Compare the Market's AI code reviewer hit 70% accurate inline comments using GitLab Orbit — against 58% with RAG, which performed worse than no context at all. Ryan Harvey, Head of AI Engineering at the U.K. price comparison platform, tested four context approaches across 79 real merge requests. RAG underperformed every other method, including feeding the agent zero extra context. The message is blunt: dumping file chunks into a vector store is the wrong way to ground an agent in a large codebase.
What RAG Gets Wrong About Code Context
Agents are good at writing code. They're terrible at navigating the system around it: the pipelines, deployments, work items, ownership, and cross-repo dependencies. In a monorepo, that gap wastes token budgets and produces code that looks correct but gets reverted. Across repos, the context window fills before the agent finds the answer, and the task fails. GitLab Orbit, now in public beta, sidesteps the whole approach. It builds a live, queryable graph of every merge request, pipeline, vulnerability, deployment, and ownership relationship — kept current via change-data-capture into ClickHouse. No RAG pipeline, no chunking, no embedding.
Orbit's Architecture: Change-Data-Capture Into a 2-Billion-Edge Graph
Orbit ingests SDLC data via change-data-capture into ClickHouse, parses code in 12 languages (Ruby, Java, Kotlin, Python, TypeScript, JavaScript, Rust, Go, C#, C, C++, PHP) through GitLab's Rails internal API, and serves the graph over a Cypher-like DSL, MCP, REST, and the GitLab CLI. At GitLab's own scale, the indexer covers over 40,000 projects, 500 million nodes, and 2 billion edges in under 45 minutes. Every change is picked up by an event-driven engine, so the graph stays current. Query traffic never hits your GitLab instance — indexing runs as a separate service. Permissions mirror GitLab's so agents see exactly what the user can see. The query engine compiles every request through validation, planning, optimization, and security passes before hitting the database. No separate infrastructure to stand up; Orbit builds on data GitLab already captures.
External Agents Hit 11x Speedup on Monorepo Tasks
Connect Claude Code to Orbit through Model Context Protocol (MCP) and it stops crawling files to figure out where things live. One or two graph queries replace thousands of file-read iterations. On the same tasks with the same model, agents run up to 11x faster, consume up to 4.5x fewer tokens, and generate up to 45x fewer hallucinations. The Data Explorer gives engineers the same graph without an agent — useful for tracing incidents, mapping vulnerability blast radius, or scoping migrations against current dependencies. One Cypher query returns every in-flight MR that will hit the same failing job across every project in your group. Security ships a remediation plan in the hour a CVE lands, not the week after.
Orbit makes the graph the source of truth for agent reasoning. That approach just made RAG look like the wrong abstraction for code context — and the numbers from Compare the Market prove it's not just theory.
Source: Introducing GitLab Orbit: Full code and lifecycle context, in one query
Domain: about.gitlab.com
Comments load interactively on the live page.