Document Support RAG Chatbot

Overview

A public engineering project that explores a complete document-support workflow without presenting the application as a hosted service.

Problem / motivation

Document answers are more useful when they are grounded in relevant retrieved context and make their sources visible.

What it implements

• Accept text, Markdown and PDF documents through a browser interface and FastAPI API.
• Create embeddings with OpenAI and persist searchable document chunks in ChromaDB.
• Return grounded answers with structured source references and support local Docker, Docker Compose and Kubernetes execution.

Architecture / workflow

• Documents enter through the browser or FastAPI interface and are converted into searchable chunks.
• OpenAI embeddings represent those chunks and ChromaDB persists the local retrieval index.
• A question retrieves relevant context before the response is returned with visible source references.

Key engineering decisions

Reliability / security considerations

• Unsupported questions use a bounded fallback instead of presenting an ungrounded answer.
• Source excerpts remain visible so a user can inspect the evidence behind a response.
• Deployment is described as local or self-hosted; no public hosted service is implied.

Validation / testing

• pytest covers the Python implementation and Ruff enforces code-quality checks.
• Captured scenarios verify grounded answers, contextual follow-ups and bounded fallback behaviour.
• Docker, Compose and single-instance Kubernetes paths document repeatable local execution.