Vitali Avagyan

Modern coding agents navigate large codebases by grepping, reading files, and following references. It works, but every structural question like "what calls this function?" becomes multiple rounds of search-and-read, burning tool calls, tokens, and latency, and the answer is only as complete as the search terms the agent happened to pick. Code is a graph, and these are graph questions. Code-Graph-RAG parses multi-language codebases with Tree-sitter into a unified knowledge graph in Memgraph, where calls, inheritance, imports, and definitions are already materialized. "What calls UserService.create_user?" becomes a single traversal returning an exact, complete answer, not one stitched together from text matches. The same structure powers AST-targeted surgical edits, and the toolset is exposed to Claude Code and other clients over MCP. This session covers the architecture: Tree-sitter parsing, the language-agnostic schema, and natural-language-to-Cypher querying. It makes the case for when precomputed structure beats search-and-read, and for when the two worlds combine: graph traversal and grep working together so agents reach for the right tool at the right time.