Graft: Fixing Code Search by Running PageRank Over Your Call Graph

Grounded / Real Inflated / Uruttu

50% real

50% uruttu

article Original Content

Ask a coding agent where something lives and it keyword-greps, so a function sharing one word with your query outranks the one you meant. Graft seeds the lexical matches, then runs personalized PageRank over the call graph the way search ranks the web: a match wired into the code your query touches rises, an isolated namesake sinks. Deterministic, no embeddings. 154 stars, MIT.

verified Validated Content

Fact-Check

Naming collision issue: "Graft" turns out to be an extremely overloaded project name on GitHub — there are at least half a dozen unrelated tools called "Graft" (a Go dependency-injection library, an in-memory graph DB, a personal knowledge graph app, a semantic memory cache for AI agents, a neuroscience image-analysis tool, and at least two different codebase-context tools for AI coding agents). This makes it hard to be certain which one the post refers to.

Closest match found: A project (amaar-mc/graft) does match much of the description — it's a local-first codebase context engine for AI coding tools via MCP, it builds a dependency graph from the codebase using tree-sitter parsing, and it runs personalized PageRank to score files/functions by structural importance, and it is MIT-licensed. This lines up well with the "seeds lexical matches, then runs personalized PageRank over the call graph" concept.

What I could not verify:

The "154 stars" figure — GitHub star counts aren't something I can reliably confirm via search at a point-in-time snapshot, and they change constantly. I can't confirm this number is currently accurate for any specific "Graft" repo.
The exact framing ("a function sharing one word with your query outranks the one you meant," "match wired into the code your query touches rises, an isolated namesake sinks") — this is a paraphrase/marketing framing that doesn't appear verbatim in what I found, though it's consistent with how personalized PageRank conceptually works (a well-established, real algorithm used in web search and now applied to code graphs).
"No embeddings" — plausible and consistent with the deterministic, graph-based (not vector-similarity-based) approach described in the matching repo, but not something I could independently confirm as a specific claim.

Verdict: The underlying concept is real and technically sound (personalized PageRank applied to code call graphs to fix keyword-search ranking problems is a legitimate, sensible approach, and a matching open-source tool does appear to exist). However, I can't confirm this post refers unambiguously to one specific repo, and I can't verify the exact star count or that every phrase is a faithful, unembellished description of that project's actual behavior.