpenAIInstrumentor().instrument(tracer_provider=tracer_provider, config=config)

# Make a call and check console output for REDACTED_VALUE

Next Steps

• See TraceConfig for complete configuration reference
• See Environment Variables for deployment options