perplexity-observability

name	perplexity-observability
description	Set up monitoring for Perplexity Sonar API with latency, cost, citation quality, and error tracking. Use when implementing monitoring dashboards, setting up alerts, or tracking Perplexity API health in production. Trigger with phrases like "perplexity monitoring", "perplexity metrics", "perplexity observability", "monitor perplexity", "perplexity dashboard".
allowed-tools	Read, Write, Edit
version	1.12.0
license	MIT
author	Jeremy Longshore <[e-mail]>
tags	["saas","perplexity","monitoring","observability","dashboard"]
compatibility	Designed for Claude Code, also compatible with Codex and OpenClaw

Perplexity Observability

Overview

Monitor Perplexity Sonar API performance, cost, and quality. Key signals unique to Perplexity: citation count per response (quality indicator), search latency variability (web search is non-deterministic), and per-model cost differences.

Key Metrics

Metric	sonar (typical)	sonar-pro (typical)	Alert Threshold
Latency p50	1-2s	3-5s	p95 > 15s
Citations/response	3-5	5-10	0 for 10min
Error rate	<1%	<1%	>5%
Cost/query	$0.005	$0.02	>$0.10

Prerequisites

• Perplexity API integration running
• Metrics backend (Prometheus, Datadog, or custom)
• Alerting system configured

Instructions

Step 1: Instrument the Perplexity Client

import OpenAI from "openai";

interface SearchMetrics {
  model: string;
  latencyMs: number;
  status: "success" | "error";
  citationCount: number;
  totalTokens: number;
  cached: boolean;
  errorCode?: number;
}

const metrics: SearchMetrics[] = [];

async function instrumentedSearch(
  client: OpenAI,
  : ,
  :  = ,
  :  = 
): <{ : ; :  }> {
   start = performance.();
   : ;

   {
     response =  client...({
      model,
      : [{ : , : query }],
    });

    searchMetrics = {
      model,
      : performance.() - start,
      : ,
      : (response  ).?. || ,
      : response.?. || ,
      cached,
    };

    metrics.(searchMetrics);
     { response, : searchMetrics };
  }  (: ) {
    searchMetrics = {
      model,
      : performance.() - start,
      : ,
      : ,
      : ,
      cached,
      : err.,
    };

    metrics.(searchMetrics);
     err;
  }
}