The RAG Search skill enables Claude to perform advanced semantic and hybrid searches across processed document databases, including research papers, technical documentation, and source code. By leveraging vector similarity, Hypothetical Document Embeddings (HyDE), and reranking optimizations, it ensures high-precision retrieval of relevant information from large-scale knowledge bases to enhance AI-driven insights and context-aware responses.

When should I use rag-search?

rag-search is useful in the following scenarios: • Academic Research: Conduct semantic queries over vast collections of research papers and documents to find specific findings or methodologies. • Codebase Navigation: Search through source code chunks to locate specific implementation details, middleware, or logic using hybrid (vector + keyword) search. • Technical Documentation Retrieval: Quickly find relevant sections in complex manuals or markdown files to answer specific 'how-to' questions. • Visual Information Discovery: Search and retrieve specific figures or image-based content from processed documents using specialized image chunk tables. • High-Precision Knowledge Retrieval: Use advanced optimizations like Reranking and HyDE to improve the accuracy of answers for complex knowledge-based questions.

name	rag-search
description	Search RAG database for relevant content. Use for semantic queries over processed documents, code, or papers.

RAG Search

This skill helps you search processed document databases using semantic similarity and retrieval-time optimizations.

Quick Search

# Basic vector search
uv run processor search ./lancedb "how does the caching work"

# Hybrid search (vector + keyword)
uv run processor search ./lancedb "ConfigParser yaml loading" --hybrid

# Search code
uv run processor search ./lancedb "authentication middleware" --table code_chunks

Available Tables

Table	Content
`text_chunks`	Documents, papers, markdown (default)
`code_chunks`	Source code
`image_chunks`	Figures from papers
`chunks`	Unified table (if created with --table-mode unified)

MCP Server

Start the RAG MCP server for programmatic access:

uv run rag-mcp

Generate a config template:

uv run rag-mcp --config_generate

Configure in Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "rag": {
      "command": "uv",
      "args": ["run", "rag-mcp"],
      "cwd": "/path/to/processor"
    }
  }
}

Available MCP Tools

search - Vector/hybrid search with optimizations
search_images - Search image chunks
list_tables - List available tables
generate_config - Create config template

Retrieval Optimizations

Enable these for better results at the cost of latency:

Optimization	Flag	Latency	Best For
Hybrid Search	`hybrid=True`	+10-30ms	Keyword-heavy queries
HyDE	`use_hyde=True`	+200-500ms	Knowledge questions
Reranking	`rerank=True`	+50-200ms	High precision needs
Parent Expansion	`expand_parents=True`	+5-20ms	Broader context

Recommended Combinations

Fast search (default): No optimizations - pure vector similarity

Better recall:

search(query="...", hybrid=True)

Knowledge questions:

search(query="what is...", use_hyde=True, rerank=True)

Code search:

search(query="...", table="code_chunks", hybrid=True)

Maximum precision:

search(query="...", hybrid=True, use_hyde=True, rerank=True)

Configuration

Edit rag_config.yaml to set defaults:

# Embedding profiles (must match processor config)
text_profile: "low"
code_profile: "low"
ollama_host: "http://localhost:11434"

# Default search behavior
default_limit: 5
default_hybrid: false

# HyDE settings (uses Claude SDK by default, falls back to Ollama)
hyde:
  enabled: false
  backend: "claude_sdk"  # claude_sdk (default) or ollama
  claude_model: "haiku"  # haiku, sonnet, opus
  ollama_model: "llama3.2:latest"  # fallback

# Reranking settings
reranker:
  enabled: false
  model: "BAAI/bge-reranker-v2-m3"
  top_k: 20
  top_n: 5

Understanding Results

Results include:

content: Matched chunk text
source_file: Original file path
score: Similarity (0-1, higher is better)
metadata: Additional fields (section, language, etc.)

Troubleshooting

No results found

Check database exists: uv run processor stats ./lancedb
Verify table has data: --table text_chunks
Try broader query terms

Poor quality results

Enable hybrid search: --hybrid
Check embedding profiles match processor config
Consider reranking: rerank=True

Slow searches

Disable HyDE if not needed
Reduce rerank_top_k
Check Ollama server performance

rag-search

When & Why to Use This Skill

Use Cases