PolicyLens - Medical Benefit Drug Policy Tracker

Query and compare drug coverage policies across health insurance payers using Vision-Language Models, a knowledge graph, and Model Context Protocol (MCP).

---

Table of Contents

1. Overview
2. Architecture
3. Prerequisites
4. Quick Start
5. Configuration
6. Usage
7. MCP Servers
8. Example Queries
9. Project Structure

---

Overview

Health insurance payers publish hundreds of Medical Benefit Drug Policy PDFs every year. These documents define which drugs are covered, under what conditions (prior authorisation, step therapy, quantity limits), and for which diagnoses. Manually comparing policies across payers or tracking changes over time, is error-prone and time-consuming.

PolicyLens automates this workflow:

1. Ingest - Upload a policy PDF. A Vision-Language Model (Claude) reads every page (including scanned tables) and extracts structured data.
2. Graph - Extracted entities (drugs, indications, criteria, plans) are stored in a Neo4j knowledge graph with rich relationships.
3. Retrieve - Raw policy excerpts are indexed in ChromaDB for semantic search with page-level citations.
4. Query - A Claude agent connects to four MCP servers and answers natural language questions with citations and a transparent reasoning trace.

---

Architecture

flowchart TD
    subgraph Ingestion
        A[PDF Upload] --> B[PDF Loader\npymupdf]
        B --> C[VLM Extractor\nClaude claude-sonnet-4-6]
        C --> D[ExtractedPolicy\nPydantic model]
        D --> E[Graph Builder\nNeo4j]
        D --> F[Excerpt Store\nChromaDB]
    end

    subgraph Knowledge Store
        E --> G[(Neo4j\nGraph DB)]
        F --> H[(ChromaDB\nVector Store)]
    end

    subgraph MCP Layer
        G --> I[Policy Graph\nMCP Server]
        H --> J[Document Retrieval\nMCP Server]
        K[RxNorm / ICD-10\nAPIs] --> L[Reference Data\nMCP Server]
        G --> M[Change Detection\nMCP Server]
    end

    subgraph API & UI
        N[FastAPI\nBackend] -->|MCP client| I
        N -->|MCP client| J
        N -->|MCP client| L
        N -->|MCP client| M
        O[Streamlit\nDemo UI] --> N
        P[Claude Desktop] -->|stdio| I
        P -->|stdio| J
    end

    User -->|question| O
    O -->|answer + citations| User

Loading

---

Prerequisites

Requirement	Version
Python	3.11+
Docker + Docker Compose	24+
Anthropic API key	-

---

Quick Start

# 1. Clone and enter the repo
git clone <repo-url> policy-lens
cd policy-lens/medical-policy-tracker

# 2. Create a virtual environment
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

# 3. Install dependencies
pip install -e ".[dev]"

# 4. Copy and fill in the env file
cp .env.example .env
# Edit .env - set ANTHROPIC_API_KEY and NEO4J_PASSWORD at minimum

# 5. Start the databases
docker compose up -d

# 6. Initialise the Neo4j schema (constraints + indexes)
python scripts/init_neo4j.py

# 7. Download sample policies (edit URLs in the script first)
python scripts/download_sample_pdfs.py

# 8. Ingest a policy PDF
python scripts/ingest_policy.py --pdf data/raw_pdfs/sample_aetna.pdf

# 9. Start the API
uvicorn api.main:app --reload

# 10. Start the UI (separate terminal)
streamlit run ui/streamlit_app.py

---

Configuration

All settings are read from environment variables (see .env.example).

Variable	Default	Description
ANTHROPIC_API_KEY	-	Required. Anthropic API key
NEO4J_URI	bolt://localhost:7687	Neo4j Bolt URI
NEO4J_USER	neo4j	Neo4j username
NEO4J_PASSWORD	-	Required. Neo4j password
CHROMA_PERSIST_DIR	./data/chroma	ChromaDB persistence path
VLM_MODEL	claude-sonnet-4-6	Claude model for extraction
LOG_LEVEL	INFO	Logging level
API_HOST	0.0.0.0	FastAPI bind host
API_PORT	8000	FastAPI bind port

---

Usage

Running the API

uvicorn api.main:app --host 0.0.0.0 --port 8000 --reload

Interactive docs at http://localhost:8000/docs.

Running the Streamlit UI

streamlit run ui/streamlit_app.py

Opens at http://localhost:8501.

Running MCP Servers standalone

# Run any server over stdio (for direct MCP client use)
python scripts/run_mcp_server.py policy_graph
python scripts/run_mcp_server.py document_retrieval
python scripts/run_mcp_server.py reference_data
python scripts/run_mcp_server.py change_detection

Connecting Claude Desktop to the MCP servers

Add the following to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "policy-graph": {
      "command": "python",
      "args": ["/absolute/path/to/scripts/run_mcp_server.py", "policy_graph"],
      "env": {
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USER": "neo4j",
        "NEO4J_PASSWORD": "changeme"
      }
    },
    "document-retrieval": {
      "command": "python",
      "args": ["/absolute/path/to/scripts/run_mcp_server.py", "document_retrieval"]
    }
  }
}

---

Example Queries

"Does Aetna cover pembrolizumab for NSCLC, and what are the prior auth criteria?"

"Compare step therapy requirements for adalimumab between Cigna and UnitedHealthcare."

"Which payers updated their GLP-1 agonist policies in the last 90 days?"

"What ICD-10 codes does Humana accept for dupilumab coverage in atopic dermatitis?"

"Show me all drugs in the anti-VEGF class covered by Anthem's commercial plans."

---

Project Structure

medical-policy-tracker/
├── data/                    # PDFs and extracted JSON (gitignored)
├── src/
│   ├── config.py            # Centralised settings (pydantic-settings)
│   ├── models/              # Pydantic data models
│   ├── ingestion/           # PDF loading, VLM extraction, graph building
│   ├── graph/               # Neo4j client, schema, query functions
│   ├── vector_store/        # ChromaDB excerpt store
│   ├── reference/           # RxNorm and ICD-10 API wrappers
│   └── mcp_servers/         # Four MCP servers (stdio transport)
├── api/                     # FastAPI backend
├── ui/                      # Streamlit demo
├── scripts/                 # CLI helpers and one-shot setup scripts
└── tests/                   # pytest test suite

---

Built with Claude · Neo4j · MCP