Graphify Turns Codebases and Documents Into Queryable Knowledge Graphs

Graphify is expanding an open-source workflow that maps codebases, documentation and other files into a queryable knowledge graph for AI coding assistants. The project produces an interactive HTML graph, a machine-readable graph.json file and a plain-language GRAPH_REPORT.md, giving agents a structured alternative to repeatedly searching raw files.\n\nThe tool is available as a skill for Claude Code, Cursor, Codex, Gemini CLI and several other coding environments. Its code-focused pipeline uses tree-sitter for local, deterministic parsing, resolving relationships such as calls, imports and inheritance across roughly 40 programming languages. The project says the code pass requires no language-model calls and sends nothing off the machine; documents and media can use a configured model for semantic extraction. Graphify’s repository also describes support for PDFs, images, video, audio, SQL schemas and configuration files.\n\nGraphify’s central distinction is that it is not a conventional vector index. It builds a traversable graph in which concepts, files and relationships become nodes and edges. Users can query a concept, trace a path between two entities or ask the system to explain a relationship. Community detection identifies subsystems, while “god nodes” highlight highly connected concepts. Each inferred relationship is labeled as extracted, inferred or ambiguous, providing an audit trail for what the system found directly and what it resolved.\n\nThe repository reports 0 LLM credits for graph construction in its benchmark table, alongside a 45.3% question-answering accuracy score on LOCOMO and 76% on LongMemEval-S. On LOCOMO recall@10, Graphify reports 0.497, compared with 0.048 for mem0 and 0.149 for supermemory; on QA accuracy, supermemory is listed at 49.7% against Graphify’s 45.3%. The project notes that the tests used the same harness, model and budgets, with blind validation by a second judge reaching 90.6% agreement. These are project-reported results rather than an independent industry benchmark.\n\nThe workflow is aimed at persistent context for coding agents. After installation, a user can run /graphify . to generate the graph artifacts, then install platform-specific guidance so an assistant prefers scoped graph queries before broad file searches. Optional commands can update only changed files, export an Obsidian vault or wiki, generate GraphML, connect to Neo4j or FalkorDB, and expose an MCP server.\n\nDevelopers describe the approach as useful for navigating large repositories, understanding cross-file dependencies and combining code knowledge with an Obsidian-style personal knowledge base. A separate setup guide says the initial run can take roughly two to five minutes depending on corpus size, with AST extraction remaining local while semantic extraction consumes model tokens for documents and images. A Graphify and Obsidian guide documents that split and warns that the semantic pass is where token costs arise.\n\nThe limitations are material. Non-code inputs require a model backend, so privacy, latency and cost depend on configuration. The project’s own documentation warns that large graphs can become difficult to open as HTML beyond 5,000 nodes, and it includes troubleshooting for incomplete extraction, duplicate nodes, oversized model responses and stale graphs after repository changes. Some developers favor a simpler Markdown-based memory system with hooks and direct file reading, arguing that a graph can add complexity without reliably improving answers. Others report that the Graphify-and-Obsidian combination makes agent context easier to browse and maintain.\n\nGraphify’s significance is therefore less about replacing retrieval systems outright than about making relationships, provenance and architecture visible to both humans and agents. Its next test will be whether the graph remains accurate and maintainable as repositories change, while its reported gains hold across independent datasets and real production workloads.