Best Practices

Use this checklist when preparing a SupportGPT agent for production.

Keep Sources Clean

• Upload only current policies, docs, and FAQs.
• Delete old versions instead of leaving conflicting files in the workspace.
• Prefer concise TXT or DOCX files when PDFs contain complex layouts.
• Break large policy sets into smaller topic-based files.
• Use Q&A pairs for high-value questions that need exact phrasing.

Make Website Crawls Focused

Website crawling works best on public HTML pages with clear content.

• Start with a small set of high-signal pages.
• Use include paths for product, docs, pricing, or FAQ sections.
• Use exclude paths for blogs, changelog archives, marketing pages, or noisy sections.
• Re-crawl when your website changes.

Tune Retrieval

In Settings -> AI, use Match Threshold to control how strict source matching should be.

Tune Generation

Use Temperature to control response variability.

For support agents, start conservative and raise temperature only when responses feel too rigid.

Write Practical Instructions

Good instructions usually include:

• The agent's role.
• Brand tone.
• What to do when sources are missing.
• Topics the agent must not answer.
• When to escalate to a human.
• Formatting preferences, such as short paragraphs or bullets.

Example:

You are the SupportGPT support assistant for Acme.
Answer using the workspace knowledge sources when possible.
If the source does not contain the answer, say you are not sure and ask for a human handoff.
Keep answers concise and friendly.
Do not invent pricing, legal, medical, or security details.

Configure Human Handoff

Enable Settings -> Human Handoff when a human team will monitor the inbox.

Useful settings:

• Show the "Talk to Human" button.
• Require visitor data such as email, user ID, or name before showing the button.
• Escalate on negative sentiment.
• Add trigger keywords like speak to human or real person.
• Set a waiting message and offline message.

Team members should keep the Live Inbox open and set their status online when they are available.

Use Chat Logs As Training Feedback

Review Activity -> Chat logs every week.

When a response is weak:

1. Check whether the source actually contains the answer.
2. Delete stale or contradictory source files.
3. Add a precise Q&A item when wording matters.
4. Update instructions for repeated behavior issues.
5. Use Revise on a message when you want to preserve a corrected answer pattern.

Test Before Deployment

Use Playground for internal testing, then use the embedded widget on a staging page before going live.

Test:

• Common customer questions.
• Follow-up questions in the same conversation.
• Questions outside your source data.
• Human handoff trigger phrases.
• Widget appearance on desktop and mobile.
• Identified visitor data if your site calls window.supportgpt.identify().