Two settings, easy to confuse, doing quite different things.
| Welcome message | System prompt | |
|---|---|---|
| Read by | The visitor | The model |
| Where | Sohay → AI Settings | Sohay → KB Settings → Settings |
| Limit | 280 characters | 4,000 characters |
The welcome message
The first thing a visitor sees when they open the chat. It is doing one job: telling them what this thing is good at.
A specific one gets better questions than an empty box does.
- Weak: “Hi! How can I help you today?”
- Better: “Ask me about delivery, returns, or anything about our products.”
Naming what you cover also sets expectations. A visitor told the chatbot knows about shipping is less disappointed when it does not know about your hiring process.
The system prompt
The instructions the model follows when it writes answers. Leave it blank and Sohay’s default is used, which is tuned for grounded support answers — most sites never need to touch it.
Reach for it when you want to change how answers are written, not what they know. Tone, length, formality, a house rule about never quoting prices without a link.
Two placeholders are replaced before the prompt is sent:
{site_name}— your site’s title{current_date}— today’s date
{current_date} is more useful than it looks. Models have no inherent sense
of the current date, so a prompt that includes it can reason about “this
month’s sale” or “since last week” correctly.
What the system prompt cannot do
It cannot add knowledge. Facts you put in the prompt are not searchable, are sent with every single request, and eat into your 4,000 characters.
If you find yourself writing product details or opening hours into the system prompt, that content belongs in a Knowledge Base article. It will retrieve better, cost less, and be editable by anyone who can write a post.
It cannot make the chatbot know something the articles do not. Instructing it to “always answer confidently” produces confident wrong answers, not correct ones.
A prompt worth writing
Additions that genuinely earn their space:
- Tone. “Reply in two or three sentences unless asked for detail.”
- A routing rule. “For anything about an existing order, direct the visitor to support@example.com.”
- A boundary. “Never estimate delivery dates; link to the shipping page instead.”
Each of those changes behaviour the articles cannot express on their own.
If you overrun the limit
Input over 4,000 characters is truncated on save, with a notice telling you. Truncation lands mid-sentence, so check what survived rather than assuming the important part did.
Where to go next
Writing a good Knowledge Base article, which is where most “the chatbot should say X” problems actually get solved.
