Regulatory Knowledge Assistant (RKA)

A domain-specific Retrieval-Augmented Generation (RAG) system designed to answer regulatory and policy-related queries using official documents such as Indian labour laws, RBI banking guidelines, and HR policy documents, with web search fallback for real-time information.

Live App: https://regulatory-knowledge-assistant-rka-q2q7paevuqr2lqvakz7hx9.streamlit.app/

Live Demo: https://drive.google.com/file/d/1CsFoiKaPk6VhNLedKjsKfVRoCtR8H5f4/view?usp=sharing

---

Project Name

Regulatory Knowledge Assistant (RKA)

---

Why I Chose This Project

NeoStats works with organizations that often need accurate answers from regulatory documents such as labour laws, banking guidelines, and HR policies. These documents are typically lengthy and complex, making it difficult for users to quickly retrieve relevant information.

To address this challenge, I designed a domain-specific Retrieval-Augmented Generation (RAG) application that can retrieve relevant knowledge from regulatory documents and provide precise answers.

The core idea behind the project was:

• Reuse publicly available regulatory data and make it useful for clients.
• Focus on documents relevant to NeoStats’ potential client ecosystem, such as banking regulations and HR policies.
• Build a system that can answer questions such as:

Example query:

What is the gratuity rule after 5 years?

To achieve this, the system:

1. Uses publicly available documents such as:

   • Indian Labour Law Handbook
   • RBI Guidelines
   • Payment of Gratuity Act
   • HR Policy documents

2. Converts these documents into embeddings and stores them in a vector database.

3. Retrieves relevant context when a user asks a question.

4. Sends the retrieved context to a Large Language Model (LLM) for generating accurate responses.

5. Falls back to web search if the knowledge is not present in the document corpus.

6. Provides a simple web interface using Streamlit for user interaction.

This approach demonstrates how NeoStats could build intelligent knowledge assistants for regulatory or compliance-heavy industries, enabling faster access to important information.

The complete solution — from development to deployment — was designed to be implemented within a short development cycle.

---

Basic Architecture of the Model

The system follows a Retrieval-Augmented Generation (RAG) architecture.

User Question
      │
      ▼
Embedding Query
      │
      ▼
Vector Database Search
      │
      ├── Relevant documents found → Send to LLM
      │
      └── No relevant documents
              │
              ▼
         Web Search (Tavily)
              │
              ▼
         Send results to LLM
              │
              ▼
           Final Answer

This architecture ensures that:

• The system answers using trusted regulatory documents whenever possible
• If knowledge is missing, real-time web search provides updated information

---

Overall Configuration (Project Structure)

The system is organized into modular components for scalability and maintainability.

NeoStats AI Engineer Use Case
│
└── AI_UseCase
    │
    ├── config
    │     └── config.py
    │
    ├── data
    │     ├── gratuity_act.pdf
    │     ├── hr_policy.pdf
    │     ├── labour_law.pdf
    │     └── rbi_guidelines.pdf
    │
    ├── models
    │     ├── embeddings.py
    │     └── llm.py
    │
    ├── utils
    │     ├── document_loader.py
    │     ├── rag_pipeline.py
    │     └── web_search.py
    │
    ├── app.py
    ├── requirements.txt
    └── NeoStats AI Engineer Case Study.pdf

Module Explanation

config/

• Stores environment configuration and API keys.

data/

• Contains regulatory documents used for knowledge retrieval.

models/

• Contains embedding models and LLM configuration.

utils/

• Handles document loading, RAG pipeline logic, and web search functionality.

app.py

• Streamlit application that provides the user interface.

---

Environment Requirements

The system requires the following Python packages:

streamlit
langchain-openai
langchain-groq
langchain-google-genai
langchain-core
langchain-community
sentence-transformers
faiss-cpu
pypdf
tavily-python
python-dotenv
groq

These can be installed using:

pip install -r requirements.txt

---

How to Run the Project Locally

1. Clone the repository

git clone https://github.com/Divyaanshvats/Regulatory-Knowledge-Assistant-RKA.git

Navigate into the project:

cd Regulatory-Knowledge-Assistant-RKA

---

2. Install dependencies

pip install -r requirements.txt

---

3. Set API Keys

Set environment variables for the required APIs.

Example (Windows):

setx GROQ_API_KEY "your_key_here"
setx TAVILY_API_KEY "your_key_here"

Restart the terminal after setting environment variables.

---

4. Run the Streamlit Application

streamlit run app.py

Open the browser at:

http://localhost:8501

---

Features Implemented

• Retrieval-Augmented Generation (RAG)
• Vector database using FAISS
• SentenceTransformer embeddings
• LLM integration using Groq
• Web search fallback using Tavily
• Streamlit-based web interface
• Domain-specific regulatory knowledge base

---

Example Queries

What is the gratuity rule after 5 years?

What deductions are allowed from wages?

Can banks give loans to their directors?

What is the latest RBI repo rate?

---

Deployment

The application is deployed using Streamlit Cloud.

Live Deployment:

https://regulatory-knowledge-assistant-rka-q2q7paevuqr2lqvakz7hx9.streamlit.app/

---

License

MIT License