RAG Knowledge API

A full-stack Retrieval-Augmented Generation (RAG) application. Upload PDFs or web pages, then ask questions and get answers grounded in your documents — powered by OpenAI GPT-3.5-Turbo.

Project structure

rag-knowledge-api/
├── backend/                  # FastAPI application
│   ├── api/main.py           # REST endpoints
│   ├── rag/
│   │   ├── generate.py       # OpenAI GPT-3.5 answer generation
│   │   └── retrieve.py       # FAISS vector search
│   ├── ingest/
│   │   ├── loader.py         # PDF + URL loaders
│   │   ├── chunker.py        # Text chunking
│   │   └── embed.py          # Sentence-transformer embeddings
│   ├── requirements.txt
│   └── Dockerfile
├── frontend/                 # Angular 17 SPA
│   ├── src/app/
│   │   ├── pages/upload/     # Knowledge Base page (PDF + URL upload)
│   │   ├── pages/chat/       # Ask AI chat page
│   │   └── services/         # HTTP API service
│   ├── nginx.conf            # Reverse-proxy config (production)
│   └── Dockerfile
├── docker-compose.yml
├── .env.example
└── README.md

API endpoints

Method	Path	Description
GET	/health	Health check
POST	/upload/pdf	Index a PDF (multipart/form-data: user_id, file)
POST	/upload/url	Index a web page ({ user_id, url })
POST	/ask	Ask a question ({ user_id, question })

Each user's knowledge base is isolated by user_id.

Prerequisites

• An OpenAI API key
• Docker — for the production setup
• Python 3.11 + Node 20 — for local development

---

Option A — Docker (recommended)

cp .env.example .env
# Edit .env and set your OPENAI_API_KEY

docker compose up --build

The Angular UI is served at http://localhost. All /api/* requests are proxied to the backend by nginx automatically.

---

Option B — Local development

Backend (runs from the backend/ folder so Python imports resolve correctly)

cd backend
python -m venv venv
source venv/Scripts/activate   # Windows
# source venv/bin/activate     # macOS / Linux
pip install -r requirements.txt

uvicorn api.main:app --reload
# API:   http://localhost:8000
# Docs:  http://localhost:8000/docs

Frontend (proxies /api → http://localhost:8000)

cd frontend
npm install
npm start
# UI: http://localhost:4200

---

How it works

1. Ingest — Documents are loaded (PDF pages or scraped HTML), split into 500-word overlapping chunks, and embedded with all-MiniLM-L6-v2. Vectors are stored in a per-user FAISS index under backend/data/.
2. Retrieve — The question is embedded and the top-5 nearest chunks are fetched from the user's index.
3. Generate — Retrieved chunks are passed as context to GPT-3.5-Turbo with a system prompt that prevents answers outside the provided context.

Environment variables

Variable	Required	Description
OPENAI_API_KEY	Yes	Your OpenAI secret key (sk-...)