searching-codebases
Find code in any codebase by regex pattern or natural language concept. Auto-routes between n-gram indexed regex search (2-20x faster than ripgrep) and TF-IDF semantic search. Expands results to full functions via tree-sitting AST data. For Python sources, a binding-resolved reference/definition tier (pyright, via python-lsp) finds real callers and definitions without same-name false positives.
Features
- Dual search modes — regex (pattern/identifier) and semantic (natural language concepts), auto-routed per query
- N-gram indexed regex — sparse inverted index narrows candidate files by 90-99% before ripgrep verification
- TF-IDF semantic search — cosine similarity ranking over code chunks (functions, classes)
- Binding-resolved Python tier —
--refs/--def/--hovervia pyright: true find-all-callers and go-to-definition that exclude same-named, unrelated symbols and follow imports. Engaged lazily; degrades to the regex text path when pyright/node is unavailable - AST context expansion —
--expandreturns complete function/class bodies instead of line fragments - Flexible sources — accepts GitHub URLs, local directories, uploaded files/archives, or project knowledge
- Mixed queries — multiple queries with different modes in a single invocation; indexes built once per mode
Dependencies
- ripgrep — required for regex verification
- tree-sitting — auto-installs the bare
tree-sitterpackage when needed: for--expandcontext and for the symbol→position resolution that seeds the binding-resolved tier (grammars ship bundled). Regex and semantic search work without it - scikit-learn — required for semantic mode (auto-installs)
- python-lsp — provides the binding-resolved tier (
--refs/--def/--hover); self-bootstraps pyright on first use and needs systemnode(v18+). Without it those flags degrade to the regex text path