All guides
    Setup

    Training an AI support agent on your addon documentation

    Turn your existing FAQ, product docs, and source code into an AI support agent that answers customer questions correctly.

    An AI support agent is only as good as what you feed it. This guide covers exactly what to include in training data, what to exclude, and how to structure it for accurate, on-brand replies.

    The training data hierarchy

    Not all training data is equal. In order of impact:

    1. FAQs — direct question/answer pairs. Highest accuracy per character.
    2. Product documentation — install guides, config references, changelogs.
    3. Ticket history — past resolved tickets, cleaned up.
    4. Source code — for technical questions. Context only; the agent will never quote it back.

    Start with 1 and 2. Add 3 once you have volume. Add 4 only if customers ask hook/API questions.

    What to include

    Include:

    • Every FAQ you've written, in Q&A format
    • Complete installation instructions (per platform, per host, per game version)
    • Configuration reference — every option, what it does, valid values
    • Known errors with exact messages and fixes
    • Compatibility matrices (game version, other addons, server hosts)
    • Your refund/support policy in your own words

    Exclude:

    • Marketing copy from your product page
    • Changelog entries older than 6 months (unless still relevant)
    • Duplicate information rephrased three different ways
    • Anything vague, aspirational, or "coming soon"

    Vague content is worse than no content — the agent will confidently repeat it.

    Format for accuracy

    AI agents parse structured content far better than prose. A well-formatted training doc looks like:

    ## Installation

1. Download the .zip file from your purchase page
2. Extract to your server's `addons/` folder
3. Restart the server

## Common errors

### "Attempt to index nil value (global 'DarkRP')"

Cause: DarkRP is not installed or not loaded before this addon.
Fix: Install DarkRP first, then reload the addon.

### "Config file not found"

Cause: The addon looks for `myaddon/config.lua` in `data/`.
Fix: Rename `config.example.lua` to `config.lua`.

    Named errors, exact quotes, explicit fixes. That's the pattern.

    Character budget

    Arnvio caps extracted training data at 50,000 characters per file and roughly 8,000 characters in the active prompt budget per reply. This means:

    • More documentation ≠ better answers. Past a certain point, the agent has to choose what to include per reply.
    • Split large products into topic-scoped documents ("installation.md", "configuration.md", "api-reference.md"). Arnvio's retrieval will pick the most relevant one per query.
    • Delete stale content aggressively. Old changelog entries, deprecated features, and abandoned config options all dilute quality.

    System prompt: the amplifier

    Alongside training data, your agent has a system prompt that sets tone, scope, and rules. A good one:

    You are the support agent for [Product Name] by [Your Name]. Answer customer questions about installation, configuration, and troubleshooting based only on the provided documentation. If you're not confident in an answer, say so and offer to escalate. Never quote source code. Never make promises about features not documented. Match the customer's tone — casual, professional, or technical.

    Explicit boundaries produce more accurate agents than open-ended ones. "Answer based only on the provided documentation" is doing a lot of work in that prompt.

    Iterating

    The first week of training is 80% of the win. The workflow:

    1. Enable draft mode for one week
    2. Review every drafted reply the agent produced
    3. For each miss, note the question and add the correct answer to training
    4. Deploy updated training, re-run

    After a week you'll have covered nearly every real-world question. New question categories emerge slowly after that — usually only when you ship a new feature.

    What "trained enough" looks like

    • 80%+ of drafts are approved as-is
    • Escalations are actually things you should be handling (bugs, refunds, feature requests)
    • New tickets get answered in under 2 minutes, not 2 hours

    At that point, flip to auto-send. Keep an eye on the weekly summary Arnvio emails and re-tune when patterns emerge.

    Next steps