Nemesis

GraphRAG context engine for AI coding agents.

Nemesis parses your codebase into a semantic knowledge graph, creates vector embeddings for natural-language search, and exposes everything as MCP tools — giving AI agents deep, structured understanding of your code.

Your Code  ──>  Rust Parser  ──>  Code Graph (Kuzu)  ──>  MCP Tools  ──>  AI Agent
                                  Vector DB (LanceDB)
                                  Memory (Rules/Decisions)

Features

• Code Graph — builds a rich knowledge graph with classes, functions, imports, inheritance, and call relationships. Backed by Kuzu (embedded) or Neo4j (remote).
• Semantic Search — vector similarity search over code chunks using OpenAI embeddings (text-embedding-3-small) or local models via sentence-transformers. Powered by LanceDB.
• AST-Aware Chunking — Rust-based parser (Tree-sitter) splits code at natural boundaries, not arbitrary line counts.
• Delta Updates — tracks file hashes and only re-indexes what changed. Auto-compacts the vector store after each update.
• Memory System — stores coding rules, architectural decisions, and conventions in the graph for persistent context across sessions.
• File Watcher — watches your project for changes and auto-reindexes with configurable debouncing.
• Multi-Project — index and query multiple projects independently with isolated namespaces.
• 15+ Languages — Python, TypeScript, JavaScript, Rust, Go, Java, C/C++, C#, Ruby, PHP, Swift, Kotlin, Scala, Zig.

Quick Start

Installation

pip install nemesis-ai

With optional backends:

pip install nemesis-ai[neo4j]             # Neo4j graph backend
pip install nemesis-ai[local-embeddings]  # Local embeddings (no API key needed)

Configuration

Set your OpenAI API key (or use local provider):

echo "NEMESIS_OPENAI_API_KEY=sk-..." > ~/.nemesis/.env

All settings via environment variables with NEMESIS_ prefix:

Variable	Default	Description
NEMESIS_DATA_DIR	~/.nemesis	Data storage directory
NEMESIS_GRAPH_BACKEND	kuzu	Graph backend (kuzu or neo4j)
NEMESIS_VECTOR_PROVIDER	openai	Embedding provider (openai or local)
NEMESIS_VECTOR_MODEL	text-embedding-3-small	OpenAI embedding model
NEMESIS_LANGUAGES	["python"]	Languages to index

CLI

nemesis index ./my-project              # Index a project
nemesis query "authentication logic"    # Semantic search
nemesis watch ./my-project              # Watch & auto-reindex
nemesis projects                        # List indexed projects
nemesis serve                           # Start MCP server

MCP Server Setup

Add to your Claude Code config (~/.claude.json):

{
  "mcpServers": {
    "nemesis": {
      "command": "/path/to/nemesis",
      "args": ["serve"],
      "env": {
        "NEMESIS_OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

MCP Tools

Tool	Description
search_code	Natural-language semantic search over your codebase
get_context	Get full graph context for a file (classes, methods, imports, relations)
index_project	Index a project directory
update_project	Delta update — only re-indexes changed files
remember_rule	Store a coding rule for persistent context
remember_decision	Record an architectural decision
get_memory	Retrieve all stored rules, decisions, and conventions
get_session_summary	Get current session queries and results
list_projects	List all indexed projects
remove_project	Remove a project and its indexed data

Architecture

nemesis/
├── core/           # Engine, CLI, MCP server, config, file watcher
├── graph/          # Graph DB abstraction (Kuzu + Neo4j adapters)
├── vector/         # LanceDB vector store + embedding providers
├── indexer/        # Parsing pipeline, delta detection, AST chunking
├── parser/         # Rust parser bridge (PyO3 + Tree-sitter)
├── memory/         # Rules, decisions, conventions, session context
└── tools/          # MCP tool implementations

Graph Schema: 15 node types, 16 edge types covering code structure (File, Class, Function, Method, Variable, Import), relationships (CONTAINS, INHERITS, CALLS, IMPORTS), and knowledge (Rule, Decision, Convention).

Data Storage (~/.nemesis/):

• graph/ — Kuzu embedded database
• vectors/ — LanceDB vector store
• registry.json — Project metadata

Tech Stack

Component	Technology
Code Parsing	Rust + Tree-sitter (via PyO3)
Graph DB	Kuzu (embedded) / Neo4j (remote)
Vector DB	LanceDB
Embeddings	OpenAI text-embedding-3-small / sentence-transformers
Search Metric	Cosine similarity
Protocol	MCP (Model Context Protocol)
Config	Pydantic Settings
CLI	Click

Development

git clone https://github.com/CytrexSGR/Nemesis.git
cd Nemesis
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest

487 tests, Python 3.11+.

License

MIT