🏥 LLM RAG ETL Showcase: Hosto-bot

Hosto-bot is an advanced Agentic Microservices demonstration combining Large Language Models (LLMs) with a Graph Database (Neo4j) to deliver a powerful Hybrid RAG (Retrieval-Augmented Generation) system.

Traditional RAG systems often struggle with structured aggregation and relationship-heavy queries. Hosto-bot addresses this limitation using a Router Agent that dynamically selects the best tool:

• 🔎 Vector Search → qualitative insights (patient sentiment)
• 🧠 Graph Cypher Generation → quantitative analytics
• ⚙️ Python Tools → real-time simulations and computations

_{Hosto-bot UI: Real-time chat with visible Agent reasoning traces.}

🎯 Project Goals & Context

Healthcare data is inherently fragmented:

• Structured Data: billing, visits, hospitals
• Unstructured Data: patient reviews
• Real-time Data: ER wait times, bed availability

Hosto-bot provides a unified natural language interface allowing users (administrators, patients) to query across all data types without needing SQL, Cypher, or Python.

---

🏗️ System Architecture

The system follows a decoupled microservices architecture.

1. Data Layer – ETL Engine (/hospital_neo4j_etl)

Responsible for building the data foundation:

• Data ingestion: loads 6 datasets (Hospitals, Payers, Physicians, Patients, Visits, Reviews)
• Graph construction:
  • Enforces uniqueness constraints
  • Builds relationships (e.g. (:Physician)-[:TREATS]->(:Visit))
• Vectorization:
  • Creates Neo4j vector index on Review nodes
  • Uses OpenAI embeddings for semantic search
• Resilience:
  • Retry mechanism to handle startup delays

---

2. Logic Layer – Agentic API (/chatbot_api)

Core intelligence built with FastAPI + LangChain.

• Router Agent:

  • Uses create_openai_functions_agent
  • Selects optimal tool based on intent

• Hybrid Querying:

  • Graph Tool → dynamic Cypher queries
  • Experiences Tool → vector similarity search
  • Wait-Time Tool → Python-based simulation (NumPy)

• Error Handling:

  • Async retry decorator for API reliability

---

3. Presentation Layer – Chatbot UI (/chatbot_frontend)

Built with Streamlit.

• Maintains session state
• Displays chat history
• Provides full transparency:
  • Shows tool selection
  • Displays generated queries and reasoning

---

📂 Project Structure

LLM_RAG_ETL_Showcase-Hostobot/
├── .env                        # Critical environment configurations
├── docker-compose.yml          # Orchestration for all 3 services
├── requirements.txt            # Root dependencies (mostly for local dev)
│
├── img/                        # Images to display in the Readme
│   ├── interface.png
│   ├── dbscheme.png
│
├── hospital_neo4j_etl/         # DATA LAYER
│   ├── src/
│   │   ├── hospital_bulk_csv_write.py  # Bulk loader & relationship mapper
│   │   └── entrypoint.sh               # Execution script
│   └── Dockerfile
│
├── chatbot_api/                # LOGIC LAYER (FastAPI)
│   ├── src/
│   │   ├── main.py                     # API entry point & routes
│   │   ├── agents/
│   │   │   └── hospital_rag_agent.py   # ReAct Agent & Tool definitions
│   │   ├── chains/
│   │   │   ├── hospital_cypher_chain.py # Cypher generation logic
│   │   │   └── hospital_review_chain.py # Vector search logic
│   │   ├── tools/
│   │   │   └── wait_times.py            # Simulated real-time tool
│   │   ├── models/
│   │   │   └── hospital_rag_query.py    # Pydantic schemas
│   │   └── utils/
│   │       └── async_utils.py          # Retry decorators
│   └── Dockerfile
│
└── chatbot_frontend/           # PRESENTATION LAYER
    ├── src/
    │   └── main.py                     # Streamlit UI logic
    └── Dockerfile

📊 Knowledge Graph Schema

The Neo4j database is modeled to handle multi-hop questions.

Nodes

• Hospital: {id, name, state_name}
• Payer: {id, name}
• Physician: {id, name, dob, school, salary}
• Patient: {id, name, sex, blood_type}
• Visit: {id, room_number, admission_type, status, diagnosis}
• Review: {id, text, patient_name, physician_name}

Relationships

• (Patient)-[:HAS]->(Visit)
• (Physician)-[:TREATS]->(Visit)
• (Visit)-[:AT]->(Hospital)
• (Visit)-[:COVERED_BY]->(Payer)
• (Visit)-[:WRITES]->(Review)
• (Hospital)-[:EMPLOYS]->(Physician)

_{Graph Ontology: Mapping relationships between Patients, Visits, and Providers.}

🚀 Installation & Deployment

1. Configuration

Create a .env file in the root directory:

# API Keys
OPENAI_API_KEY=sk-...

# Neo4j Database (AuraDB or Local)
NEO4J_URI=neo4j+s://...
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=...
NEO4J_DATABASE=neo4j

# Data Source URLs
HOSPITALS_CSV_PATH=https://.../hospitals.csv
# ... (Fill in paths for Payers, Physicians, Patients, Visits, Reviews)

# Model Settings
HOSPITAL_AGENT_MODEL=gpt-3.5-turbo-1106
HOSPITAL_CYPHER_MODEL=gpt-3.5-turbo-1106
HOSPITAL_QA_MODEL=gpt-3.5-turbo-0125

# Internal Networking
CHATBOT_URL=http://chatbot_api:8000/hospital-rag-agent

### 2. Launch with Docker

Ensure Docker Desktop is running, then execute:

```bash
docker-compose up --build

This command builds the images and starts the containers in the correct order:

• ETL
• API
• Frontend

---

3. Usage

• Frontend UI: http://localhost:8501
• API Documentation: http://localhost:8000/docs

---

🛠️ Tech Stack

• Language: Python 3.11+
• Graph Database: Neo4j (Cypher)
• LLM Orchestration: LangChain
• LLM Provider: OpenAI GPT-3.5 / GPT-4
• Backend: FastAPI (Asynchronous)
• Frontend: Streamlit
• Containerization: Docker & Docker Compose

---

💬 Sample Queries

Type	Query
Quantitative	"What was the total amount billed to each payer during 2023?"
Qualitative	"How are patients describing the nursing team at Castaneda-Hardy?"
Real-time	"What’s the current wait time at Wallace-Hamilton Hospital?"
Comparative	"Which doctor has the shortest average visit duration?"

💬 More Sample Queries

• Which hospitals are part of the hospital network?
• What’s the current wait time at Wallace-Hamilton Hospital?
• At which hospitals are patients reporting issues related to billing or insurance?
• What’s the average length in days for completed emergency visits?
• How are patients describing the nursing team at Castaneda-Hardy?
• What was the total amount billed to each payer during 2023?
• What is the average charge for visits covered by Medicaid?
• Which doctor has the shortest average visit duration?
• What is the total billed amount for patient 789's hospital stay?
• Which state saw the biggest percentage increase in Medicaid visits from 2022 to 2023?
• What’s the average daily billing amount for patients with Aetna coverage?
• How many patient reviews have been submitted from Florida?
• For visits that include a chief complaint, what percentage also have a review?
• What percentage of visits at each hospital include patient reviews?
• Which physician has received the highest number of reviews for their visits?
• What is the unique identifier for Dr. James Cooper?
• Show all reviews associated with visits handled by physician 270 — include every one.