Document Processing Pipeline

A FastAPI service that processes insurance claim PDFs using LangGraph to orchestrate document segregation and multi-agent data extraction powered by the Gemini API.

---

Project Structure

├── main.py                  # FastAPI app and endpoint
├── graph.py                 # LangGraph workflow definition
├── agents/
│   ├── segregator.py        # Classifies pages into document types
│   ├── id_agent.py          # Extracts patient & policy identity info
│   ├── discharge_agent.py   # Extracts discharge summary details
│   ├── bill_agent.py        # Extracts itemized billing information
│   └── aggregator.py        # Combines all agent outputs
├── utils/
│   └── pdf_parsing.py       # PDF → pages (text + base64 image)
├── config/
│   └── llm_setup.py         # Gemini client initialisation
├── .env                     # API keys (not committed)
└── requirements.txt

---

Setup

1. Clone & install dependencies

git clone <your-repo-url>
cd <repo-folder>
pip install -r requirements.txt

2. Configure environment variables

Create a .env file in the root:

GEMINI_API_KEY2=your_gemini_api_key_here

3. Run the server

uvicorn main:app --reload

The API will be available at http://localhost:8000.

---

API

POST /api/process

Processes a PDF claim file and returns structured extracted data.

Request — multipart/form-data

Field	Type	Description
claim_id	string	Unique identifier for the claim
file	file	PDF file to process

Example using curl

curl -X POST http://localhost:8000/api/process \
  -F "claim_id=CLM001" \
  -F "file=@/path/to/claim.pdf"

Response — application/json

{
  "metadata": {
    "processed_at": "2026-04-15T10:00:00Z",
    "status": "success",
    "errors": null
  },
  "patient": {
    "name": "Ravi Kumar",
    "date_of_birth": "12/05/1985",
    "id_numbers": ["XXXX-XXXX-4321", "ABCPK1234Q"],
    "policy_number": "HDFC123456789",
    "policy_holder": "Ravi Kumar",
    "insurance_company": "HDFC Ergo"
  },
  "hospitalization": {
    "hospital_name": "Apollo Hospitals",
    "admission_date": "10/03/2026",
    "discharge_date": "14/03/2026",
    "length_of_stay_days": 4,
    "final_diagnosis": ["Dengue Fever"],
    "treating_physician": "Dr. Suresh Reddy",
    "procedures_performed": [],
    "medications_on_discharge": ["Paracetamol 500mg", "Dolo 650"]
  },
  "billing": {
    "line_items": [
      { "description": "Room Charges", "amount": 3000 },
      { "description": "Medicine Charges", "amount": 1500 },
      { "description": "Lab Charges", "amount": 2000 }
    ],
    "total_amount": 11800,
    "currency": "INR",
    "payment_mode": "Cash"
  }
}

---

LangGraph Workflow

START
  │
  ▼
segregation_node        ← Classifies every PDF page into one of 9 document types
  │
  ▼
id_agent_node           ← Extracts patient name, DOB, policy & ID numbers
  │
  ▼
discharge_agent_node    ← Extracts diagnosis, admission/discharge dates, physician
  │
  ▼
bill_agent_node         ← Extracts all line items, totals, payment details
  │
  ▼
aggregator_node         ← Merges all agent outputs into final JSON
  │
  ▼
END

All three extraction agents read from classified_pages produced by the segregator. Each agent only receives the pages relevant to it — the ID agent never sees billing pages, etc.

Document Types (9)

Type	Description
identity_document	Patient ID, Aadhaar, PAN, policy card
claim_forms	Insurance claim forms
discharge_summary	Hospital discharge records
itemized_bill	Hospital bills with line items
prescription	Doctor prescriptions
investigation_report	Lab / diagnostic reports
cash_receipt	Payment receipts
cheque_or_bank_details	Bank account / cheque info
other	Anything that doesn't fit above

---

Requirements

fastapi
uvicorn
python-multipart
langgraph
google-genai
pymupdf
Pillow
python-dotenv
typing_extensions

---

Development Tips

• Use mock_segregation() in segregator.py during development to avoid spending API quota — it classifies pages with simple keyword matching, no API calls.
• The Gemini free tier has per-minute rate limits. The safe_generate / safe_segregate functions use exponential backoff (15s → 30s → 60s) and only retry on known transient errors (429, 500, 503).
• Image-only PDF pages (scanned documents) are handled by converting the page to PNG bytes and passing them to Gemini via types.Part.from_bytes().