Minima

What is Minima about?

Minima enables you to index local documents and query them through LLMs that run either locally (via Ollama) or as remote services (ChatGPT, Claude). It packages the entire RAG pipeline – document ingestion, embedding, vector storage, reranking, and LLM inference – inside Docker Compose setups, keeping data on your own machine or private cloud.

How to use Minima?

1. Clone the repository and copy env.sample to .env.
2. Fill required variables (LOCAL_FILES_PATH, EMBEDDING_MODEL_ID, EMBEDDING_SIZE, OLLAMA_MODEL, RERANKER_MODEL, optionally USER_ID and PASSWORD for ChatGPT).
3. Choose a Docker Compose file:
   • docker-compose-ollama.yml for fully local deployment.
   • docker-compose-chatgpt.yml for ChatGPT‑enabled custom GPTs.
   • docker-compose-mcp.yml for MCP integration with Anthropic Claude.
4. Run the stack: docker compose -f <file> --env-file .env up --build.
5. For ChatGPT, copy the OTP printed in the terminal and use the provided Minima GPT link.
6. For Claude, add the MCP entry to claude_desktop_config.json as shown in the README.
7. To launch the Electron UI, navigate to the electron folder, run npm install and npm start.

Key Features

• Three operation modes: fully isolated on‑premises, ChatGPT‑backed custom GPT, Claude‑backed MCP.
• Containerized architecture with Docker Compose for easy deployment.
• Support for multiple models: Sentence‑Transformer embeddings, BAAI rerankers, any Ollama LLM.
• Flexible indexing of PDFs, DOCX, TXT, MD, CSV, XLS recursively.
• Vector store powered by Qdrant, sized according to the embedding dimension.
• Electron UI for local interaction and a web UI at http://localhost:3000.
• MCP compatibility allowing integration with Claude Desktop or VS Code Copilot.

Use Cases

• Enterprises needing secure, on‑premises document search without sending data to external APIs.
• Developers building custom GPTs that answer questions from proprietary knowledge bases.
• Researchers who want a plug‑and‑play local RAG stack for literature review.
• Teams using Anthropic Claude who want to augment the model with their own files via MCP.

FAQ

Q: Do I need an internet connection? A: Only if you use remote LLM services (ChatGPT or Claude). The isolated mode runs entirely offline.

Q: Which embedding models are supported? A: Any Sentence‑Transformer model; the README uses sentence-transformers/all-mpnet-base-v2 as an example.

Q: How do I choose an Ollama model? A: Use any LLM listed on the Ollama model catalog (e.g., qwen2:0.5b). Ensure it is an LLM, not an embedding model.

Q: Can I run Minima on Windows? A: Yes, as long as Docker Desktop and the required Python >= 3.10 with uv are installed.

Q: How is authentication handled for custom GPTs? A: Provide USER_ID (your email) and PASSWORD in .env; these create a temporary Firebase account for ChatGPT integration.

Minima is an open source RAG on-premises containers, with ability to integrate with ChatGPT and MCP. Minima can also be used as a fully local RAG.

Minima currently supports three modes:

1. Isolated installation – Operate fully on-premises with containers, free from external dependencies such as ChatGPT or Claude. All neural networks (LLM, reranker, embedding) run on your cloud or PC, ensuring your data remains secure.

2. Custom GPT – Query your local documents using ChatGPT app or web with custom GPTs. The indexer running on your cloud or local PC, while the primary LLM remains ChatGPT.

3. Anthropic Claude – Use Anthropic Claude app to query your local documents. The indexer operates on your local PC, while Anthropic Claude serves as the primary LLM.

---

Running as Containers

1. Create a .env file in the project’s root directory (where you’ll find env.sample). Place .env in the same folder and copy all environment variables from env.sample to .env.

2. Ensure your .env file includes the following variables:

3. For fully local installation use: docker compose -f docker-compose-ollama.yml --env-file .env up --build.

4. For ChatGPT enabled installation use: docker compose -f docker-compose-chatgpt.yml --env-file .env up --build.

5. For MCP integration (Anthropic Desktop app usage): docker compose -f docker-compose-mcp.yml --env-file .env up --build.

