--- name: Code Search description: Semantic code search using local embeddings and sqlite-vec. Use for natural language queries against indexed codebases. Supports indexing, incremental updates, hybrid vector+keyword search, and background indexing for large repos. --- # Code Search Skill Semantic code search using local embeddings (sentence-transformers) and sqlite-vec. ## Setup Setup has already been completed. If commands fail with import errors or missing dependencies, re-run setup: ```bash # Check if setup is intact python3 ~/.claude/skills/code-search/scripts/setup.py --check # Re-run setup if needed python3 ~/.claude/skills/code-search/scripts/setup.py ``` Do NOT check setup before every command — assume it is ready unless an error indicates otherwise. ## Commands ### /index-codebase Index a codebase for semantic search. Runs incrementally - only re-indexes changed files. ```bash python3 ~/.claude/skills/code-search/scripts/index.py # Force re-index all files python3 ~/.claude/skills/code-search/scripts/index.py --force # Run in background for large codebases python3 ~/.claude/skills/code-search/scripts/index.py --background ``` ### /index-status Check the status of background indexing or list indexed codebases. ```bash # Check current indexing status python3 ~/.claude/skills/code-search/scripts/status.py # List all indexed codebases python3 ~/.claude/skills/code-search/scripts/status.py --list # JSON output python3 ~/.claude/skills/code-search/scripts/status.py --json ``` ### /search Search indexed codebases with natural language queries. ```bash python3 ~/.claude/skills/code-search/scripts/search.py "" --codebase # Limit results python3 ~/.claude/skills/code-search/scripts/search.py "" --codebase --limit 5 # Pretty output for humans python3 ~/.claude/skills/code-search/scripts/search.py "" --codebase --pretty ``` **Output format** (JSON): ```json [ { "file": "app/Services/PaymentService.php", "lines": "45-78", "content": "public function processPayment...", "score": 0.94, "language": "php" } ] ``` ### /reindex Force full re-index of a codebase (e.g., after model change): ```bash python3 ~/.claude/skills/code-search/scripts/index.py --force ``` ## How It Works 1. **Indexing**: Scans files, splits into ~50-line chunks with overlap, generates embeddings via sentence-transformers 2. **Change Tracking**: SHA256 manifest tracks file hashes - only changed files are re-indexed 3. **Hybrid Search**: Combines vector similarity (semantic) with FTS5 (keyword) using RRF fusion 4. **Storage**: Single SQLite database with sqlite-vec extension at `~/.claude/code-search-data/index.db` ## Data Locations - **Database**: `~/.claude/code-search-data/index.db` - **Manifests**: `~/.claude/code-search-data/manifests/` - **Status**: `~/.claude/code-search-data/status.json` - **Log**: `~/.claude/code-search-data/indexing.log` - **Model**: `~/.cache/huggingface/` (standard HuggingFace cache) ## Supported File Types Python, PHP, TypeScript, JavaScript, HTML, CSS, SCSS, JSON, Markdown, SQL, Shell, YAML, Go, Rust, Ruby, Java, Kotlin, Swift, C/C++, C#, and more. ## Performance Notes - **Initial index**: ~30-60 minutes for large codebases (runs in background) - **Incremental updates**: Seconds (only processes changed files) - **Search**: ~1 second per query (includes model load on first query) - **Embedding model** (default: all-MiniLM-L6-v2, ~80MB, fast) - **Alternative model**: `--model nomic` for nomic-embed-text-v1.5 (~275MB, slower but higher quality) - **RAM usage**: ~300MB-500MB (minilm) or ~500MB-1GB (nomic) - **Switching models requires**: `rm ~/.claude/code-search-data/index.db` then `--force` reindex

Code Search

Scan your own AI components