ickyMCP

RAG MCP Server for Document Search. Built for legal professionals and business users who need to search across large document collections.

Features

• Semantic Search: Find relevant content based on meaning, not just keywords

• Document Support: PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), Markdown, Text

• 4K Token Chunks: Large chunks preserve context for legal and business documents

• Incremental Indexing: Only re-index changed files

• Local Embeddings: Uses nomic-embed-text-v1.5 (no API costs)

• SQLite Storage: Single portable database file

Installation

# Clone or copy the project
cd ickyMCP

# Create virtual environment
python -m venv venv
source venv/bin/activate  # or `venv\Scripts\activate` on Windows

# Install dependencies
pip install -r requirements.txt

# Or install as package
pip install -e .

Configuration

Environment Variables

Claude Code Configuration

Add to your claude_desktop_config.json or MCP settings:

{
  "mcpServers": {
    "ickyMCP": {
      "command": "python",
      "args": ["/path/to/ickyMCP/run.py"],
      "env": {
        "ICKY_CHUNK_SIZE": "4000",
        "ICKY_CHUNK_OVERLAP": "500",
        "ICKY_DB_PATH": "/path/to/icky.db"
      }
    }
  }
}

Usage

Tools Available

index

Index documents from a file or directory.

index(path="/contracts/2024", patterns=["*.pdf", "*.docx"])

search

Semantic search across indexed documents.

search(query="indemnification clause", top_k=10, file_types=["pdf"])

similar

Find chunks similar to a given text.

similar(chunk_text="The parties agree to...", top_k=5)

refresh

Re-index only files that have changed.

refresh(path="/contracts")

list

List all indexed documents.

list(path_filter="/contracts")

delete

Remove documents from the index.

delete(path="/contracts/old")
delete(all=true)  # Clear entire index

status

Get server status and statistics.

status()

How It Works

1. Indexing: Documents are parsed, split into 4K token chunks with 500 token overlap

2. Embedding: Each chunk is embedded using nomic-embed-text-v1.5 (768 dimensions)

3. Storage: Embeddings stored in SQLite with sqlite-vec for fast vector search

4. Search: Query is embedded, compared against all chunks using cosine similarity

5. Results: Top-K most similar chunks returned with full text and metadata

System Requirements

• Python 3.10+

• 4GB RAM (2GB for model + headroom)

• ~1GB disk space (model + database)

License

MIT