Code Semantic Search

Overview

Semantic code search using Phase 1 vector embeddings and Phase 2 hybrid search (semantic + structural). Find code by meaning, not just keywords.

Core principle: Search code by what it does, not what it's called.

Phase 2: Hybrid Search

This skill now supports three search modes:

1. Hybrid (Default):

   • Combines semantic + structural search
   • Best accuracy (95%+)
   • Slightly slower but still <150ms
   • Recommended for all searches

2. Semantic-Only:

   • Uses only Phase 1 semantic vectors
   • Fastest (<50ms)
   • Good for conceptual searches
   • Use when structure doesn't matter

3. Structural-Only:

   • Uses only ast-grep patterns
   • Precise for exact matches
   • Best for finding function/class definitions
   • Use when you need exact structural patterns

Performance Comparison

Mode	Speed	Accuracy	Best For
Hybrid	<150ms	95%	General search
Semantic-only	<50ms	85%	Concepts
Structural-only	<50ms	100%	Exact patterns
Phase 1 only	<50ms	80%	Legacy (fallback)

When to Use

Always:

• Finding authentication logic without knowing function names
• Searching for error handling patterns
• Locating database queries
• Finding similar code to a concept
• Discovering implementation patterns

Don't Use:

• Exact text matching (use Grep instead)
• File name searches (use Glob instead)
• Simple keyword searches (use ripgrep instead)

Usage Examples

Hybrid Search (Recommended)

// Basic hybrid search
Skill({ skill: 'code-semantic-search', args: 'find authentication logic' });

// With options
Skill({
  skill: 'code-semantic-search',
  args: 'database queries',
  options: {
    mode: 'hybrid',
    language: 'javascript',
    limit: 10,
  },
});

Semantic-Only Search

// Fast conceptual search
Skill({
  skill: 'code-semantic-search',
  args: 'find authentication',
  options: { mode: 'semantic-only' },
});

Structural-Only Search

// Exact pattern matching
Skill({
  skill: 'code-semantic-search',
  args: 'find function authenticate',
  options: { mode: 'structural-only' },
});

Implementation Reference

Hybrid Search: .claude/lib/code-indexing/hybrid-search.cjs

Query Analysis: .claude/lib/code-indexing/query-analyzer.cjs

Result Ranking: .claude/lib/code-indexing/result-ranker.cjs

Integration Points

• developer: Code exploration, implementation discovery
• architect: System understanding, pattern analysis
• code-reviewer: Finding similar patterns, consistency checks
• reverse-engineer: Understanding unfamiliar codebases
• researcher: Research existing implementations

Iron Laws

1. ALWAYS use hybrid mode (semantic + structural) for general searches — semantic-only misses exact matches; structural-only misses conceptual variants; hybrid provides 95% accuracy vs 85% for single mode.
2. ALWAYS use ripgrep/keyword search first for fast keyword discovery — semantic search is for meaning-based queries; exact strings, function names, and filenames are found faster with ripgrep.
3. NEVER use semantic search without a meaningful natural-language query — single-word or code-syntax queries produce poor semantic results; describe what the code does, not what it's called.
4. ALWAYS combine with code-structural-search for precision refinement — start broad with semantic discovery, then use ast-grep patterns to find exact structural matches from the semantic results.
5. NEVER ignore low-similarity results without checking them — similarity scores are approximations; a 0.7 score result may be more relevant than a 0.9 score result for uncommon patterns.

Anti-Patterns

Anti-Pattern	Why It Fails	Correct Approach
Semantic search for exact string matching	Slower and less accurate than text search	Use ripgrep/Grep for exact keyword matching
Single-word queries ("auth")	Too vague for semantic matching; returns noise	Use natural-language descriptions ("authentication token validation logic")
Using semantic-only mode for general searches	Misses structural variants; 85% vs 95% accuracy	Use hybrid mode (default) for general queries
Ignoring search results that don't match expectations	Semantic results find surprising-but-relevant code	Read all results; unexpected matches are often the most valuable
Not combining with structural search	Finds concepts but not exact patterns	Use semantic for discovery → structural for precision

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.