Never train an AI support bot directly on raw past tickets without heavy curation. Past tickets contain obsolete product features, expired discount promises, and agent mistakes. Instead, train your AI on live website URLs, structured Notion/Confluence docs, and curated knowledge base articles, which represent verified single sources of truth.
The 'Garbage In, Garbage Out' Training Dilemma
When engineering teams decide to implement an AI support agent, the most common instinct is to dump their entire 5-year Zendesk or Freshdesk ticket archive into a vector database: 'We have 50,000 real conversations, the AI will learn everything!'
This is almost always a catastrophic mistake. Here is why raw ticket archives pollute AI models:
- Obsolete Policies & Pricing: A support ticket from 2022 might show an agent granting a special grandfathered $19/month plan or explaining a legacy UI workflow that was deprecated two years ago. The AI will cite this outdated information as current fact.
- Human Agent Errors & Hallucinations: Human agents occasionally give incorrect answers, misinterpret terms of service, or make unauthorized commitments. Ingesting raw tickets codifies human errors into the AI's permanent memory.
- Contradictory Information: Two different agents may have answered the same edge-case question differently. When fed conflicting data, vector retrieval models produce erratic, unpredictable responses.
| Knowledge Source | Data Hygiene Level | Risk of Outdated Policy Hallucination | Maintenance Overhead |
|---|---|---|---|
| Raw Zendesk / Helpdesk Tickets | Very Poor (Dirty, unvetted data) | Extreme (Quotes old discounts & bugs) | Heavy (Requires manual data scrubbing) |
| Static PDF Handbooks | Moderate (Fixed point in time) | High (Becomes stale quickly) | Moderate (Must re-upload PDFs) |
| Live Website URLs & Help Center | Highest (Single source of truth) | Zero (Reflects current production) | Zero (Auto-crawled & synced live) |
| Structured Notion / Confluence Docs | Very High (Curated by team) | Very Low (Direct editorial oversight) | Low (Auto-synced via webhook) |
4 Steps to Build a Pristine Knowledge Base for Support AI
- Establish Single Sources of Truth: Designate your public help center or an internal Notion workspace as the sole authoritative repository of company policy.
- Structure Content with Clear Headings: Format documentation in semantic Markdown using clear `H2` and `H3` questions that mirror real customer phrasing.
- Automate Dynamic Web Crawling: Configure your AI engine to re-index live documentation URLs whenever your product team publishes an update.
- Mine Tickets for Questions, Not Answers: Use past tickets to discover what questions customers ask most frequently, but write verified editorial answers in your knowledge base.
Keep your support responses 100% accurate. Seatext AI Support Assistant syncs directly with live URLs and structured docs, eliminating outdated ticket clutter.
Sync Your Knowledge Base →Frequently Asked Questions
How often does Seatext re-crawl our website for changes?
Seatext can crawl automatically on a daily schedule, or you can trigger an instant re-sync via webhook whenever you publish product updates.
Can we include private internal knowledge that isn't on our public website?
Yes. You can connect private Notion pages, Google Docs, or upload curated Markdown files accessible only to the AI assistant.
How does the AI handle formatting like code snippets or tables?
Seatext natively renders syntax-highlighted code blocks, clean tables, and clickable markdown links in chat responses.
