Nimbus RAG Forge

Enterprise-grade RAG (Retrieval-Augmented Generation) application built on Azure. Upload documents, ingest and index them into Azure AI Search, then chat with your data through a conversational AI that answers strictly from your sources — with citations.

Based on Microsoft's Chat With Your Data Solution Accelerator, customized for production use.

Architecture

Four deployable components:

Component	Stack	Role
Web App	React (Vite/TypeScript) + Flask (Python)	Chat UI and conversation API
Admin App	Streamlit (Python)	Data ingestion, exploration, deletion, prompt config
Backend	Azure Functions (Python)	Document processing pipeline, embedding generation
Infrastructure	Azure Bicep (IaC)	Full Azure resource provisioning via azd

Users → Web App (Flask + React) → Azure OpenAI (GPT-3.5/4)
                                 → Azure AI Search (hybrid vector + keyword)
                                 → Azure Blob Storage (raw documents)

Admin → Streamlit App → Azure Functions → Document Intelligence (OCR)
                                        → Azure AI Search (indexing)
                                        → Azure Blob Storage

Features

• Multi-format ingestion — PDF, DOCX, TXT, HTML, Markdown, JPG, PNG, URLs
• Multiple chunking strategies — Layout-based, page-based, paragraph-based, fixed-size-overlap
• Hybrid search — Vector + keyword search with optional semantic ranking via Azure AI Search
• Citation-grounded answers — Every response traces back to source documents with [docN] references
• Three orchestration strategies (pluggable via env var):
  • openai_function — OpenAI function calling
  • semantic_kernel — Microsoft Semantic Kernel
  • langchain — LangChain agent
• Two conversation flows:
  • custom — Full RAG pipeline (chunking → embedding → retrieval → generation)
  • byod — Azure OpenAI "on your data" API with Azure Search as data source
• Integrated vectorization — Optional Azure AI Search integrated vectorization (indexer + skillset)
• Content safety — Azure AI Content Safety for filtering harmful queries/responses
• Post-answer fact-check — Optional validation that the generated answer aligns with sources
• Speech-to-text — Azure Speech Services for voice input
• Conversation logging — Interactions and token usage logged to a dedicated search index
• Observability — Azure Application Insights + OpenTelemetry
• Config-driven — Prompts, chunking, processors, orchestration strategy all defined in JSON config

Tech Stack

Backend: Python 3.10+, Flask, Streamlit, Azure OpenAI SDK, Azure AI Search, Azure Form Recognizer, LangChain, Semantic Kernel, OpenTelemetry

Frontend: React 18, Vite, TypeScript, Fluent UI, React Markdown, Azure Speech SDK

Infrastructure: Azure Bicep, Azure Developer CLI (azd), Docker Compose, App Service, Azure Functions

Testing: pytest, Cypress (E2E), Vitest (frontend), pre-commit hooks (flake8 + black)

Project Structure

code/
├── app.py, create_app.py        # Flask web app (chat API + static files)
├── frontend/                    # React chat UI (Vite + TypeScript)
├── backend/
│   ├── Admin.py                 # Streamlit admin app
│   ├── pages/                   # Streamlit pages (ingest, explore, delete, config)
│   └── batch/                   # Azure Functions (document processing)
│       └── utilities/           # Core: chunking, loading, embedding, search, orchestration
├── tests/                       # Unit + functional tests
infra/                           # Azure Bicep IaC
docker/                          # Dockerfiles + docker-compose
data/                            # Sample documents for testing
docs/                            # Documentation and ADRs

Getting Started

Local (Docker Compose)

cp .env.sample .env   # Fill in Azure resource values
make docker-compose-up

Services: web on :8080, admin on :8081, backend on :8082.

Local Development

poetry install
cd code/frontend && npm install
make build-frontend
make unittest

Deploy to Azure

azd auth login
azd up

Provisions all Azure resources (App Service, Functions, AI Search, OpenAI, Key Vault, Storage, etc.) and deploys all three services.

Admin Panel

The Streamlit admin app provides four pages:

1. Ingest Data — Upload and process documents into the search index
2. Explore Data — Inspect how documents were chunked and indexed
3. Delete Data — Remove indexed documents
4. Configuration — Adjust prompts, logging, and orchestration settings

License

All rights reserved.

Author

Cherif Benham