mcp-neo4j-lexical-graph

MCP server for creating rich lexical graphs from PDF documents in Neo4j. Designed for Neo4j sales engineers to quickly build PDF-to-graph and GraphRAG agent chatbot POCs.

Supports four parsing strategies (PyMuPDF, Docling, page-image, VLM block ordering), pluggable chunking, document versioning, VLM-based description generation, and vector/fulltext search with Neo4j 2026.01 native VECTOR type and document-name prefiltering.

Graph Model

graph LR
    Doc[Document] -->|HAS_PAGE| Page
    Doc -->|HAS_ELEMENT| Img[Image]
    Doc -->|HAS_ELEMENT| Tbl[Table]
    Doc -->|HAS_SECTION| Sec[Section]
    Sec -->|HAS_SUBSECTION| Sec
    Chunk -->|PART_OF| Doc
    Chunk -->|NEXT_CHUNK| Chunk
    Chunk -->|HAS_ELEMENT| Img
    Chunk -->|HAS_ELEMENT| Tbl
    Page -->|NEXT_PAGE| Page

Node types depend on the parse mode used. See Parse Modes below.

Parse Modes

Quick Start

cd mcp-neo4j-lexical-graph
uv sync

Cursor MCP Configuration

Add to your .cursor/mcp.json:

{
  "mcpServers": {
    "neo4j-lexical-graph": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/mcp-neo4j-lexical-graph",
        "run",
        "mcp-neo4j-lexical-graph"
      ],
      "env": {
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USERNAME": "neo4j",
        "NEO4J_PASSWORD": "your-password",
        "NEO4J_DATABASE": "neo4j",
        "EMBEDDING_MODEL": "text-embedding-3-small",
        "EXTRACTION_MODEL": "gpt-5.4-mini",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Tools

Tools must be called in a specific order — which tools to call depends on the parse mode and document type. See the workflow table below.

Workflow Order

¹ Recommended when extract_images=True or extract_tables=True (pymupdf) or when the document contains images/tables (docling/vlm_blocks). Without descriptions, Image/Table nodes are invisible to semantic search.

Tool Reference

Environment Variables

Requirements

• Neo4j 2026.01+ (native VECTOR type, vector search with filters)
• Python 3.10+
• API key for your embedding provider (OpenAI, Azure, Cohere, Voyage, Ollama, etc.)
• API key for VLM if using vlm_blocks mode, generate_chunk_descriptions, or assign_section_hierarchy