# Code knowledge graph

Band: emerging — Real practice, thin evidence. Mostly vendor-stated or reported rather than measured.
Group: Unsurveyed water

Indexing a repository as symbols and relationships — call edges, imports, inheritance, tests — so an agent traverses dependencies rather than matching strings. This half has the actual evidence: RepoGraph, CodexGraph, LocBench results, and a controlled three-arm study reporting a localisation gain within a harness and no regression against agentic grep. The finding that matters most for anyone building one is behavioural, not technical: in CodeCompass's trials, 58% of runs with graph access made zero tool calls, and agents needed explicit prompting to use it. Building the index is the easy half.

## See also

- [Agentic search vs. semantic index](https://clickai.dev/legend/agentic-search)
- [Graph engineering](https://clickai.dev/legend/graph-engineering)
- [Context pointer](https://clickai.dev/legend/context-pointer)

---

Full legend: https://clickai.dev/legend.md