Building Real-Time Agentic Websites with RAG

The Breakthrough

Carlos Sanchez demonstrates a system that generates hyper-personalized web pages in under 1.2 seconds by dynamically composing pre-defined content blocks based on user intent, rather than generating entire pages from scratch.

What Actually Worked

• Block-Based Composition: Instead of generating raw HTML, the system treats the website as a collection of modular blocks (hero, product lists, navigation, CTAs). The LLM acts as an orchestrator that selects and arranges these blocks based on the user's current intent.
• RAG-Grounded Generation: To prevent hallucinations and maintain strict brand guidelines, the system uses the entire existing website as a corpus for a Retrieval-Augmented Generation (RAG) pipeline. The LLM only fills in content within the constraints of the retrieved site data.
• Per-Site Model Evaluation: The team uses the Prompu tool to continuously evaluate various LLM providers and models against specific site requirements, prioritizing inference speed over model size to ensure the generation process stays under the critical 1-2 second latency threshold.
• Intent-Driven Bucketing: User browsing signals (pages visited, time spent, search queries) are captured and used to bucket users into personas. This data is fed into the LLM to tailor the page structure and recommendations for an "audience of one."

Before / After

• Latency Comparison: Using Cerebras with the Gemma 4 model, the system achieved an average generation latency of 1.1 seconds, compared to 4.6 seconds for the runner-up configuration.
• Throughput: The Cerebras implementation sustained approximately 2,300 tokens per second during page generation.

Context

Adobe Experience Manager is exploring how to move beyond static, one-size-fits-all web experiences. By leveraging edge delivery services and fast inference, they aim to provide marketers with tools to automate site authoring while maintaining brand integrity. The architecture focuses on intent-driven personalization, where the site adapts its layout and copy in real-time as the user navigates or performs specific searches.

Content References

• Tool: Prompu, mentioned as the evaluation framework for models and prompts.