Enterprise Knowledge Assistant

Production-style AI knowledge assistant built with Retrieval-Augmented Generation (RAG) for enterprise document search and question answering.

Live Demo

• Frontend: http://98.92.140.134
• API Docs: http://98.92.140.134/docs

Note: Demo may be usage-limited to control API costs.

---

Overview

This project ingests PDF documents, extracts and chunks text, generates embeddings with OpenAI, stores vectors in FAISS, and answers questions using retrieved context through a FastAPI backend and a React frontend.

It is designed as a production-style AI application, not a toy demo, with a clean backend architecture, modular services, metadata persistence, vector retrieval, Dockerized deployment, and a polished frontend experience.

---

Features

• PDF upload and ingestion pipeline
• Text extraction, cleaning, and chunking
• OpenAI embeddings
• FAISS vector search
• Retrieval-Augmented Generation (RAG)
• Source-aware answers with citations
• Query across all documents or a selected document
• SQLite metadata tracking with SQLAlchemy
• Duplicate document detection via content hashing
• Rebuildable vector index
• Document delete + reindex workflow
• FastAPI backend with modular architecture
• React + Vite + Tailwind frontend
• Dockerized frontend/backend deployment
• AWS EC2 deployment
• Structured logging and request tracing

---

Screenshots

Main Workspace

Query Result

Documents View

API Docs

---

Tech Stack

Backend

• Python 3.11
• FastAPI
• SQLAlchemy
• SQLite
• FAISS
• OpenAI API
• PyPDF

Frontend

• React
• Vite
• TypeScript
• Tailwind CSS
• Axios

DevOps / Deployment

• Docker
• Docker Compose
• AWS EC2
• GitHub Actions

---

Architecture

Client (React + Vite + Tailwind)
        │
        ▼
FastAPI Backend
  ├── Document Upload / Ingestion
  │     ├── PDF parsing
  │     ├── Text cleaning
  │     ├── Chunking
  │     ├── OpenAI embeddings
  │     └── FAISS indexing
  │
  └── Query / RAG Service
        ├── Query embedding
        ├── Vector retrieval
        ├── Prompt assembly
        └── OpenAI answer generation

Storage Layer
  ├── Raw PDF files
  ├── SQLite metadata DB
  └── FAISS vector index

Project Structure

app/
  api/
  core/
  db/
  embeddings/
  ingestion/
  llm/
  repositories/
  retrieval/
  schemas/
  services/
  storage/
  utils/
frontend/
data/
scripts/
tests/
docs/

Core API Endpoints

• GET /api/v1/health
• GET /api/v1/health/ready
• GET /api/v1/documents
• GET /api/v1/documents/{document_id}
• POST /api/v1/documents/upload
• DELETE /api/v1/documents/{document_id}
• POST /api/v1/documents/rebuild-index
• POST /api/v1/query

Example Query Request

{
  "question": "What is this document about?",
  "top_k": 5,
  "document_id": null
}

Example Query Response

{
  "answer": "The document discusses ...",
  "sources": [
    {
      "chunk_id": "chunk-id",
      "document_id": "doc-id",
      "filename": "sample.pdf",
      "page_number": 1,
      "chunk_index": 0,
      "text_preview": "..."
    }
  ]
}

Local Setup

1. Clone the repository

git clone https://github.com/oday198/enterprise-knowledge-assistant.git
cd enterprise-knowledge-assistant

2. Create environment

python -m venv .venv
source .venv/bin/activate

3. Install dependencies

pip install -e ".[dev]"

4. Configure environment

Copy .env.example to .env and set OPENAI_API_KEY

5. Run backend

uvicorn app.main:app --reload

6. Run frontend

cd frontend
npm install
npm run dev

---

Docker

docker compose up --build

• Frontend: http://localhost
• API Docs: http://localhost/docs

---

Deployment

Recommended deployment: AWS EC2 with Docker Compose

High-level deployment flow:

1. Launch Ubuntu EC2 instance
2. Install Docker + Docker Compose
3. Clone repository
4. Create .env from production template
5. Run docker compose up --build -d
6. Expose port 80

---

Testing

pytest -q

GitHub Actions CI is included for automated test runs on push/pull request.

---

Key Engineering Decisions

• Modular monolith architecture for simplicity and production clarity
• FAISS abstraction layer to keep vector store replaceable
• Metadata DB + vector index separation
• Duplicate detection by content hash
• Document-scoped querying in addition to global querying
• Dockerized deployment for reproducibility
• Frontend + backend separation for cleaner architecture

---

Future Improvements

• Background async ingestion
• Authentication / access control
• Rate limiting
• Managed vector DB option
• Cloud object storage for PDF files
• LLM response caching
• Better observability and metrics
• Reranking for retrieval quality

---

Author

Oday Soueidan LinkedIn: https://www.linkedin.com/in/odaysoueidan GitHub: https://github.com/oday198