🤖 InnovateTech — AI Customer Support Bot

A production-grade, full-stack chatbot demonstrating robust session management, high-accuracy RAG (Retrieval-Augmented Generation), escalation simulation, conversation summarization, and deep LLM integration.

---

🎯 Project Summary

InnovateTech is a full-stack AI customer support bot designed as a professional proof-of-concept. It focuses on:

• ✅ Conversational accuracy using RAG (knowledge base grounding)
• ✅ Contextual memory via persistent sessions stored in PostgreSQL
• ✅ LLM-powered features: response generation, conversation summarization, and suggested next actions
• ✅ Escalation simulation for human handoffs
• ✅ A modern, deployable single-file frontend with a floating, expandable chat widget

This README contains an easy Quick Start, API examples, prompt guidance, deployment notes, and recommended next steps for production hardening.

---

🛠️ Tech Stack

Category	Technology & Libraries	Purpose
Backend	Python, FastAPI, Uvicorn	For creating a high-performance, modern REST API.
Database	PostgreSQL	Robust, relational data storage for sessions.
ORM	SQLModel	Type-safe, modern data interaction with the DB.
LLM Service	Google Gemini API	Powering the bot's conversational intelligence.
DevOps	Docker, Docker Compose	For containerizing the database and ensuring a reproducible environment.
Frontend	Vanilla HTML, CSS, JavaScript (Single File)	To create a lightweight, universally compatible, and stunning user interface.

---

🎥 Live Demo

[Watch the 3-Minute Video Demo Here](https://drive.google.com/file/d/10QHK-pePmT1Kgj5eMa1K2gekwcXTnrwb/view?usp=sharing)

---

📚 Table of contents

• Quick Start

  • Prerequisites
  • Clone & configure
  • Run with Docker Compose
  • Launch frontend

• How it works (high level)

• API Reference

• Knowledge Base & RAG

• Prompt Engineering

• Future Enhancements

• Production checklist & additional requirements

• Troubleshooting

• Credits & Contact

---

🚀 Quick Start

Prerequisites

• Docker & Docker Compose
• A Google AI / Gemini API Key (or other LLM provider key if you adapt the handler)
• Git (optional but recommended)

1. Clone & configure

git clone https://github.com/[YOUR-USERNAME]/ai-customer-support-bot.git
cd ai-customer-support-bot

2. Create a .env from the example and add your keys:

cp .env.example .env
# then open .env and set:
# GOOGLE_API_KEY="YOUR_SECRET_API_KEY_HERE"

🔐 Security note: Do not commit .env. Use a secrets manager for production.

3. Run with Docker Compose

Build and run backend + DB with a single command:

docker-compose up --build

4. Run the Backend Server

Set up a Python virtual environment and run the FastAPI server.

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn src.main:app --reload

• Backend: The API will now be running at - http://127.0.0.1:8000
• API docs (interactive): http://127.0.0.1:8000/docs

5. Launch the frontend

Navigate to the frontend/ directory and open the index.html file directly in your web browser. . Ensure the API_BASE variable in the file points to your backend host (http://127.0.0.1:8000 by default).

---

🏗️ Architecture Diagram

The system uses a clean, decoupled architecture for scalability and maintainability.

graph TD
  A["🌐 Frontend: Floating Chat Widget"] <--> B["🚀 FastAPI Backend API"]
  B <--> C["🧠 Google Gemini API\n(for Generation & Summarization)"]
  B <--> D["📦 PostgreSQL Database\n(in Docker Container)"]
  B -- RAG --> E["📚 knowledge_base.json"]

  subgraph "User's Browser (Client)"
    A
  end

  subgraph "Local Server (Host)"
    B
    D
    E
  end

  subgraph "Third-Party Cloud Service"
    C
  end

Loading

---

🔎 How it works (high level)

1. Frontend (single index.html) provides a floating, expandable chat widget. On first message the frontend calls POST /api/sessions/ to create a session and stores session_id in localStorage.

2. Backend (FastAPI) handles REST endpoints:

   • POST /api/sessions/ → create session
   • POST /api/sessions/{session_id}/messages/ → receive user message, save it, run RAG + LLM to produce bot reply, save bot message
   • POST /api/sessions/{session_id}/summarize → instruct LLM to summarize the session and save the result

3. Database (PostgreSQL) stores sessions and messages (sender, content, timestamp). The conversation history is used to provide contextual memory.

4. RAG pipeline scans knowledge_base.json for matching FAQ entries and injects those facts into the LLM prompt to guarantee grounded answers.