6. In case of ChatGPT enabled installation copy OTP from terminal where you launched docker and use Minima GPT

7. If you use Anthropic Claude, just add folliwing to /Library/Application\ Support/Claude/claude_desktop_config.json

{
    "mcpServers": {
      "minima": {
        "command": "uv",
        "args": [
          "--directory",
          "/path_to_cloned_minima_project/mcp-server",
          "run",
          "minima"
        ]
      }
    }
  }

8. To use fully local installation go to cd electron, then run npm install and npm start which will launch Minima electron app.

9. Ask anything, and you'll get answers based on local files in {LOCAL_FILES_PATH} folder.

---

Variables Explained

LOCAL_FILES_PATH: Specify the root folder for indexing (on your cloud or local pc). Indexing is a recursive process, meaning all documents within subfolders of this root folder will also be indexed. Supported file types: .pdf, .xls, .docx, .txt, .md, .csv.

EMBEDDING_MODEL_ID: Specify the embedding model to use. Currently, only Sentence Transformer models are supported. Testing has been done with sentence-transformers/all-mpnet-base-v2, but other Sentence Transformer models can be used.

EMBEDDING_SIZE: Define the embedding dimension provided by the model, which is needed to configure Qdrant vector storage. Ensure this value matches the actual embedding size of the specified EMBEDDING_MODEL_ID.

OLLAMA_MODEL: Set up the Ollama model, use an ID available on the Ollama site. Please, use LLM model here, not an embedding.

RERANKER_MODEL: Specify the reranker model. Currently, we have tested with BAAI rerankers. You can explore all available rerankers using this link.

USER_ID: Just use your email here, this is needed to authenticate custom GPT to search in your data.

PASSWORD: Put any password here, this is used to create a firebase account for the email specified above.

---

Examples

Example of .env file for on-premises/local usage:

LOCAL_FILES_PATH=/Users/davidmayboroda/Downloads/PDFs/
EMBEDDING_MODEL_ID=sentence-transformers/all-mpnet-base-v2
EMBEDDING_SIZE=768
OLLAMA_MODEL=qwen2:0.5b # must be LLM model id from Ollama models page
RERANKER_MODEL=BAAI/bge-reranker-base # please, choose any BAAI reranker model

To use a chat ui, please navigate to http://localhost:3000

Example of .env file for Claude app:

LOCAL_FILES_PATH=/Users/davidmayboroda/Downloads/PDFs/
EMBEDDING_MODEL_ID=sentence-transformers/all-mpnet-base-v2
EMBEDDING_SIZE=768

For the Claude app, please apply the changes to the claude_desktop_config.json file as outlined above.

To use MCP with GitHub Copilot:

1. Create a .env file in the project’s root directory (where you’ll find env.sample). Place .env in the same folder and copy all environment variables from env.sample to .env.

2. Ensure your .env file includes the following variables:

   • LOCAL_FILES_PATH
   • EMBEDDING_MODEL_ID
   • EMBEDDING_SIZE

3. Create or update the .vscode/mcp.json with the following configuration:

{
  "servers": {
    "minima": {
      "type": "stdio",
      "command": "path_to_cloned_minima_project/run_in_copilot.sh",
      "args": [
        "path_to_cloned_minima_project"
      ]
    }
  }
}

Example of .env file for ChatGPT custom GPT usage:

LOCAL_FILES_PATH=/Users/davidmayboroda/Downloads/PDFs/
EMBEDDING_MODEL_ID=sentence-transformers/all-mpnet-base-v2
EMBEDDING_SIZE=768
USER_ID=user@gmail.com # your real email
PASSWORD=password # you can create here password that you want

Also, you can run minima using run.sh.

---

Installing via Smithery (MCP usage)

To install Minima for Claude Desktop automatically via Smithery:

npx -y @smithery/cli install minima --client claude

For MCP usage, please be sure that your local machines python is >=3.10 and 'uv' installed.

Minima (https://github.com/dmayboroda/minima) is licensed under the Mozilla Public License v2.0 (MPLv2).