Adaptive KG-RAG

Local-first knowledge graph repo memory for coding agents.

Adaptive KG-RAG indexes a workspace into a typed code knowledge graph, then lets coding agents query compact repo context through MCP instead of reading large parts of the repository into the prompt.

What It Does

• Indexes files, symbols, imports, calls, tests, docs, and repo-memory facts across Python, JavaScript/TypeScript, Java, and C/C++ (functions, structs/unions/enums, typedefs, macros, #include edges).
• Shows graph status, node/edge counts, storage footprint, and export tokens in a VS Code sidebar.
• Provides a graph explorer for inspecting files, symbols, and relationships.
• Configures MCP for Codex, Claude Code, Cursor, Antigravity, and Kimi.
• Lets agents query adaptive_kg_status, adaptive_kg_query, and related MCP tools before falling back to direct source reads. Graph retrieval is hybrid (vector + lexical scoring, then graph expansion) and falls back to flat vector chunk RAG when the graph is sparse.
• Uses offline Hash/KG mode by default (private, no API). Optional local semantic embeddings (sentence-transformers) or API providers are selectable via the adaptiveKg.embeddingProvider setting.

Important Beta Note

This extension is currently a lightweight VS Code wrapper around the local Adaptive KG Python package. For best results, install the Python package from the project repository first:

git clone https://github.com/YijunSun/KG-RAG-CodeKG
cd KG-RAG-CodeKG
python -m pip install -e .[dev]

If you are developing from source, set adaptiveKg.sourceRoot to the repository root when VS Code cannot discover src/adaptive_kg automatically.

Quick Use

1. Open a repository in VS Code.
2. Open the Adaptive KG activity bar icon.
3. Run Index Workspace or Rebuild Graph.
4. Run Configure Agent MCP and choose Codex, Claude Code, Cursor, Antigravity, or Kimi.
5. Restart the selected agent or open a fresh agent session.

MCP Targets

Use Adaptive KG: Uninstall Agent MCP to remove the managed MCP entry and managed instruction block from one target or all configured targets.

Commands

• Adaptive KG: Index Workspace
• Adaptive KG: Rebuild Workspace Graph
• Adaptive KG: Query Workspace Graph
• Adaptive KG: Configure Agent MCP
• Adaptive KG: Uninstall Agent MCP
• Adaptive KG: Show Graph Status
• Adaptive KG: Open Graph Explorer
• Adaptive KG: Prepare LLM Graph Audit
• Adaptive KG: Open Agent Protocol
• Adaptive KG: Review Pending Updates
• Adaptive KG: Start Local API Server
• Adaptive KG: Evaluate Local Model

Embeddings

Set adaptiveKg.embeddingProvider:

• hashing (default) — offline deterministic, zero dependencies, fully private.
• sentence-transformers — local semantic embeddings (MiniLM); free and private, needs the embed extra installed, and falls back to hashing if the model isn't available.
• openai / gemini / openrouter / custom — API embeddings. These send code snippets to a third party and are subject to its rate limits.

For code retrieval the graph's edge structure (who-calls-whom, imports, includes) does most of the work, so the local providers are recommended; API embeddings mainly help fuzzy natural-language queries.

Privacy

Default indexing and querying are fully local — hashing and sentence-transformers never send your code off the machine. Only the API embedding providers transmit code, and only when you explicitly select one. MCP mode does not ask Adaptive KG for your model API key; the host agent keeps using its own account and model session.

Project

Source, documentation, benchmarks, and release notes:

https://github.com/YijunSun/KG-RAG-CodeKG