How Graphify works.
Shared codebase context for your agents and your team: a code graph parsed from source, with the Wiki, Memory and PR reviews built on top of it.
One graph, four ways in.
The graph sits underneath. Your agents and your team read it through the Wiki, Memory and reviews.
Code graph
Functions, files and the calls and imports between them, parsed from your source, with a tag on every edge saying how it was found.
Wiki
Docs generated from the graph for every indexed repository. Each page keeps a confidence trail, so you can see what was read from the code and what was inferred.
Memory
Decisions, incidents and lessons your agents record, kept per code entity rather than per conversation, so an agent editing a function recalls what is known about it.
PR reviews
Graphify reviews a pull request when it is opened, reopened or marked ready, and uses the graph to show how the change affects the code around it.
Two stages.
One is deterministic. The other says so when it is not.
Code is parsed
Tree-sitter grammars walk every source file, SQL schemas and Terraform included. The open-source Graphify CLI supports about 40 languages, and hosted Graphify lists 63 grammars. Functions, imports and calls become edges because a parser found them.
Everything else is read
Docs, PDFs and images go through an optional pass with the model backend you configure, which marks a link it is unsure of AMBIGUOUS.
Three files, all yours.
With the open-source Graphify CLI, in graphify-out in your repo: one to open, one to read, one for a machine.
graphify-out/graph.html an interactive graph
graphify-out/GRAPH_REPORT.md god nodes, communities, surprises
graphify-out/graph.json the graph itselfAn edge is a claim.
Every claim carries a tag saying how it was arrived at, which a similarity score cannot offer.
- EXTRACTED
- Straight from the syntax tree. Parse it again and the parser finds it again.
- INFERRED
- Worked out by Graphify, with or without a model, and given a confidence score.
- AMBIGUOUS
- Flagged by the model pass as uncertain. The parser leaves out a call it cannot pin to one target.
What text search misses.
Embeddings and grep are built to find text, not connections, and connections are what the hard questions about code are made of. What breaks if I change this is a question about the caller of a caller of a caller. Chunking cuts the links before retrieval begins; grep answers only the hop you can phrase as a string.
The retrieval literature has reached the same diagnosis. GraphRAG (Edge et al., 2024) puts a knowledge graph on top of chunk retrieval to recover the structure chunking threw away. Graphify starts from the graph, so an answer arrives as a path through your real code.
The honest limit: for fuzzy search over large volumes of prose, embeddings are still the better tool. Graphify vs RAG covers when each one fits, and what is Graphify gives the short version.
Three verbs.
Each one answers with a path you can open.
$ graphify query AuthService # everything joined to a symbol
$ graphify path AuthService Db # the route between two of them
$ graphify explain <edge> # why this edge exists at all