Most code search looks fine on a small repository. Twenty thousand lines, a few hundred files, one naming convention — almost anything works. Grep works. Embeddings work.
That is not where enterprise software lives, and it is not where search has to hold up.
So we ran the benchmark at the sizes that actually matter: a 200K-line codebase and a 2M-line codebase, comparing Vyazen Continuum against Claude Code's search agent on the same questions, at both scales.
The results
| Vyazen Continuum | Claude Code Search Agent | |
|---|---|---|
| Adherence — 200K | 98 | 85 |
| Adherence — 2M | 95 | 72 |
| Accuracy — 200K | 92 | 82 |
| Accuracy — 2M | 87 | 62 |
Vyazen Continuum leads on both axes at both sizes. On the 200K codebase, by 13 points on adherence and 10 on accuracy. On the 2M codebase, by 23 and 25.
What the numbers say
The margins are the headline, but the interesting part is what happens between the two columns.
Scale it ten times, and Vyazen Continuum gives up 3 points of adherence and 5 points of accuracy. It answers a 2M-line codebase about as well as it answers a 200K-line one.
Claude Code gives up 13 points of adherence and 20 points of accuracy. At 2M lines its accuracy falls to 62 — meaning roughly two in five answers are no longer right.
So the gap is not a fixed advantage. It roughly doubles on adherence and more than doubles on accuracy as the codebase grows. Whatever separates these two systems is not visible at small scale, and becomes the dominant factor at large scale.
Why the curves diverge
A search agent that works from the filesystem rebuilds its understanding of the codebase from scratch on every question. It greps, it reads, it infers. That is a reasonable strategy on a small repository, where the search space is small enough that guessing well is enough.
It does not survive growth. At two million lines, the same query returns far more near-identical matches, so the agent reads more files to tell them apart, and the useful fraction of what it has gathered keeps falling. The work grows with the codebase, every single time you ask.
Vyazen Continuum does that work once. Relationships across the codebase are resolved when the code is indexed, not rediscovered per question — so answering does not get harder as the repository gets bigger. That is the whole reason the curve stays flat, and it is why the two lines pull apart exactly where it costs the most.
What this means in practice
Benchmarks on small repositories will tell you that every tool is roughly as good as every other tool. They are measuring the case where the problem is easy.
The number worth asking any vendor for is the one at your size. If a search tool's quality is quoted on a repository an order of magnitude smaller than yours, you have not been told how it will behave on your codebase — you have been told how it behaves on someone else's.
At two million lines, that difference is 25 points of accuracy.