multilingual-rag-mcp

Multilingual RAG MCP server for local document search. Query in one language, find content in another.

Built for the common case where you talk to AI agents in English but your documents are in Russian (or any other language). The multilingual embedding model maps semantically similar concepts across 50+ languages to the same vector space — no translation step needed.

Install

pip from PyPI

pip install multilingual-rag-mcp

One-liner with uvx (no install needed)

uvx --from multilingual-rag-mcp rag-mcp index ./docs/

pip from GitHub (latest)

pip install git+https://github.com/aliaksandr-kazarez/multilingual-rag-mcp.git

Quick start

1. Index your documents

rag-mcp index ~/documents/

2. Add to Claude Code

claude mcp add rag -- uvx --from multilingual-rag-mcp rag-mcp

Set the document paths via env vars:

claude mcp add rag \
  -e RAG_DOCS=$HOME/documents \
  -- uvx --from multilingual-rag-mcp rag-mcp

If installed locally (pip install), use the simpler form:

claude mcp add rag -e RAG_DOCS=$HOME/documents -- rag-mcp

3. Search

From Claude Code, the search tool handles cross-lingual queries automatically:

• "search for protein recommendations" finds Russian articles about белок
• "найди рецепты" finds recipe content regardless of language

MCP tools

Configuration

Supported formats

Markdown (.md), plain text (.txt), PDF (.pdf), JSON (.json), CSV (.csv).

Markdown files with YAML frontmatter have their metadata (title, date, etc.) extracted automatically.

CLI

rag-mcp index <dir> [<dir> ...]   Index documents
rag-mcp stats                     Print index statistics
rag-mcp                           Start MCP server (stdio)
rag-mcp --help                    Show help
rag-mcp --version                 Show version

How it works

1. Documents are parsed and split into chunks (markdown-aware: respects headers, code blocks)
2. Each chunk is embedded with a multilingual model via fastembed (ONNX, no PyTorch needed)
3. Embeddings are stored in ChromaDB (local, file-based)
4. Queries are embedded with the same model and matched by cosine similarity
5. Cross-lingual retrieval works because the model places semantically similar text from different languages near each other in vector space

License

MIT