Perplexity Reference Architecture

Contents

• Overview
• Prerequisites
• Instructions
• Output
• Error Handling
• Examples
• Resources

Overview

Production architecture for AI-powered research and search with Perplexity Sonar API. Covers model routing for cost/quality tradeoffs, citation extraction, multi-query research pipelines, and conversational search.

Prerequisites

• Perplexity API key (Sonar access)
• OpenAI-compatible client library
• Understanding of search models (sonar, sonar-pro, sonar-reasoning)

Architecture

Application Layer (Research Agent, Fact Checker, Content Writer)
        |
Search Router (sonar/sonar-pro/sonar-reasoning)
        |
Citation Pipeline (Extract URLs, Validate, Store, Render)
        |
Cache Layer (Query Hash -> Result, TTL by freshness need)

Instructions

Step 1: Set Up Search Service with Model Routing

Use OpenAI-compatible client pointed at api.perplexity.ai. Route queries by depth: quick/standard (sonar), deep (sonar-pro), reasoning (sonar-reasoning).

Step 2: Build Citation Extraction Pipeline

Parse response text for [N] URL patterns and inline URLs. Deduplicate and store citations with metadata.

Step 3: Create Research Pipeline

Multi-phase: broad overview (fast model) -> identify subtopics -> deep dive each (sonar-pro) -> deduplicate citations.

Step 4: Implement Conversational Search

Maintain message history for follow-up questions that build on previous context.

See detailed implementation for search service, citation extraction, multi-query research pipeline, conversational session, and fact-check service code.

Output

• Search service with model routing by query depth
• Citation extraction from responses
• Multi-query research pipeline
• Conversational search with context

Error Handling

Examples

Quick Fact Check

const result = await factCheck("The Earth is approximately 4.5 billion years old");
console.log(result.verdict);  // Accurate, with sources
console.log(result.sources);  // Citation URLs

Resources

• Perplexity API Docs
• Perplexity Model Guide