Red Hat ships ripwire and says coding agents need a deterministic repo map, not a vector database

Red Hat's Emerging Technologies group published ripwire, an Apache-2.0 zero-dependency C++23 binary with vendored tree-sitter grammars for 21 languages, exposed as both a CLI and an MCP server (GitHub). It reached 969 stars on Hacker News on September 7. It returns ranked symbols, blast radius, tests-to-run and quality deltas at about 5% of the tokens a grep-and-read pass costs, and claims 0.31 seconds indexing against 34-plus seconds for graph-database approaches. No embeddings. No vector store. No LLM indexer. No daemon.

A vendor Red Hat's size publicly rejecting embeddings for code context is an architecture argument aimed at how most code-search products are built. The reasoning holds up: code has an AST, ASTs give you exact call graphs, and an exact call graph answers "what breaks if I change this" in a way cosine similarity structurally cannot. Embeddings shine when you don't know the vocabulary of what you're looking for. Inside a repo, you usually do.

There's corroborating evidence from a completely different direction. A study holding the environment fixed across 100 multi-file RefactorBench tasks and four model families found AST-aware chunking beat naive token-window chunking by 25-30% across prompt modes, and naive retrieval scored below the retrieval-free baseline (arXiv 2609.04898). Bad retrieval is worse than none. The same paper reports a lean retrieval-augmented single agent at 86%, beating the multi-agent sub-agent configuration, which is a direct argument against reflexively fanning out subagents on repo-scale refactors.

The tell that the category is shifting: ripwire and an independent MIT project called agentmap both benchmark against grep rather than against each other (agentmap). agentmap claims 100% precision on blast radius against grep's 60%, and 100% top-1 symbol location against grep's 32%, using ts-morph for TypeScript, JavaScript and Vue SFC. When every new entrant measures itself against the free Unix tool, the commercial incumbent has stopped being the reference point.

I've built RAG over code with pgvector and I'd take a deterministic symbol graph over it for anything where "what calls this" matters. Where embeddings still earn their keep is natural-language questions about intent, comments and docs. Two different jobs. The mistake was ever selling one product for both.