5. LLM handler (provider-agnostic) communicates with Google Gemini / OpenAI based on configuration and returns structured responses.

---

🧭 API Reference (examples)

All endpoints live under /api by default. Replace host/port if needed.

1. Create a session

POST /api/sessions/
Content-Type: application/json

# Response (201)
{ "id": 1 }

2. Post a message (gets bot response)

POST /api/sessions/1/messages/
Content-Type: application/json

{ "content": "Hi, how can I track my order?" }

# Response (200)
{
  "id": 2,
  "content": "You can track your order by visiting ...",
  "sender": "bot"
}

3. Summarize a session

POST /api/sessions/1/summarize
Content-Type: application/json

# Response (200)
{
  "session_id": 1,
  "summary": "User asked about returns and order tracking; bot provided FAQ-based answers and recommended escalation for account issues."
}

---

📚 Knowledge Base & RAG

knowledge_base.json contains an array of faqs objects with this structure:

{
  "question": "return policy",
  "keywords": ["return", "refund", "30-day"],
  "answer": "InnovateTech offers a 30-day, no-questions-asked money-back guarantee..."
}

RAG behaviour: The backend searches for FAQ entries whose keywords match the user query (or uses a semantic search if you upgrade). Matching answer text is placed in the prompt as context so the LLM grounds its answer.

---

🧠 Prompt Engineering (short guide)

Use carefully designed prompts to control persona and behavior.

RAG System Prompt (example)

SYSTEM:
You are a friendly and highly knowledgeable customer support assistant for InnovateTech. Use the following context verbatim if it is relevant.
--- CONTEXT FROM KNOWLEDGE BASE ---
{context_from_kb}
USER:
{user_message}

Summarization Prompt (example)

Please analyze the following customer support conversation and provide a concise, one-paragraph summary describing the issue and the final resolution or next steps.
--- CONVERSATION TRANSCRIPT ---
{transcript}

Tip: Keep system prompts short but explicit. Always instruct the model how to use or ignore the provided context.

---

🔮 Future Enhancements (recommended)

• Semantic RAG: Replace keyword matching with vector-based semantic search (Chroma, Pinecone, Weaviate).
• User Authentication: Add JWT-based auth to associate sessions with user accounts and pull personalized data (orders, subscriptions).
• Admin Dashboard: A manager UI to review conversations, edit the knowledge base, and re-run summaries.
• Async workers & queueing: Offload LLM calls to worker queues for better throughput and retries.
• Monitoring & observability: Add tracing (OpenTelemetry), structured logs, and usage/cost monitoring for LLM API calls.

---

🛡️ Production checklist & additional requirements

Infrastructure

• TLS / HTTPS (Let's Encrypt)
• Managed DB with backups (RDS / Cloud SQL)
• Secrets manager for API keys (Vault / AWS Secrets Manager / GitHub Secrets)
• Reverse proxy (Nginx / Traefik) for routing and protection

Application

• Rate limiting and caching for frequent queries
• PII redaction before sending to external LLMs
• Data retention & deletion endpoints for compliance (GDPR/CCPA)

Scaling

• Docker Compose for quick demos; Kubernetes for production scaling
• Use a vector DB for semantic search at scale
• Autoscaling and worker pools for LLM request handling

---

🩺 Troubleshooting

• 500 on LLM call: Verify your API key, billing, and LLM provider availability. Check backend logs (docker-compose logs -f backend).
• Frontend cannot reach API: Ensure API_BASE is correct and backend allows CORS.
• DB connection refused: Check DATABASE_URL and that Postgres is running (docker-compose ps).
• Session issues: Clear browser localStorage.chat_session_id to reset a session.

---

📁 Suggested .env.example

DATABASE_URL=postgresql://postgres:postgres@db:5432/innovate
LLM_PROVIDER=google_gemini
GOOGLE_API_KEY=REPLACE_ME
OPENAI_API_KEY=REPLACE_ME
DEBUG=true
PORT=8000

---

📦 Deployment recommendations

• For staging: Docker Compose on a small VM, fronted by Nginx/Traefik with TLS.
• For production: Kubernetes, managed DB, separate worker pool, vector DB for RAG, CI/CD pipeline, and secrets manager.

---

🙋‍♂️ Credits & Contact

• Author: Ishan Peshkar — [email protected]
• Repo: https://github.com/ishanpeshkar/Ai-Customer-Support-Chatbot
• License: MIT

---