kgrag

name

kgrag

description

Expert knowledge for KGRAG — the unified cross-KG registry and federated query layer for CodeKG, MemoryKG, and MetaKG. Use this skill when: (1) Setting up or configuring KGRAG in projects, (2) Querying across multiple knowledge graphs simultaneously, (3) Extracting code/doc snippets for LLM context, (4) Integrating with Claude Code or Claude Desktop, (5) Running architectural analyses, (6) Managing the KG registry, or (7) Troubleshooting KG-related issues.

KGRAG Skill

KGRAG is the unified abstraction over CodeKG, MemoryKG, and MetaKG. It provides a single registry and federated query interface to search across multiple knowledge graphs simultaneously.

What is KGRAG?

KGRAG (Knowledge Graph Registry and Query) orchestrates multiple KG backends:

• CodeKG — Semantic + structural analysis of Python code
• MemoryKG — Semantic + structural analysis of markdown documentation
• MetaKG — Domain-specific knowledge graphs
• KGRAG — The unifying registry, CLI, and federated query layer

One registry manages all KG instances. One CLI queries all of them. One philosophy: "One registry. Many KGs. Infinite queries."

Core Concepts

Registry

The KGRAG registry (~/.kgrag/registry.sqlite by default) is the source of truth:

• Tracks all KG instances across your projects
• Stores metadata: names, kinds, paths, build status
• Enables federated queries and analysis

KG Layers

A KG layer is a single knowledge graph for a repository:

• Code layer (CodeKG) — Python codebase semantic indexing
• Doc layer (MemoryKG) — Markdown documentation semantic indexing
• Multiple layers can coexist in the registry

One project can have both code and doc layers registered simultaneously.

Federated Queries

Query all KGs in one command:

kgrag query "database connection"
# Results ranked globally, sourced from all registered KGs

Quick Start

Initialize a project

cd ~/repos/myproject
kgrag init
# Auto-detects applicable layers, builds them, registers in registry

Query all KGs

kgrag query "error handling patterns"
kgrag query "REST API" --kind code -k 5

Extract snippets for LLM

kgrag pack "authentication flow" --out context.md

Visualize interactively

kgrag viz
# Launches Streamlit UI: Registry, Federated Query, Analysis, Snippets

Check registry health

kgrag status
kgrag list

CLI Commands

See cli-reference.md for detailed documentation on all commands:

Command	Purpose
kgrag init	Initialize and register KG layers for a project
kgrag query	Federated semantic search across all KGs
kgrag pack	Extract source snippets for LLM context
kgrag analyze	Show cross-KG statistics and health metrics
kgrag viz	Launch Streamlit visualizer (interactive UI)
kgrag list	Show all registered KG instances
kgrag status	Quick registry health check
kgrag info	Detailed metadata for a specific KG
kgrag register	Manual KG registration (rarely needed)
kgrag unregister	Remove KG from registry
kgrag scan	Auto-discover existing KG databases
kgrag mcp	Launch MCP server for Claude Code/Desktop integration

Quick reference: kgrag <command> --help for any command.

Workflows

See workflows.md for detailed multi-step workflows:

• First-time setup — Initialize KGRAG for one or more projects
• Preparing LLM context — Extract snippets for code review or AI analysis
• Cross-project analysis — Query patterns across multiple repositories
• Team collaboration — Share visualizer with non-technical stakeholders
• Claude Code integration — Set up MCP for interactive exploration

Setup MCP for Claude Code (5 min)

1. Create .mcp.json in your project root:

{
  "mcpServers": {
    "kgrag": {
      "command": "kgrag",
      "args": ["mcp", "--registry", "/path/to/registry.sqlite"]
    }
  }
}

2. Restart Claude Code — MCP tools appear automatically

3. Now available in prompts:

   • kgrag_query(q, k, kinds) — Semantic search
   • kgrag_pack(q, k, context, kinds) — Snippet extraction
   • kgrag_list() — List registered KGs
   • kgrag_info(name) — KG metadata
   • kgrag_stats() — Registry health

Environment Variables

Variable	Purpose	Default
KGRAG_REGISTRY	Path to registry SQLite file	~/.kgrag/registry.sqlite
CODEKG_MODEL_DIR	Cache embedding model (CodeKG)	.codekg/models
DOCKG_MODEL_DIR	Cache embedding model (MemoryKG)	.memorykg/models

Set these once in your shell profile for permanent configuration.

Troubleshooting

See troubleshooting.md for detailed solutions.

Common issues:

• Query returns no results — Check if KG is built: kgrag status
• KG marked "not built" — Rebuild: kgrag init <repo-path> --wipe
• MCP tools not appearing — Verify .mcp.json paths are absolute; restart Claude Code
• Stale results in visualizer — Click "🔄 Refresh Registry" or rebuild

Best Practices

Registry management:

• Keep a single registry (~/.kgrag/registry.sqlite) for your workspace
• Use meaningful names: project-name-code, project-name-doc
• Run kgrag status monthly to catch unbuilt KGs

Querying:

• Use specific queries ("JWT validation" > "authentication")
• Filter by kind: --kind code or --kind doc
• Adjust -k: Start with 5 for focused results, 12 for broad surveys

Rebuilding:

• After major refactoring: kgrag init --wipe (full rebuild)
• After renames: kgrag init (incremental, no --wipe)
• Monthly health check: Re-run kgrag init to catch stale data

Philosophy

One registry. Many KGs. Infinite queries.

KGRAG unifies CodeKG, MemoryKG, and MetaKG under a single contract:

• Apps register what they know (a KG instance)
• Users query what they want (federated search)
• Agents orchestrate across domains (MCP tools)

This enables:

• Cross-domain searches — Find code and matching documentation in one query
• Unified CLI — Same commands work across all KG types
• Extensibility — New KG types integrate seamlessly
• Composability — Agents layer KGRAG with other tools

Learn More

• CodeKG skill — Semantic + structural indexing of Python code
• MemoryKG skill — Semantic + structural indexing of markdown docs
• CLI reference — See cli-reference.md
• Workflows — See workflows.md
• Troubleshooting — See troubleshooting.md