I initially built cs (codespelunker) as a way to answer the question, can BM25 relevance search work without building an index?

Turns out it can, and so I iterated on the idea, building it into a full CLI tool. Recently I wanted to improve it by adding relevance of tools like Sourcegraph or Zoekt but again without adding an index.

cs uses scc https://github.com/boyter/scc to understand the structure of the file on the fly. As such it can filter searches to code, comments or strings. It also applies a weighted BM25 algorithm where matches in actual code rank higher than matches in comments (by default).

I also added a complexity gravity weight using the cyclomatic complexity output from scc as it scans. So if you're searching for a function, the implementation should rank higher than the interface.

    cs "authenticate" --gravity=brain           # Find the complex implementation, not the interface
    cs "FIXME OR TODO OR HACK" --only-comments  # Search only in comments, not code or strings
    cs "error" --only-strings                   # Find where error messages are defined
    cs "handleRequest" --only-usages            # Find every call site, skip the definition

v3.0.0 adds a new ranker, along with a interactive TUI, HTTP mode, and MCP support for use with LLMs (Claude Code/Cursor).

Since it's doing analysis and complexity math on the fly, it's slower than any grep. However, on an M1 Mac, it can scan and rank the entire 40M+ line Linux kernel in ~6 seconds.

Live demo (running over its own source code in HTTP mode): https://codespelunker.boyter.org/ GitHub: https://github.com/boyter/cs