mcp-name: io.github.nipunkhanderia/golden-dataset-mcp

golden-dataset-mcp

An MCP server wrapping golden-dataset-studio — version-controlled golden dataset management and semantic evaluation for RAG/LLM pipelines.

This is a thin protocol layer over the existing golden_dataset library (DatasetStore, Evaluator). It does not reimplement any logic — it exposes the library's existing Python API as MCP tools so an agent (Claude Desktop, Claude Code, or any MCP client) can manage golden datasets conversationally.

No LLM API key required. Evaluation uses TF-IDF cosine similarity (scikit-learn), not an LLM call.

Why a separate package from golden-dataset-studio?

golden-dataset-studio is a CLI tool — designed for a human typing golden add, golden commit, etc. in a terminal. golden-dataset-mcp exposes the same underlying operations as MCP tools so an LLM agent can drive them programmatically, e.g. as part of an automated RAG evaluation pipeline. Keeping them as separate PyPI packages means CLI users aren't forced to pull in fastmcp as a dependency, and MCP users get a clean, protocol-focused package.

Tools

Design: every tool takes an explicit dataset_path

Unlike the CLI (which operates on the current working directory), every tool here requires an explicit dataset_path parameter. This keeps the server fully stateless between calls — no hidden "current dataset" session state to lose track of, and safe for one server instance to manage multiple datasets or serve multiple concurrent clients.

Installation

pip install golden-dataset-mcp

This pulls in golden-dataset-studio and scikit-learn automatically as dependencies.

Usage with Claude Desktop / Claude Code

{
  "mcpServers": {
    "golden-dataset": {
      "command": "golden-dataset-mcp"
    }
  }
}

No environment variables needed — no API key, no config.

Example flow

1. init_dataset(dataset_path="./my-rag-eval", name="support-bot-eval")
2. add_entry(dataset_path="./my-rag-eval", question="...", answer="...")
   [repeat for each golden Q&A pair]
3. commit_version(dataset_path="./my-rag-eval", description="initial 50 questions")
4. [run your RAG pipeline, collect actual answers]
5. evaluate_answers(dataset_path="./my-rag-eval", actual_answers=[...])
   -> avg_semantic_similarity, per-entry scores, pass/fail

As your RAG pipeline changes over time, commit_version again after edits and use diff_versions to see exactly what changed in your golden set between releases.

Relationship to the underlying library

If you want the human-driven CLI, use golden-dataset-studio directly. If you want an agent to drive it, use this package.

Development

git clone https://github.com/nipunkhanderia/golden-dataset-mcp
cd golden-dataset-mcp
pip install -e ".[dev]"
pytest -v

Validate the MCP-facing contract:

npx @modelcontextprotocol/inspector golden-dataset-mcp

Limitations

• evaluate_answers uses TF-IDF cosine similarity, which captures lexical overlap better than deep semantic meaning. For embedding-based or RAGAS-style metrics, call the underlying library's Evaluator.ragas_evaluate() directly (requires pip install "golden-dataset-studio[ragas]" — not exposed as an MCP tool in this version).
• Very short or stop-word-only answers will raise an error. scikit-learn's TF-IDF vectorizer raises ValueError: empty vocabulary on inputs like a bare "4" or "the a an". Avoid single-token golden answers, or expect evaluate_answers to fail on them.
• All state is filesystem-backed JSON/JSONL under <dataset_path>/.golden_dataset/; this server does no remote storage or syncing.

License

MIT