A support bot that won't make things up
For the ElitHome Instagram bot, the knowledge base is a folder of Markdown notes in Git. Staff edit it in Obsidian, and the bot answers only from what's written there.
A customer-support bot for a real-estate business has one job it must never get wrong: it cannot invent a price, a floor or an apartment that doesn't exist. For the ElitHome Instagram bot I built the knowledge base around that rule.
The knowledge base is just Markdown
Everything the bot knows lives in a Git repository, in a faq/ folder of Markdown files — apartments, prices, districts, general information. One topic or district per file, with clear headings and concrete facts: district, rooms, area in m², floor, price, status.
The staff who maintain it are not developers, so they don't edit it like developers. They open the repository as an Obsidian vault, on a phone or a desktop, edit a note, and commit and push with a Git plugin. The bot picks up the change within about two minutes. Nobody has to redeploy anything.
How the bot reads it
The bot runs on Railway and reads the notes through the GitHub API with a read-only token. Its configuration is three variables: the repository, the folder, and the token.
FAQ_GITHUB_REPO=IkboljonMe/elithome-faq
FAQ_GITHUB_PATH=faq
FAQ_GITHUB_TOKEN=<read-only token>The rule that matters
If a fact isn't written down, the bot does not invent it — it offers a callback from a manager instead.
That one rule changes how everyone treats the notes. Missing information is no longer a bot problem; it's a gap in the knowledge base that someone can fill in a minute. And customers get either a correct answer or a person, never a confident guess.