DocsChat - RAG Document Assistant

The document Q&A application using RAG. Upload documents, ask questions, get answers backed from the uploaded document content.

Implementation Note

Built without LangChain or similar frameworks - all RAG components implemented from scratch using direct API calls to demonstrate understanding of core concepts:

• Custom document chunking logic
• Direct integration with Sentence Transformers for embeddings
• Manual vector similarity search with pgvector
• Direct LLM API calls (OpenAI, Ollama)
• Custom streaming response handling

Features

• Switch between OpenAI (gpt-4o-mini) and Ollama (qwen2.5:7b) models
• Upload/delete documents (PDF, DOCX, TXT)
• Ask questions about uploaded documents
• Clear chat history
• Light/dark theme toggle
• Source citations for AI responses

Tech Stack

Backend:

• FastAPI
• PostgreSQL + pgvector
• Sentence Transformers (embeddings)
• SQLAlchemy

Frontend:

• Vanilla JavaScript
• Server-Sent Events (SSE streaming responses)

AI:

• Ollama (local LLM)
• OpenAI API (optional)

Setup

Prerequisites:

• Docker & Docker Compose
• 8GB RAM (for Ollama)
• OpenAI API key

Install:

# Clone repo
git clone https://github.com/yourusername/rag-chatbot.git
cd rag-chatbot

# Create .env file
cp .env.example .env

# Edit .env and add your OpenAI API key

.env configuration:

POSTGRES_USER=postgres
POSTGRES_PASSWORD=postgres
POSTGRES_DB=doc_rag_db
POSTGRES_HOST=db
POSTGRES_PORT=5432
OPENAI_MODEL=gpt-4o-mini
OPENAI_API_KEY=your-openai-api-key-here

Start services:

docker-compose build
docker-compose up

Access:

• UI: http://localhost:8000
• API docs: http://localhost:8000/docs

Note: First startup takes 10-15 minutes while Ollama downloads the model (~2.8GB)

Usage

1. Go to Documents tab
2. Upload your files
3. Switch to Chat tab
4. Ask questions
5. View answers with source citations

Development

View logs:

docker-compose logs -f api

Access database:

docker-compose exec db psql -U postgres -d doc_rag_db

How It Works

1. Documents split into chunks (~1200 chars, 150 char overlap)
2. Each chunk converted to 384-dimensional vector
3. Vectors stored in PostgreSQL with pgvector
4. User question embedded and similar chunks retrieved
5. LLM generates answer using retrieved chunks
6. Response streamed back in real-time