Doc-Bot — RAG-Powered Document Chatbot

A production-deployed AI assistant that lets you upload PDFs and chat with them using a full Retrieval-Augmented Generation (RAG) pipeline. Built end-to-end as an 8-week ML/AI Engineer portfolio project.

Live: http://18.227.122.170:8000

---

Architecture

User Browser
     │
     │  HTTP
     ▼
FastAPI (EC2 t3.micro, us-east-2)
     │
     ├── /upload ──► Local disk + S3 (doc-bot-pdfs-ahmadissa)
     │
     └── /chat
           │
           ├── ChromaDB (vector store, MMR retrieval)
           │       └── nomic-embed-text embeddings (Ollama)
           │
           └── llama3.2 LLM (Ollama)
                   └── Strict document-only prompt

Deployment pipeline:

git push → GitHub Actions → SSH into EC2 → git pull → docker build → docker run

---

Tech Stack

Layer	Tool
LLM	Llama 3.2 via Ollama (runs locally, free)
Embeddings	nomic-embed-text via Ollama
Vector Store	ChromaDB
Retrieval	MMR (Maximal Marginal Relevance, k=6, fetch_k=20)
Framework	LangChain
API	FastAPI + Pydantic
UI	Embedded dark-theme HTML/CSS/JS (mobile-responsive)
Container	Docker (multi-stage build)
Cloud	AWS EC2 + S3 + IAM + Elastic IP
CI/CD	GitHub Actions (auto-deploy on every push)
ML Tracking	MLflow (14 experiment runs, SQLite backend)

---

Features

• Upload PDFs via drag & drop or file picker
• Ask questions in plain English — answers cite exact page numbers
• Session-based conversation memory (last 3 turns)
• Strict document-only answers — refuses to answer from outside knowledge
• Every uploaded PDF backed up to AWS S3
• /metrics endpoint tracks questions, response times, errors
• API key authentication on all write endpoints
• Auto-deploys on every GitHub push via CI/CD

---

Project Structure

doc-bot/
├── main.py                          # FastAPI app — REST endpoints + embedded chat UI
├── rag.py                           # RAG logic — load, index, retrieve, answer
├── chatbot.py                       # CLI version of the chatbot
├── train.py                         # MLflow experiment tracking (14 runs)
├── Dockerfile                       # Multi-stage Docker build
├── docker-compose.yml               # Local development setup
├── requirements.txt
├── .github/workflows/deploy.yml     # GitHub Actions CI/CD pipeline
├── docs/                            # PDF storage (gitignored)
├── chroma_db/                       # Vector store (gitignored)
└── mlflow.db                        # MLflow SQLite backend (gitignored)

---

API Endpoints

Method	Endpoint	Auth	Description
GET	/	No	Chat UI
POST	/upload	Yes	Upload and index a PDF
POST	/chat	Yes	Ask a question
GET	/documents	No	List indexed documents
DELETE	/session/{id}	No	Clear conversation history
GET	/metrics	No	Usage stats
GET	/health	No	Health check
GET	/docs	No	Swagger API docs

Authentication: Pass X-API-Key: docbot-secret-123 header on upload and chat requests.

---

Local Setup

1. Install Ollama and pull models

# Download from https://ollama.ai
ollama pull llama3.2
ollama pull nomic-embed-text

2. Clone and install dependencies

git clone https://github.com/issa89ai/doc-bot.git
cd doc-bot
pip install -r requirements.txt

3. Run the API

python -m uvicorn main:app --reload

4. Open the chat UI

Go to http://localhost:8000

Docker (alternative)

docker-compose up

---

MLflow Experiment Tracking

Trained a 4-class topic classifier on the 20 Newsgroups dataset (sci.med, sci.space, rec.sport.hockey, talk.politics.guns) with 14 tracked runs:

Model	Accuracy
NaiveBayes (alpha=0.5)	91.26% — best
LinearSVC bigrams	90.8%
LogisticRegression bigrams	90.4%
LinearSVC	89.9%
LogisticRegression C=1.0	89.1%
LogisticRegression C=0.1	85.3%

python train.py
mlflow ui --backend-store-uri sqlite:///mlflow.db
# Open http://localhost:5000

---

AWS Infrastructure

Resource	Details
EC2	t3.micro, Amazon Linux 2023, us-east-2
Elastic IP	18.227.122.170 (permanent)
S3 Bucket	doc-bot-pdfs-ahmadissa (us-east-2)
IAM User	doc-bot-app (AmazonS3FullAccess)
Security Group	Ports 22 (SSH), 8000 (app)

---

Weekly Progress

Week 1 — First RAG App

• PDF loading with LangChain PyPDFLoader
• Text splitting (1000 chars / 100 overlap)
• ChromaDB vector store + nomic-embed-text embeddings
• Deliverable: CLI chatbot (chatbot.py)

Week 2 — Vector DB & RAG Deep Dive

• Multi-document support
• MMR retrieval (fetch 20, return 6 diverse results)
• Conversation history (last 3 turns)
• Source citations with filename + page number
• Deliverable: Multi-doc CLI chatbot with memory

Week 3 — REST API & Chat UI

• FastAPI backend with 7 endpoints
• Session-based conversation history
• Embedded dark-theme chat UI with drag & drop
• API key authentication
• Deliverable: localhost:8000 chat interface

Week 4 — Docker

• Multi-stage Dockerfile (builder → slim runtime)
• docker-compose with volume mounts
• OLLAMA_HOST env var for Docker → host Ollama bridge
• Deliverable: docker-compose up spins up everything

Week 5 — MLOps & Experiment Tracking

• 14 MLflow runs across 3 model families
• NaiveBayes alpha hyperparameter sweep (8 runs)
• Best: NaiveBayes alpha=0.5 → 91.26% accuracy
• Deliverable: MLflow dashboard with metric charts

Week 6 — Cloud Deployment (AWS)

• S3 bucket for PDF backup on every upload
• EC2 t3.micro deployment with Docker
• Deliverable: App live at http://18.227.122.170:8000

Week 7 — CI/CD Pipeline

• GitHub Actions workflow: push → SSH → rebuild → redeploy
• Elastic IP for permanent fixed address
• Docker --restart always for auto-start on boot
• Deliverable: Push code → deployed in ~15 seconds automatically

Week 8 — Monitoring & Polish

• Strict document-only prompt (no hallucination)
• Request logging to requests.log
• /metrics endpoint: total questions, avg response time, errors
• Deliverable: Production-hardened, monitored, polished repo

---

Skills Demonstrated

Skill	Where
Python	Throughout
LangChain / RAG / LLMs	Weeks 1–2, rag.py
Vector databases (ChromaDB)	Weeks 1–2
FastAPI / REST APIs	Week 3, main.py
Docker	Week 4, Dockerfile
Scikit-learn / MLflow	Week 5, train.py
AWS (EC2, S3, IAM)	Week 6
CI/CD (GitHub Actions)	Week 7
Monitoring / Observability	Week 8