Offline-Support-ChatBot.

# Offline Customer Support Chatbot (Ollama + Llama 3.2)

## Overview

This project demonstrates how to build an **offline AI-powered customer support chatbot** using **Ollama** and the **Llama 3.2 (3B)** language model. The chatbot answers common e-commerce support queries such as order tracking, returns, payments, and account management.

Unlike cloud-based AI services, this system runs entirely **locally**, ensuring that no customer data leaves the machine. This approach improves **privacy, security, and cost efficiency** by eliminating external API usage.

The project also evaluates two prompt engineering techniques:

* **Zero-Shot Prompting**
* **One-Shot Prompting**

The responses generated by the model are logged and manually evaluated based on relevance, coherence, and helpfulness.

---

## Project Structure

```
offline-support-chatbot
│
├── prompts
│   ├── zero_shot_template.txt
│   └── one_shot_template.txt
│
├── eval
│   └── results.md
│
├── chatbot.js
├── setup.md
├── report.md
└── README.md
```

### Description of Components

**chatbot.js**
Node.js script that sends prompts to the Ollama API and retrieves responses from the Llama 3.2 model.

**prompts/**
Contains the prompt templates used for zero-shot and one-shot prompting.

**eval/results.md**
Stores generated responses and manual evaluation scores.

**report.md**
Detailed analysis comparing the performance of different prompting strategies.

**setup.md**
Instructions for installing dependencies and running the project.

---

## Features

* Fully **offline AI chatbot**
* Uses **Ollama local LLM runtime**
* Powered by **Llama 3.2 (3B)**
* Implements **prompt engineering experiments**
* Evaluates responses using **manual scoring metrics**
* Generates structured evaluation logs in Markdown

---

## Technologies Used

* **Node.js**
* **Ollama**
* **Llama 3.2**
* **Axios (HTTP requests)**
* **Markdown for evaluation logs**

---

## Example Customer Queries

The chatbot processes realistic e-commerce support questions such as:

* How can I track the shipping status of my order?
* My discount code is not working at checkout.
* I received a damaged product. How can I request a replacement?
* How can I reset my account password?

Each query is processed twice using different prompt strategies.

---

## Evaluation Metrics

Responses are evaluated using the following criteria:

**Relevance (1–5)**
How well the response addresses the customer query.

**Coherence (1–5)**
Clarity, grammar, and readability.

**Helpfulness (1–5)**
Whether the response provides actionable guidance.

---

## Results

The experiment showed that **one-shot prompting generally produces more structured and helpful responses compared to zero-shot prompting**.

Full evaluation logs are available in:

```
eval/results.md
```

---

## Conclusion

This project demonstrates the feasibility of building **offline AI customer support systems** using locally deployed language models. By combining Ollama with prompt engineering techniques, developers can create privacy-preserving AI tools without relying on external APIs.

Future improvements could include integrating product databases, retrieval-based knowledge systems, and a web-based user interface.