Axiom AI SDK provides flexible redaction policies to control what data is captured in OpenTelemetry spans. This allows you to balance observability needs with privacy and compliance requirements.

## Built-in redaction policies

Axiom AI SDK provides two built-in redaction policies:

| Policy                              | What gets captured                                                  | What gets excluded                                             | When to use                               |
|-------------------------------------|--------------------------------------------------------------------|--------------------------------------------------------------|-------------------------------------------|
| [AxiomDefault](/content/docs/ai-engineering/redaction-policies#axiomdefault-policy/index.html) | Full data                                                           | –                                                            | Full observability                         |
| [OpenTelemetryDefault](/content/docs/ai-engineering/redaction-policies#opentelemetrydefault-policy/index.html) | Model metadata, token usage, error info                            | Prompt text, AI responses, tool args and results             | Privacy-first                              |

If you don’t specify a redaction policy, Axiom AI SDK applies `AxiomDefault`. To determine which redaction policy fits your needs, see the following comparison:

- AxiomDefault

- OpenTelemetryDefault

### AxiomDefault policy

By default, Axiom AI SDK captures all data for maximum observability. **What gets captured:**

- Full prompt text and AI responses in chat spans
- Complete tool arguments and return values on tool spans
- All standard OpenTelemetry attributes (model name, token usage, etc.)

Capturing full message content increases span size and storage costs.

**When to use:**

- You need full visibility into AI interactions
- Data privacy isn’t a concern
- Debugging complex AI workflows

### OpenTelemetryDefault policy

The OpenTelemetry default policy excludes sensitive content. **What gets captured:**

- Model metadata (name, provider, version)
- Token usage and performance metrics
- Error information and status codes

**What gets excluded:**

- Prompt text and AI responses
- Tool arguments and return values

**When to use:**

- Handling sensitive or personal data
- Compliance requirements restrict data capture
- You only need performance and error metrics

### What gets captured

To determine which redaction policy fits your needs, see the following examples about what gets captured with each default policy:

- AxiomDefault

- OpenTelemetryDefault

- Chat spans

- Tool spans

```
{
  "gen_ai.operation.name": "chat",
  "gen_ai.request.model": "gpt-4o-mini",
  "gen_ai.input.messages": "[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"Hello, how are you?\"}]}]",
  "gen_ai.output.messages": "[{\"role\":\"assistant\",\"content\":\"I'm doing well, thank you for asking!\"}]",
  "gen_ai.usage.input_tokens": 12,
  "gen_ai.usage.output_tokens": 15,
  "gen_ai.usage.total_tokens": 27
}
```

```
{
  "gen_ai.tool.name": "weather_lookup",
  "gen_ai.tool.description": "Get current weather for a location",
  "gen_ai.tool.arguments": "{\"location\":\"San Francisco\",\"units\":\"celsius\"}",
  "gen_ai.tool.message": "{\"temperature\":18,\"condition\":\"partly cloudy\"}"
}
```

- Chat spans

- Tool spans

```
{
  "gen_ai.operation.name": "chat",
  "gen_ai.request.model": "gpt-4o-mini",
  "gen_ai.usage.input_tokens": 12,
  "gen_ai.usage.output_tokens": 15,
  "gen_ai.usage.total_tokens": 27
}
```

Message content (`gen_ai.input.messages` and `gen_ai.output.messages`) is excluded for privacy.

```
{
  "gen_ai.tool.name": "weather_lookup",
  "gen_ai.tool.description": "Get current weather for a location"
}
```

Tool arguments and results (`gen_ai.tool.arguments` and `gen_ai.tool.message`) are excluded for privacy.

## Global configuration

Set a default redaction policy for your entire application using `initAxiomAI`:

Full data capture

Privacy-first

```
import { trace } from '@opentelemetry/api';
import { initAxiomAI, RedactionPolicy } from 'axiom/ai';

const tracer = trace.getTracer("my-tracer");

initAxiomAI({ tracer, redactionPolicy: RedactionPolicy.AxiomDefault });
```

In [Quickstart](/content/docs/ai-engineering/quickstart/index.html), `initAxiomAI` is called in your instrumentation file (`/src/instrumentation.ts`).

## Per-operation override

You can configure different policies for each operation. Axiom resolves redaction policies in the following order (from highest to lowest precedence):

1. Per-operation policy
2. Global policy
3. Default policy

Override the global or default policy for specific operations by passing a `redactionPolicy` to `withSpan`:

```
import { withSpan, RedactionPolicy } from 'axiom/ai';
import { generateText } from 'ai';

const result = await withSpan(
  { capability: 'customer_support', step: 'handle_sensitive_query' },
  async (span) => {
    span.setAttribute('user.id', userId);
    return generateText({
      model: wrappedModel,
      prompt: 'Process this sensitive customer data...'
    });
  },
  { redactionPolicy: RedactionPolicy.OpenTelemetryDefault }
);
```

## Custom redaction policies

Create custom policies by defining an `AxiomAIRedactionPolicy` object:

```
import { trace } from '@opentelemetry/api';
import { initAxiomAI, AxiomAIRedactionPolicy } from 'axiom/ai';

const tracer = trace.getTracer("my-tracer");

// Custom policy: capture messages but not tool payloads
const customPolicy: AxiomAIRedactionPolicy = {
  captureMessageContent: 'full',
  mirrorToolPayloadOnToolSpan: false
};

initAxiomAI({ tracer, redactionPolicy: customPolicy });
```

The `AxiomAIRedactionPolicy` object has two properties:

- **captureMessageContent**

'full' | 'off'

default:"'full'"

Controls whether prompt and response text is included in chat spans.

- `'full'`: Include complete message content
  - `'off'`: Exclude all message content

- **mirrorToolPayloadOnToolSpan**

boolean

default:"true"

Controls whether tool arguments and results are duplicated on tool spans.

- `true`: Mirror tool data for easier querying
  - `false`: Only capture tool metadata (name, description)

The built-in policies configure the `AxiomAIRedactionPolicy` object in the following way:

| Default policy          | captureMessageContent | mirrorToolPayloadOnToolSpan |
|------------------------|-----------------------|------------------------------|
| AxiomDefault           | 'full'                | true                         |
| OpenTelemetryDefault    | 'off'                 | false                        |