TraceConfig

Control what data gets captured. Useful for privacy compliance, reducing payload size, or masking sensitive data.

from fi_instrumentation import TraceConfig

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

# Pass to instrumentors
OpenAIInstrumentor().instrument(
    tracer_provider=trace_provider,
    config=config,
)

TraceAI.init(TraceConfig.builder()
    .baseUrl("https://api.futureagi.com")
    .apiKey("your-key")
    .projectName("my-project")
    .hideInputs(true)
    .hideOutputs(true)
    .hideInputMessages(true)
    .hideOutputMessages(true)
    .build()
);

import { OpenAIInstrumentation } from "@traceai/openai";
import { registerInstrumentations } from "@opentelemetry/instrumentation";

registerInstrumentations({
  tracerProvider,
  instrumentations: [
    new OpenAIInstrumentation({
      traceConfig: {
        hideInputs: true,
        hideOutputs: true,
        hideInputImages: true,
        hideEmbeddingVectors: true,
        base64ImageMaxLength: 16000,
        piiRedaction: true,
      },
    }),
  ],
});

var tracer = TraceAI.Register(opts =>
{
    opts.ProjectName = "my-project";
    opts.TraceConfig = TraceConfig.Builder()
        .HideInputs(true)
        .HideOutputs(true)
        .HideInputImages(true)
        .HideEmbeddingVectors(true)
        .Base64ImageMaxLength(16000)
        .Build();
});

Each field maps to an environment variable with the FI_ prefix (e.g. hide_inputs -> FI_HIDE_INPUTS, hide_error_details -> FI_HIDE_ERROR_DETAILS).

Errors and privacy

When a call fails, traceAI marks the span ERROR and records the exception. Error text can carry the same data you hid with the other hide_* fields: a provider error body can quote your prompt back or include provider and gateway request IDs, and a framework error can include a stack trace with local file paths. hide_error_details keeps the fact that a span failed and what kind of failure it was, and drops the text.

The status description excludes tracebacks in either mode. When error details are enabled, stack traces belong in exception.stacktrace on the exception event.

Default

You rarely need to set hide_error_details yourself. When you don’t set it in code or through FI_HIDE_ERROR_DETAILS, traceAI derives it from the content fields:

The content fields count whether you set them in code or through their own environment variables, so FI_HIDE_INPUTS=true on its own also hides error details. hide_input_images, hide_embedding_vectors, hide_llm_invocation_parameters, and pii_redaction don’t affect it. As with the other fields, a value set in code takes precedence over the environment variable.

Scope

hide_error_details applies to spans created through FITracer and instrumentors that use the shared traceAI span wrapper. In Python, this includes exceptions recorded automatically when they escape a wrapped span’s context manager. Released instrumentors such as traceai-chromadb 0.2.0 use this wrapper; they do not need a separate privacy adapter.

Content masking also covers raw.input, prompt-template variables (llm.prompt_template.variables / gen_ai.prompt.template.variables), and raw.output. In TypeScript, attributes supplied at span creation and link attributes are masked too. Environment variables apply even when you omit traceConfig.

Inside TypeScript FITracer.startActiveSpan, use the OpenTelemetry span interface. The active span is an FISpan wrapper, not the SDK’s concrete Span class.

Example

Hide inputs, outputs, and error details. Configure your Future AGI credentials before registering the provider:

from fi_instrumentation import register, TraceConfig
from fi_instrumentation.fi_types import ProjectType
from traceai_openai import OpenAIInstrumentor

trace_provider = register(
    project_name="my-chatbot",
    project_type=ProjectType.OBSERVE,
)

config = TraceConfig(
    hide_inputs=True,
    hide_outputs=True,
    # Explicitly hide errors, even if the content flags change later
    hide_error_details=True,
)

OpenAIInstrumentor().instrument(
    tracer_provider=trace_provider,
    config=config,
)

import { register, ProjectType } from "@traceai/fi-core";
import { OpenAIInstrumentation } from "@traceai/openai";
import { registerInstrumentations } from "@opentelemetry/instrumentation";

const tracerProvider = register({
  projectName: "my-chatbot",
  projectType: ProjectType.OBSERVE,
});

registerInstrumentations({
  tracerProvider,
  instrumentations: [
    new OpenAIInstrumentation({
      traceConfig: {
        hideInputs: true,
        hideOutputs: true,
        // Explicitly hide errors, even if the content flags change later
        hideErrorDetails: true,
      },
    }),
  ],
});

export FI_HIDE_INPUTS=true
export FI_HIDE_OUTPUTS=true
# Explicitly hide errors, even if the content flags change later
export FI_HIDE_ERROR_DETAILS=true

To hide error details while still capturing content, set hide_error_details=True (or FI_HIDE_ERROR_DETAILS=true) on its own.

Cap attribute size

TraceConfig’s hide_* fields redact or drop values; only base64_image_max_length actually truncates, and only for base64 images. Everything else, an oversized input, output, or tool payload, is sent at full size and can be rejected outright (see ingestion request limits).

To cap the size of every attribute value before it is exported, set an OpenTelemetry span limit instead. This applies to all instrumentors and all attributes, not just images.

# Truncates every span attribute value to 100,000 characters
export OTEL_ATTRIBUTE_VALUE_LENGTH_LIMIT=100000

from fi_instrumentation import register, SpanLimits

trace_provider = register(
    project_name="my_project",
    span_limits=SpanLimits(max_attribute_length=100000),
)

span_limits is a register() argument; there is no per-instrumentor equivalent, so set it once at registration.

PII Redaction (Python)

When pii_redaction=True, the SDK automatically detects and masks 6 types of personally identifiable information:

# Enable via code
config = TraceConfig(pii_redaction=True)

# Or via environment variable
# export FI_PII_REDACTION=true

# Direct usage
from fi_instrumentation.instrumentation.pii_redaction import redact_pii_in_string

redacted = redact_pii_in_string("Email me at test@example.com")
# "Email me at <EMAIL_ADDRESS>"