How dmayboroda/minima works
Project Type and Purpose
• Minima is a modular, on-premises Retrieval-Augmented Generation (RAG) platform delivered as a set of Docker containers. It supports fully local LLMs (via Ollama), ChatGPT integration (Custom GPT), or Anthropic Claude (MCP).
• Core use case: index local documents, embed them, re-rank, and answer user queries via web UI (React), desktop UI (Electron), or MCP (Claude app/GitHub Copilot).
File-Structure Analysis
• Top-level containers orchestrated by three Docker Compose files:
• docker-compose-ollama.yml
• docker-compose-chatgpt.yml
• docker-compose-mcp.yml
• Components / microservices:
• chat/ → React web chat UI
• electron/ → Electron desktop UI
• indexer/ → Document ingestion, embedding, vector storage
• linker/ → (If used) connects question → answer sockets
• llm/ → Core RAG: async queues, reranker, LLM chain
• mcp-server/→ MCP (Claude) interface for Copilot/Claude Desktop
• Shared config: .env (LOCAL_FILES_PATH, model IDs, credentials)
• Orchestration helper scripts: run.sh, run_in_copilot.sh
README Insights
• Describes three deployment modes and environment variables.
• Emphasizes on-prem vs. ChatGPT vs. Claude.
• Mentions Firebase-based authentication for Custom GPT.
• Electron app as fully-local desktop client.
• MCP integration via uv/MCP server.
Detail
• Container as a component.
• Shared .env injecting configuration into all services.
• Optional paths (ChatGPT branch vs. Ollama branch vs. MCP branch)
• Port numbers where relevant (3000, service ports).
• Document async queue messaging and the vector store as first-class entities.