@juspay/neurolink/dist/observability/utils/safeMetadata.js

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

/**
 * Safe metadata filtering for observability exporters.
 *
 * Only these attribute keys are forwarded to third-party backends as trace
 * metadata. User prompts (input), LLM responses (output), error stacks, and
 * any other potentially sensitive data are excluded to prevent PII leaks.
 */
// Only ai.* keys are forwarded as metadata. Stream metrics (chunk_count,
// content_length) should be accessed via span attributes directly, not via
// metadata sent to third-party backends.
export const SAFE_METADATA_KEYS = new Set([
    "ai.provider",
    "ai.model",
    "ai.temperature",
    "ai.max_tokens",
]);
export function filterSafeMetadata(attributes) {
    const filtered = {};
    for (const key of SAFE_METADATA_KEYS) {
        if (attributes[key] !== undefined) {
            filtered[key] = attributes[key];
        }
    }
    return filtered;
}