GraphRelay is a graph-aware code search tool that helps coding agents understand unfamiliar codebases. By connecting code search with function relationships and dependencies, it makes it easier to find implementations, understand how components interact, and assess the impact of code changes—all from a single local tool.
See our solution in action.
The problem we set out to solve
As codebases grow, finding the right code becomes increasingly time-consuming. Coding agents struggle to see how different components connect, making code exploration harder and increasing the risk of missed dependencies and unexpected bugs.
From problem to possibility
Meet GraphRelay - a smarter way to navigate complex codebases. It helps coding agents find relevant code and uncover hidden connections, reducing manual searching and enabling faster, more confident code changes.
Key Feature & Modules
Search code finds code from a natural-language description or finds implementations of a named interface. Lexical BM25 plus graph expansion surfaces related callers and callees, so the agent needs fewer exploratory searches.
Inspect returns a symbol's definition, callers and callees, or the contents and summary of a file or module. It can return full source in place of a separate file read.
Trace impact shows what depends on a symbol before it is changed, reducing the risk of breaking callers in other files.
A file watcher keeps the index fresh as files change, and watch status reports indexing progress and health, so results stay trustworthy while editing.
Introduction
Result
Query → FTS5 BM25 lexical search → 1-hop graph expansion → personalized PageRank → enrich → rerank (graph + keyword) → ranked results
Query latency on Keycloak (older suite, 2026-09-29): p50 189 ms, p95 281 ms. The 100 ms p95 target was not met.
Agent benchmark (30 ERPNext questions, 1 repeat) coverage vs no-tool baseline — GLM 5.3 Flash 80.2% vs 77.6% (+2.6); Sonnet 5.5 70.6% vs 67.2% (+3.4). Neither difference is statistically distinguishable from zero.
Parallel full-index parsing measured about 6.1x faster than sequential on a synthetic corpus. A real large-repo indexing benchmark is still open.
