perplexity-performance-tuning

# Perplexity Performance Tuning ## Overview Optimize Perplexity Sonar API for latency, throughput, and cost. Key insight: every Perplexity call performs a live web search, so response times are inherently variable. Typical latencies: sonar 1-3s, sonar-pro 3-8s, sonar-deep-research 10-60s. ## Latency Benchmarks | Model | Typical Latency | Max Tokens | Best For | |-------|----------------|------------|----------| | `sonar` | 1-3s | 4096 | Quick answers, simple facts | | `sonar-pro` | 3-8s | 8192 | Deep research, many citations | | `sonar-reasoning-pro` | 5-15s | 8192 | Multi-step analysis | | `sonar-deep-research` | 10-60s | 8192 | Comprehensive reports | ## Prerequisites - Perplexity API key configured - Understanding of search-augmented generation latency patterns - Cache infrastructure (Redis or in-memory LRU) ## Instructions ### Step 1: Smart Model Routing ```typescript import OpenAI from "openai"; const perplexity = new OpenAI({ apiKey: process.env.PERPLEXITY_API_KEY, baseURL: "https://api.perplexity.ai", }); type QueryComplexity = "simple" | "standard" | "deep"; function classifyQuery(query: string): QueryComplexity { const words = query.split(/\s+/).length; const simplePatterns = [/^what is/i, /^who is/i, /^when did/i, /^define/i, /^how many/i]; const deepPatterns = [/compare.*vs/i, /analysis of/i, /comprehensive/i, /pros and cons/i, /in-depth/i]; if (simplePatterns.some((p) => p.test(query)) && words < 15) return "simple"; if (deepPatterns.some((...