Containerized RAG that lets ChatGPT rummage through your local files

What it does

Minima is a Dockerized retrieval-augmented generation toolkit that indexes local documents—PDFs, spreadsheets, Word files, text, markdown, and CSVs—into a Qdrant vector store. It then exposes that indexed data to a front-end chat interface or to external LLMs like ChatGPT, Claude, or any OpenAI-compatible API endpoint. The indexer stays on your machine, so your documents never need to live in a cloud vector database.

The interesting bit

Rather than building yet another web UI, Minima treats popular chat apps as interchangeable front-ends. It supports the Model Context Protocol for Claude Desktop and GitHub Copilot, offers a Custom GPT integration for ChatGPT, and still works fully offline via Ollama. That flexibility is unusual for a project this size: it acknowledges that developers already live in these tools and just want their local files to show up there.

Key highlights

• Runs entirely on-premises with Ollama, or pairs a local indexer with remote LLMs through an OpenAI-compatible API.
• Integrates with ChatGPT Custom GPTs, Anthropic Claude Desktop via MCP, and GitHub Copilot.
• Uses Sentence Transformers for embeddings and Qdrant for vector storage; Ollama mode adds a BAAI reranker step, while custom-LLM mode skips reranking to reduce overhead.
• Ships with an Electron app for a local chat UI.
• Supports recursive indexing of .pdf, .xls, .docx, .txt, .md, and .csv files.

Caveats

• Embedding model support is currently limited to Sentence Transformers; other families are not supported.
• The ChatGPT integration requires creating a Firebase account with an email and password, which feels slightly at odds with an otherwise privacy-focused tool.
• MCP setup requires Python ≥3.10 and the uv toolchain installed on the host, not just inside containers.

Verdict

Developers who want their existing ChatGPT or Claude workflows to reach into local document folders without setting up a cloud vector database will find Minima a pragmatic fit. If you need broad embedding model support or a fully unified architecture without mode-specific Docker Compose files, it may feel more like glueware than a platform.