Get the answer, or walk the graph.
Graphify turns your repo into a code knowledge graph the assistant traverses — query a node, follow a path, explain the route. LemonCrow indexes the same structure but searches it in one call, and it's a full runtime rather than a retrieval layer. Graphify is now in LemonCrow's retrieval benchmark too, so below is both: the numbers, and the architecture comparison that explains them.
| Tool | MRR | p95 | p100 |
|---|---|---|---|
| ★ LemonCrow +semantic (BGE) | 0.727 | 390ms | 1057ms |
| ★ LemonCrow lexical (default) | 0.676 | 134ms | 319ms |
| Graphify | 0.267 | 585ms | 3769ms |
Publishes LOCOMO/LongMemEval memory-QA accuracy, not code-search MRR -- first time measured on this benchmark: 0.267 MRR, 0.188 hit@1, vs. LemonCrow's 0.727 (semantic) / 0.676 (lexical). Full architecture comparison on its own page.
| LemonCrow | Graphify | |
|---|---|---|
| Core model | One call — code_search returns the symbol, its callers and callees, and ranked source. | A knowledge graph the assistant traverses — query, path, explain. |
| The agent loop | One shot — ranked results come back in a single call. | Multi-hop — the assistant walks typed edges to assemble it. |
| Scope | A full runtime: search, role-scoped agents, skills, hooks, verification, and cost tracking. | A retrieval layer: the graph plus its query tools. |
| Retrieval | Ranked by call-graph centrality; lexical by default, optional semantic. | Traversal along typed edges — calls, imports, defines, references. |
| Evidence | Published SWE-bench and 14-repo MRR runs, committed to the repo -- Graphify included. | Publishes LOCOMO/LongMemEval memory-QA numbers, not code-search MRR -- 0.267 MRR the first time it's run on this benchmark. |
| Freshness | Incremental re-parse of the file that changed. | Update a node; the surrounding edges stay intact. |
No comparison is one-sided. If you want a standalone, fully-open, human-browsable code graph with nothing sent anywhere, Graphify is a strong pick — and here's exactly where it's ahead:
| License | Apache-2.0 runtime, open-core (proprietary engine). | MIT — fully open. |
| Telemetry | Anonymous counts, on by default — one command turns them off. | None. |
| The artifact | A local SQLite index the agent searches. | A graph.html / graph.json you can open and browse yourself. |
The honest take
Same starting point — parse the repo, keep the structure. Graphify hands the assistant a graph to walk and a file you can open; LemonCrow precomputes call-graph centrality so the hop that matters is already ranked into the first result, then wraps it in agents, skills, and hooks. Different bets. Ours is one-shot resolve inside a full runtime, with the benchmarks committed.