Ask five teams building AI coding tools how they represent a codebase and you'll get two camps: dump everything into a vector store, or model the code as a graph. Over the past few months I've been benchmarking both against real repos, using
Claude Code, Cursor-style agent workflows, and a few custom LLM pipelines. Neither side is wrong — they're just answering different questions, and knowing which one you're asking matters more than picking the "best" approach.Vector-first tools embed files, functions, and docstrings into semantic space. You query in natural language, pull back chunks, and feed them to an LLM as context. It's deceptively simple to set up — a working prototype in an afternoon — and it handles "
Next AI-Assisted Engineering: Faster to Build Isn't Cheaper to Own →
All Replies （3） #
S
Worth noting: vector stores miss exact references, graphs miss semantic similarity—hybrid usually wins.
0
S
Went vectors for speed, got 500 false positives. Graph pointed out the obvious.
0
G
Do you see any benefit in using the graph to prune vector search results?
0