RAG Isn't Dead — It's Growing Up

Every few months someone declares RAG dead because context windows got longer. Yet every production enterprise AI system we build still has retrieval at its core. Here's why — and how RAG has matured.

Why long context didn't kill retrieval

Stuffing a million tokens into context is possible, but it's slow, expensive and degrades reasoning quality on the details that matter. Retrieval remains the mechanism for freshness (your data changed an hour ago), permissions (this user can't see that document) and cost (why pay for a million tokens when 4,000 well-chosen ones perform better?).

What modern RAG looks like

• Hybrid retrieval: dense vectors plus keyword search plus metadata filters, fused with reranking
• Chunking driven by document structure, not fixed token counts
• Per-user permission filtering at retrieval time, not response time
• Citations rendered in the UI so every claim is verifiable

Evaluation is the real differentiator

The teams getting value from RAG maintain golden datasets of question–answer pairs with known-correct sources, and run them on every pipeline change. Retrieval precision, answer faithfulness and citation accuracy are tracked like uptime. Without this, RAG quality silently decays as content grows.

Where it's heading

Agentic retrieval — where the model decides what to search, reformulates queries and iterates — is replacing single-shot retrieval for complex questions. But the fundamentals stay: grounded answers, verifiable sources, controlled costs. RAG didn't die; it became infrastructure.