Context is all you need

A tribute to “Attention is all you need” and the transformative power of context in the era of Large Language Models

When Vaswani et al. published their groundbreaking paper “Attention is all you need” in 2017, they revolutionized how we think about machine learning and language processing. The Transformer mechanism made it possible for models to learn where to direct their “attention.” Today, seven years later, I want to propose a thesis inspired by this insight: Context is all you need.

The Power of Context

Large Language Models (LLMs) have fundamentally changed our understanding of what’s possible with artificial intelligence. But their true strength lies not only in their sheer size or computational power, but in their ability to understand and utilize context. With the right context, even “smaller” models can achieve remarkable results that deliver specific, high-quality, and relevant responses.

The key lies in the realization: the more precise and comprehensive the context, the better the output. It’s like the difference between a conversation with someone who knows nothing about you and a conversation with a longtime friend who knows your history, preferences, and current challenges.

The Four Pillars of Contextual AI Systems

1. User-Defined Context

The simplest and most direct form: the user explicitly provides relevant information. This can be in the form of detailed prompts, uploaded documents, or specific instructions. The more relevant details the user provides, the more tailored the response becomes.

Example: Instead of asking “Write me a marketing plan,” the user provides: “Write me a marketing plan for my B2B SaaS startup in project management, targeting mid-sized companies, budget €50,000 for Q1 2025, focus on LinkedIn and content marketing.”

2. RAG - Retrieval-Augmented Generation

RAG systems extend LLMs with the ability to retrieve relevant information from external knowledge bases. Instead of relying solely on training knowledge, these systems can dynamically access specialized databases, corporate policies, or current documentation.

The advantage: The model doesn’t need to be trained on all the world’s knowledge but can situationally retrieve and contextualize the most relevant information.

3. Web-based Context Extension

Here comes the power of real-time information. Systems that search relevant web sources with each query can incorporate current developments, latest research results, or up-to-date information.

Perplexity is an excellent example of this approach. With each user query, the system automatically performs targeted web searches to deliver the most current and relevant context. The result: answers that are based not only on static training knowledge but also incorporate the latest available information.

4. Dynamic Tool Use in Conversational Flow

The most sophisticated form of contextual work: LLMs that can autonomously decide during a conversation what additional information they need and use appropriate tools – whether search engines, APIs, databases, or specialized services.

These systems can:

• Perform web searches when current information is needed
• Access internal company systems
• Perform calculations or analyze data
• Use external APIs for specific functions

Why Context Makes the Difference

Precision Instead of Hallucination

With sufficient, relevant context, LLMs dramatically reduce their tendency to “hallucinate” – that is, to invent information. Instead, they can rely on concrete, verifiable sources.

Personalization and Relevance

Contextualized responses are not only more accurate but also more relevant to the user’s specific situation. An LLM with access to company policies can provide tailored compliance advice instead of just general recommendations.

Currency

While trained models have a cutoff date, context-enhanced systems can always be up to date. The combination of learned knowledge and current information creates a powerful synergy.

The Future is Contextual

We stand at the threshold of a new era of AI, where it’s not the size of the model that matters, but the quality and relevance of available context. Systems that can intelligently combine different context sources – from user inputs to RAG databases to real-time web research – will define the next generation of AI applications.

Just as “Attention is all you need” opened up new possibilities in neural network architecture, “Context is all you need” shows an important path for the practical application of AI: from isolated, static models to networked, context-aware AI systems.

The message is clear: In a world of infinite information, the bottleneck is not knowledge itself, but the ability to make the right knowledge available at the right time in the right context.

Context is all you need – and the systems that understand this best will shape the future of AI.