Which AI algorithms are used by Perplexity?

Perplexity uses a multi-model orchestration approach rather than relying on a single AI algorithm. Its architecture integrates several leading large language models (LLMs) and retrieval systems to deliver accurate, up-to-date, and well-cited answers. Here’s how Perplexity’s AI stack stands out:

• Model-Agnostic Orchestration: Perplexity can dynamically route queries to multiple state-of-the-art LLMs—including GPT-4o, Claude 3.7 Sonnet, Gemini Flash 2.0, Llama 3, DeepSeek R1, and its own proprietary models like Sonar (based on Llama 3.3) and R1 1776 (based on DeepSeek R1). This allows it to select the best model for each task, balancing speed, cost, and quality^[1][2][3][4].
• Retrieval-Augmented Generation (RAG): At its core, Perplexity uses a sophisticated RAG pipeline. This means it combines real-time web data retrieval with LLMs, ensuring answers are both current and grounded in external, verifiable sources^[5][2].
• Hybrid Retrieval Engine: Perplexity’s retrieval system blends vector search (semantic similarity) with keyword search to maximize recall and precision. This hybrid engine surfaces the most relevant documents for each query^[5].
• Neural Re-Ranking and Context Fusion: After initial retrieval, a neural re-ranking model (e.g., DeBERTa-v3 cross-encoder) refines the candidate set. Contextual fusion—using models like T5—chunks and enriches the input with metadata before answer generation^[5].
• Transparent Citations: Perplexity’s output module attaches inline citations and confidence scores to each response, enhancing transparency and trustworthiness^[5][3].
• User Feedback Loop: Answers are continuously improved through user feedback and reinforcement learning, updating retrieval and ranking models over time^[5].
• Multi-Modal Capabilities: Perplexity supports not just text, but also image and document queries, leveraging the strengths of different LLMs for various input types^[2][3].

In summary:
Perplexity’s AI is not a single algorithm, but a flexible, orchestrated system that leverages the strengths of multiple cutting-edge LLMs (like GPT-4o, Claude, Gemini, Llama, DeepSeek, and its own Sonar), advanced retrieval-augmented generation, and hybrid search techniques to provide accurate, cited, and up-to-date answers^{[1][5][2][3][4]}.

⁂

• https://www.pymnts.com/artificial-intelligence-2/2025/perplexity-ai-gains-traction-by-understanding-business-needs/  
• https://www.keysight.com/blogs/en/tech/nwvs/2025/05/19/perplexityai-har-analysis    
• https://en.wikipedia.org/wiki/Perplexity_AI    
• https://www.queencaffeineai.com/post/perplexity-ai-review-2025  
• https://www.linkedin.com/pulse/perplexityai-architecture-overview-2025-priyam-biswas-3mekc