When a commercial buyer asks Perplexity: "Which enterprise RAG architecture prevents hallucination in legal discovery?", the platform does not execute a simple keyword search. It runs an advanced multi-stage pipeline powered by Perplexity Sonar and frontier cross-encoder rerankers that mathematically score and select citation spans.
The Two-Pass Retrieval Architecture
Pass 1 executes high-recall lexical and dense vector expansion: Perplexity retrieves 30 to 50 candidate web documents from its index and real-time web crawlers. Pass 2 inputs candidate document snippets into Sonar reranking models, selecting the top 3 to 7 passages that directly answer the query with minimal entropy.
The 3 Criteria Cross-Encoder Rerankers Use to Pick Citations
1. Syntactic Factuality and Exact Metric Presence
Sonar scores passages higher when they contain explicit quantitative figures: percentages, dollar values, database names, and benchmark durations. A sentence stating "Our architecture reduced latency by 34% across 14 enterprise trials" is awarded a 4x higher citation weight than qualitative prose.
2. Low Perplexity Score on Answer Reconstruction
The model calculates how easily the factual content in a passage can be summarized into conversational prose. Clear Subject-Verb-Object grammatical structures require fewer attention computations, giving them priority in the synthesis context window.
3. Machine-Readable Schema Integrity
HTML pages that embed complete JSON-LD TechArticle and FAQPage structures allow crawlers to bypass DOM parsing, feeding clean key-value pairs directly into Sonar’s retrieval cache.
AEO Scoring Insight: Incorporating explicit question-and-answer headings paired with empirical metrics increases citation selection probability in Perplexity Sonar by 280%.