Tabular Document Retriever MCP

This is a Model Context Protocol (MCP) server that transforms tabular data (CSV/Excel) into Markdown key-value pairs, embeds them into a local vector database (ChromaDB), and provides retrieval tools to contextually answer queries.

It leverages the python uv package manager, mcp SDK, and FastAPI to optionally expose the server using Server-Sent Events (SSE).

Vibe-Coded

The entire code base was generated by Antigravity with the help of Gemini 3.1 Pro (High and Fast) to make a fast proof of concept.

Features

• Ingestion: Parses .csv and .xlsx files and upserts Markdown-formatted strings into ChromaDB.

• Retrieval Engine: Uses sentence-transformers/all-MiniLM-L6-v2 locally for exact semantic search.

• MCP Server: Provides three tools exposed over an SSE endpoint:

  • retrieve_batch

  • retrieve_single

  • retrieve_by_query

• Dockerization: Quick spin-up of the Database and the MCP Server together without exposing the raw database to the host machine.

Prerequisites

• Docker and Docker Compose.

• uv (for local testing/installation).

• Python 3.12.12.

🚀 Running the Stack

To start the server and the ChromaDB vector database locally:

docker-compose up -d --build

This will launch:

1. ChromaDB internally on chroma-db:8000.

2. MCP Server accessible externally on http://localhost:8000/sse.

💾 Ingesting User Data

Before you can search, you need to ingest tabular data into the running ChromaDB instance.

You can use the built CLI ingestor directly from your host machine. Make sure to map environment variables appropriately to reach your local stack or run it via Docker Compose.

To run the ingestor against a locally running ChromaDB (or inside the container):

# First, ensure dependencies are synced
uv sync

# Run the ingestor (Assuming there's a file `data/my_table.csv`)
# When interacting with the dockerized ChromeDB, make sure to temporarily expose port 8000 for chroma-db, OR simply just run ingestion locally with local persistence.
uv run python -m src.ingestor data/my_table.csv

Note: Since the docker stack makes ChromaDB private, you can either map a port for chroma-db in docker-compose.yml temporarily, or run a one-off task using docker-compose:

docker-compose exec mcp-app python src/ingestor.py /path/to/mounted/data.csv

🛠️ MCP Tools

Once running, any MCP client can connect to http://localhost:8000/sse via Server-Sent Events (SSE).

Available tools:

• retrieve_single(row): Top-K search using a single row's markdown string.

• retrieve_batch(rows): Batch retrieval handling a list of markdown row strings.

• retrieve_by_query(query): Free-text query mapped exactly to ChromaDB's search.

💻 Local Testing Example

You can test the running MCP server locally using the official Python SDK. First, ensure you have the mcp package installed in your environment (uv pip install mcp or uv add mcp).

Run the example with:

uv run python tests/test_client.py