Descriptions:
Cole Medin addresses one of the most misunderstood narratives in AI development: the claim that RAG is dead. His argument is precise — traditional RAG, meaning semantic similarity search over embedded chunks in a vector database, is genuinely obsolete for AI coding tools, but remains critical for the majority of applications dealing with unstructured knowledge bases like document repositories, email archives, or SharePoint sites.
The distinction hinges on data structure. Code is perfectly spelled, organized by file paths, and identified by exact syntax — making regex and keyword search reliable without any embedding model. Unstructured natural language, by contrast, requires semantic search to surface conceptually related content that shares no exact keywords. Medin uses the example of finding all references to Star Wars spaceships — X-wings, TIE Fighters, the Millennium Falcon — in a document corpus, where keyword matching would miss most results.
To support the coding-tools case, Medin cites Claude Code maintainer Boris Churnney, who confirmed that early versions of Claude Code used a local vector database but switched to agentic search — grep, sed, cat, and file navigation tools run directly in the terminal — after finding it worked better for structured codebases. Cline co-creator Nick Pash made the same architectural decision for similar reasons. The video closes with a practical framework: evaluate your data’s structure before choosing a retrieval approach, rather than following trend-driven claims about what’s alive or dead.
📺 Source: Cole Medin · Published February 19, 2026
🏷️ Format: Deep Dive
