🇧🇩 Bangla FAQ Bot using RAG (Retrieval-Augmented Generation)

---

🚀 Project Overview

This project demonstrates a Bangla-language FAQ assistant built using Retrieval-Augmented Generation (RAG) inside a Google Colab notebook.

The system retrieves relevant Bangla FAQ content using semantic embeddings + FAISS, applies metadata-based filtering, and generates grounded responses using an LLM — preventing hallucinations and improving answer relevance for a low-resource language.

🎯 Primary Goal: Build a proof-of-concept RAG system that supports Bangla queries with topic-aware routing and safe fallback behavior.

---

🌍 Why This Project Matters

• Most RAG examples focus on English

• Bangla is a low-resource language in AI

• Hallucination is risky in FAQ and helpdesk systems

• This project shows how to ground LLMs using retrieval + metadata

Applicable Use Cases

• Bangla customer support bots

• Educational assistants

• Government or NGO information systems

• Domain-specific knowledge assistants

---

🧩 Key Features

✅ Bangla input & output support

🧠 LLM-based topic routing

🗂️ Metadata filtering before vector search

🔎 FAISS-powered semantic similarity search

🧱 Retrieval-Augmented Generation (RAG)

🚫 Safe fallback when no relevant context is found

💻 Fully implemented inside a single Colab notebook

---

🏗️ System Architecture

User Question (Bangla)
        |
        v
[ LLM Category Router ]
        |
        v
[ Metadata-Based Filtering ]
        |
        v
[ FAISS Similarity Search ]
        |
        v
[ RAG Answer Generation ]

Design Goal: Ensure the LLM never answers beyond retrieved Bangla knowledge.

---

🛠️ Tech Stack

Component	Technology
Language	Python
Notebook	Google Colab
Embeddings	Bengali SBERT (l3cube-pune)
Vector Store	FAISS
RAG Framework	LangChain
LLM	OpenAI GPT-4.1-Nano (GitHub Inference API)

---

⚙️ How to Run (Google Colab)

1. Open the notebook:

notebooks/bangla_faq_rag.ipynb

2. Set API token in Colab:

os.environ["GITHUB_TOKEN"] = "your_api_key"

3. Run cells sequentially:

• Install dependencies

• Load and embed Bangla FAQ data

• Build FAISS vector store

• Ask questions interactively

---

▶️ Example Interaction

👩🏻‍💻 আপনার প্রশ্নটি বলুন: পরিবারের সাথে ভ্রমণের সুবিধা কী?
[LLM Router] Category: ভ্রমণ

[User Question] পরিবারের সাথে ভ্রমণের সুবিধা কী?

[Metadata Filtering] Category: ভ্রমণ

🕵🏻 [Similarity Search Results for Query] পরিবারের সাথে ভ্রমণের সুবিধা কী?
 >> পরিবারের সাথে ভ্রমণ আনন্দ বাড়ায়।
 >> ভ্রমণ মানুষকে নতুন অভিজ্ঞতা দেয়।
 >> ভ্রমণ জীবনের একঘেয়েমি দূর করে।

💡প্রত্যাশিত উত্তর: পরিবারের সাথে ভ্রমণের সুবিধা হলো এটি আনন্দ বাড়ায়, নতুন অভিজ্ঞতা দেয় এবং জীবনের একঘেয়েমি দূর করে।

If no relevant context exists:

দুঃখিত, এই বিষয়ে আমার উত্তরটি জানা নেই।

---

📊 Demo & Output

🎥 Demo Video:

👉 https://drive.google.com/file/d/1esH30Cl80EC0Y_Rs6qH8HvHFda2DvjhN/view

📌 Output Characteristics:

• Grounded answers

• Topic-aware retrieval

• No hallucinated responses

---

🧠 Learning & Engineering Highlights

• Implemented Retrieval-Augmented Generation from scratch

• Used metadata-aware vector filtering

• Applied LLM-based reasoning for intent classification

• Worked with low-resource Bangla embeddings

• Designed a safe, explainable AI assistant

---

🔮 Future Improvements

• Convert notebook into FastAPI service

• Add automatic evaluation metrics (Recall@K)

• Persist FAISS index to disk

• Add document ingestion from PDFs / web

• Deploy as a lightweight web app

---