Raw capture — raw/knowledge-graph-making-agent-dumber.md
url: https://praveenvijayan.substack.com/p/your-knowledge-graph-is-making-your title: Your Knowledge Graph Is Making Your Agent Dumber author: Praveen Vijayan date_fetched: 2026-07-25 date_published: 2026-07-24 site: Substack (praveenvijayan.substack.com)
Praveen Vijayan evaluates Graphify, a tool that converts a code folder into a knowledge graph for AI agent use. He tested it on a live TypeScript monorepo (Bun/Turbo, 605 files, ~445k words, 4 apps + 4 packages) with three real questions, comparing it against ripgrep + manual file reads.
The built graph had 1,842 nodes and 3,722 links across 103 communities. Node types: 1,520 code nodes (AST-extracted), 164 concept nodes (LLM-extracted), 110 document nodes, 48 rationale nodes. ~76% of all edges came from contains, imports, and imports_from — all AST-derived. Semantically interesting relations: 86 edges total, "under 3% of the graph."
Hyperedges over plan directories — The tool clustered ~110 markdown plan files into named workstreams (e.g., "Notifications subsystem," "Onboarding Persistence Bug Chain"). The author calls this "the standout feature," noting it "correctly ties a plan that is still open to the four completed plans it depends on."
Community labels as onboarding — 103 communities with human-readable labels that cross directory boundaries.
Import cycle detection — Returned "None detected" cleanly; "genuinely useful."
Low query-time cost — ~0.2 seconds wall time per query with no API charges.
Commit tracking — built_at_commit recorded, making staleness detectable.
Query returns nodes, not answers — A question about OAuth provider registration returned 170 nodes. "Roughly 8 of 170 nodes are on-topic." Top results were barrel files and unrelated modules. By contrast, rg returned 35 lines in 0.01 seconds with the architecture legible in the output.
Hub poisoning in barrel-export monorepos — The top three nodes by degree (176, 170, 160) were barrel files. BFS depth-2 from createApp() reaches 275 nodes (15% of the graph); from notifications.ts it reaches 465 nodes (25%). The author labels this "a graph-topology problem."
Stale graphs report dead symbols as EXTRACTED — A function renamed 40 seconds before the next commit still appeared with highest-confidence label. "Wrong data wearing a confidence badge is more dangerous than no data."
Degree centrality mislabels utility functions — cn() (a four-line Tailwind helper) ranked as a "core abstraction" due to high fan-in. Six of the top ten most-connected nodes were leaf utilities.
Community cohesion was extremely weak — All scores ranged 0.03–0.08, meaning "the partition is barely better than arbitrary." Yet LLM-generated labels sounded authoritative. "The label quality is decoupled from the cluster quality."
Inverted token economics for agents — Graphify returned ~2000 tokens with ~5% on-target; rg returned ~300 tokens at ~100% on-target. The tool "reconstructs, imperfectly, what good repos already write down" — the repo already had hand-written AGENTS.md files.
Build cost — 1.19M input tokens across two builds. Real expense on first build, though caching helped second run by 6×.
From §2.2: If a repo has barrel files, a shared types package, a DI container, a God app-factory, or a widely-imported utils.ts, import-graph traversal will produce this failure pattern. Genuinely modular codebases with narrow interfaces would fare far better.
GRAPH_REPORT.md end-to-end as an onboarding artifactrg + readingexplain on code symbols unless drift is zerobuilt_at_commit vs. HEADStrong fit: Repos with large prose corpora, undocumented legacy codebases, one-time onboarding, genuinely modular codebases, mixed-media corpora.
Poor fit: Barrel-export TypeScript monorepos, repos with strong docs culture, fast-moving codebases without automated update hooks, precise symbol lookups (use rg/LSP instead).
Closing thesis: "Graphify is an excellent document knowledge graph and a mediocre code knowledge graph."
All measurements from one private TypeScript monorepo over one session. Degree distributions and staleness blast radius computed from graph.json directly. graphify update and watch commands not exercised. Topology argument in §2.2 should generalize; precision numbers should not be assumed to transfer.