Why RAG for Voice Agents and Enterprise Search Are Fundamentally Different Problems

Retrieval-augmented generation (RAG) is commonly treated as a single engineering challenge, but developers building it for both enterprise search and voice agents have found the two environments demand near-opposite design priorities. Enterprise search is relatively forgiving — users tolerate delays of one to two seconds, can scan multiple results, and refine their queries if the first answer falls short. Voice agents, by contrast, operate under strict real-time constraints where even a brief silence mid-call disrupts the user experience and leaves no room for imperfect retrieval. The corpus challenges also differ: enterprise search must handle large, heterogeneous document collections with metadata filters, access controls, and hybrid ranking pipelines, while voice RAG must return accurate results almost instantly. This piece is the first in a three-part series examining what engineers learned while building RAG systems for both contexts.