How arm/metis works
Project type and purpose
• Metis is an open-source, AI-powered CLI tool for deep security code review.
• It’s designed as a plugin-based, extensible system that indexes source code into a vector database, uses LLMs for semantic analysis, and emits structured security findings (SARIF).
File structure
• Top-level dirs:
• src/metis: core Python package
• src/metis/cli: command-line interface entry points and commands
• src/metis/engine.py: orchestration of indexing and review workflows
• src/metis/plugins: language plugins (C, Python, Rust) + plugin metadata
• src/metis/providers: LLM providers (OpenAI, Azure)
• src/metis/vector_store: vector database backends (Chroma, pgvector)
• src/metis/sarif: SARIF result writer
• config files: metis.yaml (defaults), plugins.yaml (prompt & chunking)
• docker/docker-compose: PostgreSQL + pgvector setup
• examples, docs, tests illustrate usage and deployments
README for additional insights
• Core flow: CLI → index → review_code / review_file / review_patch → SARIF output
• RAG pattern: split code chunks → embed → store/retrieve similar contexts → prompt LLM
• Two vector backends supported (Chroma for local, Postgres+pgvector for scale)
• Single LLM provider now, pluggable to add more
Components & interactions
a. Main components
• User → CLI layer
• CLI commands → Engine
• Engine
• Loader & chunker (reads code, splits into chunks)
• Embedding client (calls LLM embed API)
• Vector store adapter (Chroma or Postgres)
• Review coordinator (retrieval, prompting via LLM provider)
• SARIF writer
• Plugins
• Language plugin logic (tokenization, prompts)
• Providers
• OpenAI/Azure client
• Data stores
• Local ChromaDB
• Remote PostgreSQL + pgvector
• External services
• OpenAI/Azure LLM endpoint
b. Relationships & data flow
1. User invokes CLI command
2. CLI calls Engine with config
3. Engine loads source files, applies language plugins
4. Engine calls embedding provider → embeddings
5. Embeddings stored in vector store
6. During review, Engine queries vector store for similar chunks
7. Engine assembles context + prompts via provider
8. LLM returns analysis → Engine formats results
9. SARIF writer outputs structured findings
10. CLI renders or writes SARIF JSON
c. Architectural patterns & principles
• Plugin-based architecture for languages & providers
• Adapter pattern around vector stores (common interface)
• Separation of concerns: CLI, engine orchestration, I/O, plugins, providers
• RAG (Retrieval-Augmented Generation) flow
• Configuration-driven via YAML (runtime parameters, prompts, chunking)
d. Key technologies
• Python 3.12+, Click or argparse for CLI
• ChromaDB (local) or PostgreSQL + pgvector (scale)
• OpenAI/Azure openai-compatible APIs
• SARIF format for output
• Docker Compose for self-hosted Postgres