🧠 HCLTech Agentic AI — Enterprise Assistant

A production-ready, "Guardrails-First" enterprise AI assistant that retrieves financial intelligence, automates IT workflows, and manages corporate actions — powered by RAG + Groq LLaMA-3.

Live Demo • API Docs • Setup Guide

---

📌 Project Overview

The HCLTech Agentic AI is a full-stack enterprise assistant built for the NLP Challenge. It combines:

• RAG pipeline over HCLTech's Annual Report 2024–25 (FAISS + SentenceTransformers)
• Groq LLaMA-3 70B for high-speed, grounded answer synthesis
• Multi-intent detection (finance, HR, IT, scheduling, access)
• Agentic actions — structured JSON output for IT tickets, access requests, meeting scheduling
• Premium Next.js dashboard with glass-morphic UI and real-time chat

---

🏗️ Architecture

User Query
    │
    ▼
┌─────────────────────────────────────────────────────────┐
│                    FastAPI Backend                       │
│                                                         │
│  ┌──────────────┐   ┌──────────────┐   ┌────────────┐  │
│  │Intent Detect │──▶│ NER Extract  │──▶│Agent Policy│  │
│  │(Groq LLaMA-3)│   │(Groq LLaMA-3)│   │   Layer    │  │
│  └──────────────┘   └──────────────┘   └─────┬──────┘  │
│                                               │         │
│         ┌──────────────────┬─────────────────┤         │
│         ▼                  ▼                 ▼         │
│   ┌──────────┐      ┌──────────┐      ┌──────────┐    │
│   │RAG (FAISS│      │ Action   │      │Clarify / │    │
│   │+ mpnet)  │      │Generator │      │Escalate  │    │
│   └────┬─────┘      └────┬─────┘      └──────────┘    │
│        ▼                 ▼                             │
│   ┌──────────────────────────────┐                    │
│   │  Groq LLaMA-3 Synthesizer    │                    │
│   └──────────────────────────────┘                    │
└─────────────────────────────────────────────────────────┘
    │
    ▼
Next.js Frontend Dashboard (Port 3000)

Backend Modules

Module	Purpose
api_server.py	FastAPI entry point — /api/chat, /health, /docs
main_assistant.py	Orchestrates the full 5-step pipeline
intent_detector.py	Groq-powered zero-shot intent classification
ner_extractor.py	Groq-powered named entity extraction
sentiment_analyzer.py	Urgency & sentiment analysis via Groq
query_assistant.py	FAISS vector retrieval with RRF re-ranking
agent_policy.py	Decides: answer / action / clarify / escalate
action_generator.py	Structured JSON for tickets, access, meetings
clarifier.py	Generates targeted clarification prompts
citation_enforcer.py	Validates citations from retrieved chunks
process_pdf.py	PDF ingestion & chunk creation
index_chunks.py	Builds FAISS index from chunks
docs_html.py	In-memory HTML project documentation page
gradio_app.py	Optional Gradio interface

Frontend Pages

Route	Description
/	Login page (username + password)
/dashboard	Main chat interface + Finance, Support, Calendar, Access panels

---

✨ Key Features

🎯 Multi-Intent Detection

• 12 intent classes: ask_finance, ask_hr, ask_it_policy, ask_compliance, ask_security, ask_procurement, ask_people, action_ticket, action_access, action_schedule, and more
• Groq LLaMA-3 as primary classifier with keyword-boost fallback
• Multi-turn context stickiness for follow-up queries

📚 Enterprise RAG

• HCLTech Annual Report 2024–25 chunked into 500-token segments
• all-mpnet-base-v2 (768-dim) embeddings → FAISS IndexFlatL2
• Hybrid retrieval: Semantic similarity + lexical overlap (RRF scoring)
• Section-aware filtering (Financial / HR / Governance / IT)
• Direct Groq fallback for queries with no RAG match

⚙️ Agentic Actions

• IT Ticket — structured JSON with priority, department, reference ID
• Access Request — scoped to application and employee
• Meeting Scheduling — extracts topic, date, time, participants

🛡️ Guardrails-First Policy

• Confidence thresholding before answer / action decision
• Entity validation against retrieved chunks
• Human escalation trigger for high-urgency / low-confidence cases

