bm25

Ranked content search over any text corpus. One CLI, in-memory BM25 index per process, with a session-local disk cache so repeat invocations against the same corpus load in tens of milliseconds instead of rebuilding.

Setup

uv pip install --system --break-system-packages bm25s

Install is sub-second on a warm uv cache. That's the entire dependency.

Usage

BM25=/mnt/skills/user/bm25/scripts/bm25.py

# Local directory
python3 $BM25 ./repo 'csrf middleware'

# Multiple queries against the same in-memory index (build once, query many)
python3 $BM25 ./repo 'csrf middleware' 'session backend' 'queryset filter'

# Cloned GitHub repo via tarball (one HTTP call)
python3 $BM25 'github.com/django/django' 'atomic transaction'
python3 $BM25 'github.com/django/django@stable/5.0.x' 'atomic transaction'

# Project knowledge or uploads
python3 $BM25 project 'RAG scaling laws'
python3 $BM25 uploads 'tax loss harvesting'

# Filters
python3 $BM25 ./repo 'auth flow' --exclude 'tests/*' --exclude '*/tests/*'
python3 $BM25 ./repo 'config' --include '*.py' --include '*.toml'

# Interactive (REPL — single corpus, many queries)
python3 $BM25 ./repo --interactive

# JSON output for piping
python3 $BM25 ./repo 'auth flow' --json

Corpus types

Options

With no --include, a default set of text/code extensions is indexed (Python, JS/TS, Go, Rust, Markdown, JSON, YAML, etc.). Standard noise dirs are skipped unconditionally: .git, node_modules, __pycache__, .venv, dist, etc.

When to use bm25

The boundary with searching-codebases: that skill is code-specific (routes between regex and TF-IDF, expands via tree-sitting AST). bm25 is the simpler general-purpose tool — any corpus, no AST awareness, no routing. Prefer searching-codebases for code; reach for bm25 when the corpus is mixed (docs + code), non-code (notes, transcripts, PDFs converted to text), or when you specifically want BM25's length-normalized scoring.

Design notes

• Session-local disk cache at /home/claude/.bm25-cache/<key>/. The

key is a hash of (resolved_corpus_path, include_globs, exclude_globs, max_file_bytes) — any change invalidates naturally. First invocation builds and saves; subsequent invocations against the same corpus and filters load in tens of milliseconds. The cache lives in /home/claude, which is ephemeral, so it expires at the session boundary — same lifetime as the corpus state itself, no cross-session staleness. ~5–35MB per cached index, depending on corpus size.

• `--no-cache` bypasses both load and save — useful only if you've

mutated the corpus mid-session (rare) or want to confirm a rebuild matches.

• Reuse within a single invocation. The retriever stays in memory

between queries in one process. Passing multiple queries positionally, or using --interactive, amortizes any rebuild cost across queries.

• No AST awareness. Chunking is per-file. For symbol-level results in

code, combine with tree-sitting queries on the same paths.

• Tokenizer. Default bm25s.tokenize with stopwords disabled — over a

small Django sample, AST-derived token streams (identifiers/strings/ comments only) gave near-identical rankings, so we don't bother.

Output format

Default (human-readable):

QUERY: csrf middleware
----------------------------------------------------------------------
  1.   5.51  django/core/checks/security/csrf.py
    def _csrf_middleware():
        return "django.middleware.csrf.CsrfViewMiddleware" in settings.MIDDLEWARE
  2.   5.34  docs/howto/csrf.txt
    ...

--json produces {"query": ..., "results": [{"path", "score", "snippet"}, ...]}.

Architecture

bm25.py CLI
  ├── resolve_corpus(spec)         → local Path (downloads tarball if github.com/...)
  ├── cache_key(...)               → 16-hex sha256 of inputs that determine the index
  ├── CorpusIndex.load(cache_dir)  → returns cached index if present, else None
  ├── CorpusIndex.build(...)       → walks files, tokenizes, indexes with bm25s
  ├── CorpusIndex.save(cache_dir)  → persists to /home/claude/.bm25-cache/<key>/
  ├── query(q, k)                  → ranked (doc_idx, score) pairs
  └── best_snippet(doc, q, lines)  → pick line w/ most query-term hits + context

Cache contents per directory:

• bm25/ — bm25s.BM25.save() output (NumPy arrays + vocab)
• corpus.pkl — pickled {paths, docs} so we can render snippets without

re-reading the source files

• manifest.json — corpus root, files count, built_at timestamp

No network beyond optional tarball fetch on github.com/... corpora. No state outside /home/claude/, which is ephemeral.