semantic-search

name	semantic-search
description	Primary code search — hybrid semantic + keyword + BM25 matching. Automatically intercepted as default search when Beacon index is healthy.
allowed-tools	["Bash","Read"]

Hybrid Code Search (Beacon)

This repo has a Beacon hybrid search index combining semantic embeddings, BM25 keyword matching, and identifier boosting. Beacon is enforced as the default search — grep is automatically intercepted and redirected to Beacon for queries it handles better.

How to search

node ${CLAUDE_PLUGIN_ROOT}/scripts/search.js "<query>"

Options

• --top-k N — number of results (default: 10)
• --threshold F — minimum score cutoff (default: 0.35)
• --no-hybrid — disable hybrid, use pure vector search only

Multi-query batch

node ${CLAUDE_PLUGIN_ROOT}/scripts/search.js "auth flow" "session handling" "token refresh"

Single HTTP round-trip for all queries. Returns grouped results.

Output

JSON array of matches, each with:

• file — file path
• lines — line range (e.g. "45-78")
• similarity — vector cosine similarity
• score — final hybrid score (when hybrid enabled)
• preview — first 300 chars of matched chunk

Grep intercept behavior

Grep is denied and redirected to Beacon unless one of these conditions is met:

Grep passes through when...	Example
Pattern has regex metacharacters	function\s+\w+
Targets a specific file	path: "src/lib/db.js"
output_mode is "count"	Counting occurrences
Pattern is <= 3 characters	fs, db
Dotted identifier	fs.readFileSync, path.join
Path-like pattern (contains / or \)	src/components
output_mode is "content"	Viewing matching lines
Quoted string literal	"use strict", 'Content-Type'
Annotation/marker pattern	TODO, FIXME, @param, #pragma
URL-like pattern	http://, localhost:3000
Beacon index is unhealthy	DB missing, empty, dimension mismatch
Intercept disabled via config	intercept.enabled: false

To disable interception entirely, set intercept.enabled: false in .claude/beacon.json.

Workflow

1. Search with Beacon → get candidate files + line ranges with scores
2. Read top 2-3 files at the indicated line ranges for full context
3. If needed, grep within those files for specifics (imports, call sites)
4. Answer the user with file:line citations