💎 Premium Dashboard

• Dark glassmorphic Next.js UI (shadcn/ui components)
• Real-time chat with structured JSON response cards
• Finance ledger, Support tickets, Calendar, Access panels
• Persistent local storage for meetings and tickets

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+
• A Groq API key

1. Clone & Configure

git clone https://github.com/Shridipa/HCL-Agentic-AI.git
cd HCL-Agentic-AI

Create .env in the project root:

GROQ_API_KEY=your_groq_api_key_here

2. Backend Setup

cd backend
pip install -r requirements.txt
python api_server.py

Backend runs at → http://localhost:8000

3. Frontend Setup

cd frontend

Create frontend/.env.local:

ML_MODEL_ENDPOINT=http://localhost:8000/api/chat
GROQ_API_KEY=your_groq_api_key_here

npm install
npm run dev

Frontend runs at → http://localhost:3000

---

🔌 API Endpoints

Method	Endpoint	Description
GET	/	Interactive project documentation
GET	/health	Health check
GET	/docs	Swagger UI
GET	/redoc	ReDoc API reference
POST	/api/chat	Main chat endpoint

/api/chat Payload

{
  "query": "What was HCLTech's revenue growth in FY25?",
  "user": "Alice",
  "history": []
}

Response Schema

{
  "reply": "{\"title\": \"Answer: Revenue Growth\", \"direct_answer\": \"...\", \"key_insights\": [...], \"citations\": [...], \"confidence_score\": \"High\"}",
  "confidence": 0.9
}

---

💬 Example Queries

Type	Example
Finance	"What was HCLTech's revenue in FY25?"
People	"Who is the CEO of HCLTech?"
HR	"What is the maternity leave policy?"
IT Ticket	"My laptop screen is flickering, raise a high priority ticket"
Access	"I need access to the SAP portal"
Scheduling	"Schedule a meeting with the finance team for next Monday at 10am"

---

📁 Project Structure

HCL-Agentic-AI/
├── .env                          # Root env (GROQ_API_KEY)
├── .gitignore
├── README.md
├── docker-compose.yml
│
├── backend/
│   ├── api_server.py             # FastAPI server
│   ├── main_assistant.py         # Pipeline orchestrator
│   ├── intent_detector.py        # Intent classification
│   ├── ner_extractor.py          # Entity extraction
│   ├── sentiment_analyzer.py     # Sentiment & urgency
│   ├── query_assistant.py        # FAISS retrieval
│   ├── agent_policy.py           # Policy decision layer
│   ├── action_generator.py       # Action JSON builder
│   ├── clarifier.py              # Clarification generator
│   ├── citation_enforcer.py      # Citation validator
│   ├── ui_formatter.py           # Response formatter
│   ├── process_pdf.py            # PDF → chunks
│   ├── index_chunks.py           # Chunks → FAISS index
│   ├── gradio_app.py             # Gradio UI (optional)
│   ├── docs_html.py              # Documentation HTML
│   ├── faq_index.faiss           # Pre-built vector index
│   ├── chunks_mapping.json       # Chunk metadata
│   ├── chunks.json               # Raw chunks
│   ├── requirements.txt
│   └── Dockerfile
│
└── frontend/
    ├── app/
    │   ├── page.tsx              # Login page
    │   ├── dashboard/page.tsx    # Main dashboard
    │   └── api/chat/route.ts     # Next.js API proxy
    ├── components/ui/            # shadcn/ui components
    ├── public/
    │   ├── avatar.png            # AI avatar
    │   └── docs.html             # Static docs page
    ├── .env.local                # Frontend env vars
    └── package.json

---

🔮 Roadmap

• Multi-document RAG (policy docs, IT manuals, FAQs)
• Persistent conversation history (PostgreSQL / Redis)
• User authentication with JWT
• Analytics dashboard (intent frequency, escalation rate)
• Slack / MS Teams integration
• Fine-tuned intent classifier on HCLTech domain data

---

📄 License

This project was built for the HCLTech NLP Challenge 2026.

---

Built with ❤️ by Shridipa Dasgupta