semantic-search（语义搜索）

Semantic Codebase Search

Find code by meaning, not just keywords. Always use this tool FIRST before falling back to grep/glob.

Projects are auto-discovered on first search — no manual projects add required.

Quick Search

PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-$(python -c "from pathlib import Path; import os; c=[Path('C:/.config/vercel-skills/skills/semantic-search'), Path.home()/'.claude/skills/semantic-search', Path.home()/'.agents/skills/semantic-search', Path('P:/aromatt/semantic-indexer/skills/semantic-search')]; print(next((str(p) for p in c if (p/'cocoindex/main.py').exists()), ''))")}"
python "$PLUGIN_ROOT/cocoindex/main.py" search "." "$ARGUMENTS"

Search Strategy

CRITICAL: Use semantic search FIRST for every code exploration task.

1. Run Semantic Search

# Search current project (auto-registers if not tracked)
python $PLUGIN_ROOT/cocoindex/main.py search "." "user authentication flow"

# Search specific project
python $PLUGIN_ROOT/cocoindex/main.py search "P:/myproject" "database connection pool"

# With limit and threshold
python $PLUGIN_ROOT/cocoindex/main.py search "." "query" 10 0.3

2. Read Results

Results include file paths, similarity scores, and daemon health status.

If daemon_status: "not_running" appears in output, start the daemon:

python $PLUGIN_ROOT/cocoindex/daemon.py start

3. Fall Back to Grep/Glob

Only use grep/glob if:

• Semantic search returns < 3 results
• You need exact string matching (regex)
• The query is a literal code pattern

Upstream/Downstream Impact Review

BEFORE making any edits, review impact:

Step 1: Find Callers (Upstream)

python $PLUGIN_ROOT/cocoindex/main.py search "." "calls functionName"
python $PLUGIN_ROOT/cocoindex/main.py search "." "imports from module-name"

Step 2: Find Dependencies (Downstream)

python $PLUGIN_ROOT/cocoindex/main.py search "." "dependencies of functionName"

Step 3: Document Impact

Before proceeding with edits, list:

1. Files that will be affected by the change
2. Functions/components that depend on the code
3. Potential breaking changes

Daemon Management

The indexer runs as persistent background daemons — one subprocess per registered project. Runtime state is shared across agent installs in ~/.semantic-indexer (override: SEMANTIC_INDEXER_HOME). Watcher fallback can be tuned with SEMANTIC_INDEXER_WATCH_FALLBACK_SECONDS (default: 60).

# Start daemons for ALL registered projects
python $PLUGIN_ROOT/cocoindex/daemon.py start

# Check per-project daemon status
python $PLUGIN_ROOT/cocoindex/daemon.py status

# View indexing logs (all projects, prefixed with [project-name])
python $PLUGIN_ROOT/cocoindex/daemon.py logs --tail 50

# Follow logs in real-time
python $PLUGIN_ROOT/cocoindex/daemon.py logs --follow

# Stop ALL daemons
python $PLUGIN_ROOT/cocoindex/daemon.py stop

Multi-Project Behavior

• start spawns one watcher subprocess per project, skipping already-running ones
• status shows per-project running/dead state with PIDs
• stop kills all daemon processes
• PIDs stored in ~/.semantic-indexer/daemon-pids.json
• Each subprocess logs with [project-name] prefix for easy filtering

Auto-Discovery

When you search a project that isn't registered yet, it's automatically added to the watch list. No manual projects add is needed for search to work (though the index must exist in pgvector).

If the daemon isn't running for a searched project, the search output includes:

• daemon_status: "not_running" — informational, search still works against existing index
• warning — human-readable message
• hint — command to start the daemon

Project Management

# Add project to watch list (for daemon)
python $PLUGIN_ROOT/cocoindex/main.py projects add "P:/myproject"

# List all projects (shows daemon status per project)
python $PLUGIN_ROOT/cocoindex/main.py projects list

# Remove project from watch list
python $PLUGIN_ROOT/cocoindex/main.py projects remove "P:/myproject"

Indexing

# Index a project once (no daemon)
python $PLUGIN_ROOT/cocoindex/main.py index "P:/myproject"

# Index with file watching (foreground)
python $PLUGIN_ROOT/cocoindex/main.py index "P:/myproject" --watch

Setup

Prerequisites

• Docker running
• Python 3.10+ with cocoindex, psycopg2
• Gemini API key

Initial Setup

# 1. Start pgvector container
cd $PLUGIN_ROOT/docker && docker-compose up -d

# 2. Install Python dependencies
pip install cocoindex psycopg2-binary

# 3. Set Gemini API key (shared state path, not plugin folder)
mkdir -p ~/.semantic-indexer
echo '{"gemini_api_key": "your-key"}' > ~/.semantic-indexer/credentials.json

# 4. Index your project
python $PLUGIN_ROOT/cocoindex/main.py index "."

Search Output Format

{
  "ok": true,
  "query": "user authentication",
  "project_root": "P:\\myproject",
  "result_count": 5,
  "daemon_status": "running",
  "results": [
    {
      "file_path": "src/auth/login.ts",
      "file_name": "login.ts",
      "content": "async function authenticateUser...",
      "location": {"start": 1234, "end": 2345},
      "similarity": 0.87,
      "preview": "async function authenticateUser(credentials: Credentials)..."
    }
  ]
}

When daemon is not running:

{
  "ok": true,
  "result_count": 0,
  "daemon_status": "not_running",
  "warning": "Daemon not running for project 'myproject'. Index may be stale.",
  "hint": "Start with: python daemon.py start"
}

Command Reference