Jancarlo Rojas

A Help Bot That Only Knows the Docs

Personal project · September 2026

A personal project: a chat bot in Microsoft Teams that answers IT questions from FSU IT documentation, and only from that documentation. Ask it something in plain language and it finds the relevant steps and writes them out. This project was built after working on a team, which was developing this exact thing. I wanted a better understanding of how it works, so I built it out myself.

How it works

The bot uses retrieval-augmented generation, or RAG. The documentation, in PDF, Word and Markdown files, is stored and indexed for search. When someone asks a question, the bot first searches for the passages most relevant to it, then gives only those passages to the model along with the question. The model writes its answer from what was retrieved instead of from memory.

Staying grounded

The model is told to rely only on the documentation. If the search turns up nothing relevant, the bot says so instead of guessing. When the documentation describes more than one way to do something, it asks a quick clarifying question before committing to an answer.

The pieces

The chat model is a small GPT nano model hosted in Microsoft Foundry. Azure AI Search handles retrieval over documents kept in Azure Storage, and an Azure Bot and an App Service connect everything to Teams.

Microsoft Teams the question comes in Azure Bot channel to Teams App Service bot logic: retrieve, then answer 1. find passages 2. write answer Azure AI Search returns the relevant passages Microsoft Foundry GPT nano model indexed Azure Storage PDF, DOCX and Markdown docs
How a question flows through the bot. It searches the indexed documents first, then asks the model to answer from what it found.

Why RAG

Documentation changes. With RAG the answers come from the documents at the moment a question is asked, so once the index is refreshed, an updated document means updated answers, with no retraining involved.

What it covers

Retrieval-augmented generation, Azure AI services, grounding a model in a fixed source of truth, and delivering it where people already work. I’m happy to go deeper on any of it, just shoot me an email.