The article emphasizes the need for tailored retrieval strategies in AI systems that handle various types of knowledge sources. It begins with an example involving a multi-functional AI that acts as a librarian, product recommender, and compliance officer. The author recounts early missteps, where treating all knowledge sources the same led to irrelevant results and missed violations. For instance, using a one-size-fits-all approach resulted in the librarian returning out-of-context paragraphs and the compliance officer failing to catch critical forbidden phrases. To address these issues, the article introduces a more nuanced retrieval system. It highlights the importance of understanding the whole document rather than relying on keyword matching. For effective recommendations, the system generates summaries of articles and products, allowing for better semantic relevance. A tiered summarization approach helps manage input size while retaining core meanings. Moreover, an LLM (Large Language Model) reranking process refines search results, focusing on the quality of related readings or product suggestions based on semantic relevance. The article also tackles compliance checks, detailing a three-layer approach to catch violations of health regulations. The system uses precise keyword matching for forbidden phrases while also incorporating contextual judgment to determine the implications of claims like “manage blood sugar.” A caching strategy is highlighted for efficiency, with calculations showing substantial cost savings by reducing redundant compliance checks. Overall, the piece provides practical insights into designing AI systems that need to navigate complex knowledge bases effectively.