RAG Internal Document AI Assistant

I’m looking to turn our scattered know-how—hundreds of PDFs, policies, detailed reports, emails, meeting minutes, and assorted documentation—into a single Retrieval-Augmented Generation assistant that answers questions instantly and finds the exact source passage behind every reply.

What I already have in mind
• Ingestion pipeline that automatically pulls new or updated PDFs, email archives, and other internal files, cleans the text, and embeds it in a vector store (FAISS, Pinecone, or a similar tool).
• A lightweight search layer so staff can type a query and receive both a concise answer and the ranked source documents.
• Chat-style interface (web or Slack/Teams bot) backed by an LLM with citations, guardrails, and role-based access.
• Modular codebase so I can swap models or storage later without a rewrite.

What I need from you
1. End-to-end architecture diagram and tech stack recommendation.
2. Working prototype deployed in our cloud (AWS or Azure) with clear setup scripts.
3. Brief hand-over guide outlining how to add new document types or retrain embeddings.

Acceptance criteria
• The assistant must correctly return the top-3 source snippets for at least 80 % of test questions drawn from our reports, policies, and emails.
• Query response time under five seconds on a mid-range VM with ~10 k documents indexed.
• No sensitive data leaves the private network; any third-party API calls must be boxed inside our VPC.

If you’ve built RAG pipelines, semantic search, or document chatbots before, I’d love to see a short demo link and hear which vector DB-LLM combo you prefer.