This chatbot allows employees to query HR policies using an AI-powered assistant. 🧠 HR Policy Q&A Assistant (RAG-powered)

A Retrieval-Augmented Generation (RAG) system that intelligently answers employee HR policy questions using PDF documents. Built with LangChain and Qdrant, the assistant retrieves, reranks, and generates accurate answers from company policy files.

Brief Description

The HR Policy Q&A Assistant is an AI-powered chatbot designed to instantly answer any queries related to an organization’s HR policies. It ensures quick, accurate, and context-aware responses without requiring manual intervention from HR personnel.

Problem Statement

In most organizations, employees often face delays when seeking clarification on HR policies, benefits, or procedures. Traditional support channels like emails or HR tickets are time-consuming, leading to inefficiencies and frustration.

What the Bot Solves for Employees and HR

This assistant eliminates the need to wait for HR responses by providing instant, reliable answers to employee queries. It empowers employees with self-service access to HR information while significantly reducing the repetitive workload on HR teams.

🔍 System Workflow

The HR Policy Q&A Assistant is built on a robust Retrieval-Augmented Generation (RAG) architecture that transforms static HR documents into an intelligent, conversational knowledge system.

A[📄 PDF Upload & Streaming] --> B[🧩 Semantic Chunking]
B --> C[🔢 Embedding Generation<br>(all-MiniLM-L6-v2)]
C --> D[🗃️ Qdrant Cloud Vector DB<br>(FLAT / HNSW / QUANTIZED)]
D --> E[🎯 Dense Retrieval]
E --> F[📚 BM25 Reranking]
F --> G[🤖 LLM Response Generation]
G --> H[💬 Conversational Memory]
H --> I[📄 DOCX / Text Output]

⚙️ Pipeline Breakdown

1. PDF Streaming & Ingestion

HR policy PDFs are dynamically streamed into the system, enabling incremental ingestion and continuous updates without downtime.

2. Semantic Chunking

Documents are broken into meaningful, context-aware chunks, preserving relationships between ideas instead of arbitrary splits.

3. Embedding Generation

Each chunk is embedded using Hugging Face’s all-MiniLM-L6-v2 — a lightweight yet high-performing model optimized for semantic similarity.

4. Vector Storage

The embeddings are stored in Qdrant Cloud, indexed under three configurations:

⚡ FLAT – For precision and baseline accuracy

🧭 HNSW – For high-speed approximate nearest neighbor search

💾 Quantized – For efficient memory usage

5. Dense Retrieval

User queries are embedded and compared against stored vectors to fetch the most relevant information — enabling contextually deep understanding rather than shallow keyword matches.

6. BM25 Reranking

The top retrieved chunks are reranked with BM25, combining both semantic and lexical relevance for balanced, high-precision results.

7. LLM Response Generation

The refined chunks are passed to an LLM, which generates concise, accurate, and human-like answers tailored to HR-related queries.

8. Conversational Memory

A memory layer maintains context across multiple turns — allowing employees and HR to have a natural, flowing chat experience.

9. Output Rendering

The final answer is displayed in the chat and can be exported as a formatted DOCX report for record-keeping or official use.

🧪 Retriever Benchmarking Results

Multiple retrieval methods — Dense, Sparse, and Hybrid — were tested extensively. After quantitative evaluation across accuracy, latency, and semantic coverage, the Dense Retriever emerged as the best-performing approach, offering both speed and contextual depth in HR-specific Q&A tasks.

├── chunking/         # Semantic chunking logic
├── data/             # HR policy PDFs
├── embedding/        # Embedding models
├── Final/            # Final runnable scripts
├── ingest/           # Incremental ingestion pipeline
├── interface/        # CLI / frontend setup (in progress)
├── llm/              # LLM interaction & prompt templates
├── Prompt/           # Prompt customization
├── render/           # DOCX response renderer
├── Reranker/         # BM25/MMR reranking
├── retrieval/        # Retriever logic (Qdrant)
├── Tracing/          # LangSmith/OpenTelemetry (observability)
├── utils/            # Common utilities (logging, config, etc.)
├── vectorstore/      # Qdrant index handling

• 📥 Incremental PDF ingestion
• ✂️ Semantic chunking + embedding
• 🧠 Multi-index vector store (Flat, HNSW, IVF) using Qdrant
• ⚖️ BM25/MMR-based reranking for relevance
• 💬 LLM-based direct answer generation
• 🧾 DOCX rendering of answers
• 🧠 Prompt templating support
• 📡 LangSmith integration
• 🧠 Multi-turn memory (WIP)
• 🌐 Streamlit interface (planned)
• 🐳 Deployed in Huggingface spaces

1. Ingest HR PDFs and split them into semantically meaningful chunks
2. Embed the chunks using OpenAI or HuggingFace models
3. Store them in Qdrant with efficient vector indexing
4. Retrieve top-k documents using similarity search
5. Rerank results using BM25 or MMR
6. Use LLM with templated prompt to generate final response
7. Export response to DOCX

# Step 1: Install dependencies
pip install -r requirements.txt

# Step 2: Run CLI
python app.py --query "Is my spouse covered under the company health insurance?"

Q: "How many casual leaves do employees get per year?" A: "Yes, your legal spouse is eligible for coverage under our medical, dental, and vision plans. You will need to provide documentation to verify their eligibility."

• LangChain
• Qdrant
• OpenAI / Ollama / HuggingFace
• BM25 / MMR
• LangSmith
• Python

• ✅ Streamlit / Gradio UI
• ✅ Redis/SQLite-based chat memory
• ✅ Docker + cloud deployment DOCKER COMMAND - {"http://localhost:8501/",docker run --env-file .env -p 8501:8501 hr-assistant-bot:latest}
• ✅ Slack/MS Teams integration

Jayandhan S — Passionate about building agentic GenAI systems and real-world AI assistants.

MIT License