OM DETTA AVSNITT
Episode 4 of Components Live—recorded on-site at La Botifarra in Valencia—continues the 11-part series with Patrick Bosek and Noz Urbina diving into one of the most important (and often misunderstood) topics in AI: context.
Rather than treating “structured vs. unstructured content” as a binary debate, Patrick and Noz reframe it as part of a broader challenge—how to effectively navigate and deliver the right context to AI systems at the right time. In this episode, they break down how context operates across different layers of AI, starting with the distinction between training and fine-tuning (post-training) versus how large language models are configured and used at question time.
From there, they unpack one of the most talked-about approaches in enterprise AI today: Retrieval-Augmented Generation (RAG). Instead of expecting an AI model to “know” everything, RAG works by retrieving relevant content from source systems and injecting it into the model’s context window—the limited working memory available during a given interaction.
That limitation is key.
Patrick and Noz explain the concept of the context window as a kind of short-term memory for AI systems—one that can become strained as more information is added, sometimes leading to degraded recall or accuracy, especially in the middle of long inputs. This makes upstream retrieval and selection incredibly important.
To make this more tangible, they compare it to how ChatGPT uses web search: it doesn’t read the entire internet on demand—it retrieves and prioritizes a subset of relevant sources. Enterprise AI systems operate the same way, except instead of the open web, they rely on internal documentation, knowledge bases, and structured content systems.
The takeaway? Success with AI isn’t just about the model—it’s about “context engineering.” By improving how content is retrieved, structured, and delivered into the context window, organizations can significantly increase the factual accuracy of AI-generated responses—often moving from roughly 70% reliability into the high 80s or beyond.
This episode sharpens the focus of the series, showing that the future of AI in technical communication isn’t just about generating content—it’s about designing the systems that ensure AI has the right information to begin with.
