# Secure Document Q&A

Compliance and policy archives are large, fragmented, and rarely searched well. The team that needs an answer in the next ten minutes settles for an answer in the next two days. We built a private RAG chatbot that lets staff ask plain-language questions over the archive, returns answers with source citations to the underlying PDFs, and keeps every byte inside the corporate perimeter.

## Key facts

- **Industry:** PUBLIC SECTOR & TOURISM
- **Client (abstracted):** A national tourism authority in the GCC
- **Timeline:** Production rollout
- **Role:** Internal RAG chatbot delivery
- **Year:** 2024

## The challenge

Public-sector and regulated organisations cannot ship document content to public LLM endpoints. The default chatbot architecture violates that constraint on day one.

Search alone is not enough. Staff need answers, not just hits. Answers must show their sources or the room will not trust them.

## The solution

Built the retrieval layer on ChromaDB and HuggingFace embeddings, kept entirely inside the customer environment. Generation runs against Gemini Pro with retrieval-augmented prompts, and every response surfaces the source documents and page numbers it drew from.

Shipped a Next.js 16 frontend with streaming responses, a ChatGPT-style conversation surface, document upload and management, and chat history. Operations staff get a familiar interface; security teams get a deployment that respects the perimeter.

## Outcomes

- A working internal Q&A surface for the corporate archive, with source citations on every answer.
- A pattern that ports to other regulated archives: legal, compliance, internal policy, regulator filings.
- A reference for Sovereign Agentic AI delivery in the GCC, where data residency is the brief, not an afterthought.

## Impact

- Local — Vector store and embeddings stay on-prem
- Cited — Every answer points to its source PDF and page
- Streaming — Real-time response generation
- Multi-doc — Conversation spans the full archive

## Tech stack

- Next.js 16
- FastAPI
- LangChain
- ChromaDB
- HuggingFace Embeddings
- Google Gemini Pro

## Related

- [All work](https://levent.ai/work/)
- [Next: Demand Forecasting and Inventory Optimiser](https://levent.ai/work/demand-forecasting/)

Client names and real outcome metrics are not published. See https://levent.ai/ai-content-policy/ for the abstraction policy used across this site.

---

**Canonical URL:** https://levent.ai/work/askive/