RAG

Compare the Market's engineering team published on 6 March 2026 an evaluation of four context-retrieval approaches for AI code review, run across 79 real merge requests and measured with MLflow on Databricks. RAG, which chunks code, embeds it, and retrieves by semantic similarity, scored worse than the no-context baseline on almost every metric, at three times the cost. GKG, the GitLab Knowledge Graph built on Tree-sitter ASTs, beat it by 21 percent on inline comment coverage, 0.696 against 0.577. The team's explanation is structural: reviewing code requires knowing who calls a function, and vector similarity returns code that merely looks alike. Noise, false positives, blindness to inter-file relations, and distraction of the model followed.

An October 2025 article makes a different case. RAG existed to work around small context windows; those went from 8K tokens toward millions, and five costs now buy nothing: chunking loses meaning, embeddings miss semantic richness, hybrid search adds complexity, reranking adds latency, and the infrastructure needs maintenance. Claude Code is cited as the alternative shape: direct, context-rich search with no retrieval machinery. The verdict there is that RAG was a temporary detour.

Leonie Monigatti traces a lineage. Vanilla RAG asked how to retrieve, Agentic RAG asks whether to retrieve, and Agent Memory adds a write path. Ashish Singh sets RAG against an Enterprise Knowledge Fabric of knowledge graphs and hybrid search. What replaces it is still contested.