🛡️ POLICY-QUERY-SYSTEM

End-to-End LLM Decision Engine for Insurance, Legal & HR Docs
Hackathon-Ready | Gemini-Powered | PDF-to-Decision in Seconds

---

🚀 QUICK DEMO

git clone https://github.com/vanstyagi/policy-llm-decision-engine.git
cd policy-query-system
pip install -r requirement.txt
echo "GEMINI_API_KEY=YOUR_KEY_HERE" > .env
python src/main.py "46-year-old male, knee surgery in Pune, 3-month policy"

✔️ Output will show in terminal and outputs/final.json like:

{
  "decision": "approved",
  "amount": null,
  "justification": "Knee surgery is covered. Age & location are valid per policy clauses.",
  "clauses": ["data_chunk95.txt", "data4_chunk56.txt", "data_chunk52.txt"]
}

---

📦 PROJECT STRUCTURE

Folder / File	Purpose
data/	Input folder for PDF / DOCX / EML
outputs/chunks/	Auto-generated document chunks (300 words)
outputs/final.json	Final structured JSON output
src/check_data.py	List available documents
src/chunk_docs.py	Split documents into chunks
src/load_docs.py	Extract raw text
src/load_chunks.py	Load all chunk files
src/parse_query.py	Parse natural query → structured JSON
src/retrieve.py	Semantic search via sentence-transformers
src/decide.py	LLM (Gemini) generates decision JSON
src/main.py	CLI entry point for full pipeline
.env	Store API Key (never commit)
requirement.txt	All dependencies
README.md

---

⚙️ INSTALLATION & USAGE

| 🧩 Step | 🛠️ Command | | 1️⃣ Clone | git clone <repo-url> | | 2️⃣ Install Deps | pip install -r requirement.txt | | 3️⃣ Set API Key | echo "GEMINI_API_KEY=YOUR_KEY" > .env | | 4️⃣ Run | python src/main.py "<your query>" | | 5️⃣ Output | cat outputs/final.json |

---

✅ FEATURE CHECKLIST

Capability	✅ Status
Natural-language query (free-text)	✅
Parse age, gender, location, procedure	✅
Ingest PDF, DOCX, and EML	✅
Semantic search using embeddings	✅
Gemini-based approve/reject reasoning	✅
Clause-level mapping	✅
JSON output for downstream usage	✅

---

💬 SAMPLE QUERIES

python src/main.py "46-year-old male, knee surgery in Pune, 3-month policy"
python src/main.py "23-year-old female, cardiac surgery, Mumbai, 6-month policy"
python src/main.py "60-year-old man, hip replacement, Delhi, 1-year policy"

---

🧠 TECH STACK

• Python 3.11
• PyPDF2, python-docx – document ingestion
• Sentence-Transformers – semantic similarity search
• Google Gemini 1.5 Flash – final decision reasoning
• dotenv – key management

---

🔄 PIPELINE OVERVIEW

🧩 Stage	🔍 Description
parse_query.py	Regex / NLP → extract query entities
retrieve.py	Embed + semantic search over document chunks
decide.py	Gemini generates final structured JSON
final.json	Output → usable by any downstream system

🧾 Flow:

Plain-English Query
        ↓
parse_query.py
        ↓
retrieve.py (embedding + search)
        ↓
decide.py (LLM reasoning)
        ↓
final.json

---

📜 LICENSE

MIT © VANSH TYAGI

---