GPBot

A bilingual (English + Bangla) FAQ chatbot for Grameenphone customer service, powered by a BiLSTM intent classifier and BM25 retrieval.

Live: gpbot.fahimabrar.com

Architecture

User Query (English / Bangla)
        |
        v
+-------------------+
|   Tokenizer       |  Lowercase, remove punctuation, split into words
+-------------------+
        |
        v
+-------------------+
|   BiLSTM Intent   |  Embedding -> Bidirectional LSTM -> Linear
|   Classifier      |  Predicts 1 of 14 intent categories
+-------------------+
        |
        v
  confidence >= 50%?
   /            \
  yes            no
  |               |
  v               v
+----------+   "Sorry, I couldn't
| BM25 FAQ |    understand..."
| Retrieval|
+----------+
  |
  v
Best matching FAQ answer

Two-Stage Pipeline

1. Intent Classification - A bidirectional LSTM neural network classifies the user query into one of 14 intents (e.g., balance_check, recharge, data_package, sim_replace). Built with PyTorch.

2. FAQ Retrieval - The FAQ knowledge base is filtered by predicted intent, then BM25Okapi ranks questions by relevance to return the best answer.

Supported Intents

Intent	Description
balance_check	Prepaid/postpaid balance inquiries
recharge	Top-up and recharge methods
data_package	Internet/data package info
call_rate	Call rate inquiries
sms_package	SMS bundle info
network_issue	Signal and connectivity problems
sim_replace	SIM replacement, 4G upgrade, eSIM
myGP_app	MyGP app usage and troubleshooting
bill_payment	Postpaid bill payment
roaming	International roaming services
account_info	Account details, number check
customer_support	Helpline, complaints
offer_promotion	Current offers and promotions
number_portability	MNP - switching operators

Project Structure

.
├── app.py              # Streamlit web UI
├── model.py            # BiLSTM model definition and text processing
├── train.py            # Training script
├── retriever.py        # BM25-based FAQ retrieval
├── data/
│   ├── intent_data.json    # ~1920 labeled training examples
│   ├── faq_kb.json         # ~250 FAQ question-answer pairs
│   └── intents.json        # Intent metadata
├── models/
│   ├── best_model.pt       # Trained model weights
│   ├── config.json         # Model hyperparameters
│   ├── word2idx.json       # Vocabulary mapping
│   └── idx2label.json      # Label mapping
├── Dockerfile
├── pyproject.toml
└── uv.lock

Setup

Prerequisites

• Python >= 3.12
• uv package manager

Install Dependencies

uv sync

Train the Model

uv run python train.py

Run the App

uv run streamlit run app.py

Open http://localhost:8501 in your browser.

Docker

Build

docker build -t gpbot .

Run

docker run -p 8501:8501 gpbot

Pull Pre-built Image

# Automatically picks your architecture (amd64 or arm64)
docker pull ghcr.io/syedfahimabrar/gpbot:latest

# Or specify platform explicitly
docker pull --platform linux/amd64 ghcr.io/syedfahimabrar/gpbot:latest
docker pull --platform linux/arm64 ghcr.io/syedfahimabrar/gpbot:latest

# Run
docker run -p 8501:8501 ghcr.io/syedfahimabrar/gpbot:latest

Testing

See test_questions.md for 42 sample queries across all 14 intents in both English and Bangla.

Tech Stack

Component	Technology
ML Model	PyTorch (BiLSTM)
Retrieval	rank-bm25 (BM25Okapi)
Frontend	Streamlit
Package Manager	uv
Container	Docker (multi-arch amd64 + arm64)