CCG-RAG: Semantic Codebase Search

Intelligent code and documentation retrieval using embeddings and knowledge graphs.

When to Use

Activate this skill when:

• Need to understand unfamiliar codebase
• Searching for code by functionality (not just text)
• Finding related code patterns
• Building context for complex tasks

Core Capabilities

1. Code Search

rag_query           - Semantic search across codebase
rag_related_code    - Find similar code patterns
rag_build_index     - Index/reindex codebase

2. Document Search

documents_search    - Search documentation
documents_find_by_type - Find docs by type (api, guide, spec)
documents_list      - List all indexed documents

Search Modes

Semantic Search

Find code by describing what it does:

"authentication middleware"
"error handling functions"
"database connection setup"

Pattern Search

Find similar implementations:

"find functions similar to validateUser"
"show me other API endpoints"
"related test patterns"

Documentation Search

"API documentation for auth"
"setup guide for database"
"architecture decisions"

How It Works

Code Chunking

• Functions and classes extracted as units
• Tree-sitter parsing for accurate boundaries
• Preserves context (imports, comments)

Embeddings

• Code-specific embeddings (UniXcoder/Qwen3)
• Natural language descriptions for each chunk
• Hybrid search: semantic + keyword

Knowledge Graph

• Function call relationships
• Import/export dependencies
• Type hierarchies

Example Usage

Find Authentication Code

User: "Find all code related to user authentication"

rag_query({ query: "user authentication login session" })

Results:
1. src/auth/login.ts:authenticate() - Main login handler
2. src/middleware/auth.ts:verifyToken() - JWT verification
3. src/services/session.ts:createSession() - Session management

Find Similar Patterns

User: "Show me code similar to the error handler in api.ts"

rag_related_code({ file: "src/api.ts", function: "handleError" })

Results:
1. src/services/db.ts:handleDbError() - 85% similar
2. src/utils/errors.ts:formatError() - 72% similar

Search Documentation

User: "Find API documentation for payments"

documents_search({ query: "payment API integration" })

Results:
1. docs/api/payments.md - Payment API Reference
2. docs/guides/stripe-integration.md - Stripe Setup Guide

Two-Stage Retrieval

For accurate results, CCG-RAG uses:

1. Vector Search - Fast semantic matching
2. LLM Rerank - Intelligent relevance scoring

Query → Embed → Top 20 candidates → LLM rerank → Top 5 results

Index Management

Build Index

rag_build_index({
  paths: ["src/", "lib/"],
  exclude: ["node_modules", "dist"],
  languages: ["typescript", "javascript"]
})

Index Status

rag_status()

{
  "indexed_files": 245,
  "chunks": 1847,
  "last_updated": "2025-12-04T08:00:00Z",
  "embedding_model": "qwen3-embedding-8b"
}

Best Practices

1. Natural language queries - Describe what you're looking for
2. Combine with memory - Store important findings
3. Use for context - Build understanding before changes
4. Keep index fresh - Rebuild after major changes

Integration with Latent Mode

RAG enhances Latent Chain Mode:

🔍 [analysis]
1. rag_query({ query: "current auth implementation" })
2. Identify hot spots from RAG results
3. Build comprehensive codeMap

📋 [plan]
1. rag_related_code({ function: "targetFunction" })
2. Find similar patterns to follow
3. Plan patches based on existing conventions