Chatbot settings explained: model, prompt and security
A beginner-friendly field-by-field walkthrough with a real EU-model configuration, exact domain rules and two verified answer examples.
The right chatbot settings connect model quality, a clear system prompt, domain security and predictable fallback behavior.
The Settings tab controls the assistant identity, where its widget may run and how every answer is generated. This guide explains each visible field in the current English Light Mode interface and shows a complete, realistic configuration instead of isolated placeholders.
Our verified example is Northstar Support. It uses two exact production hostnames, Europe/Berlin, English, a Google Vertex EU model, 0.2 answer creativity, streaming and a short role with a strict no-invention rule. We then test one known fact and one missing policy.
Model availability changes over time, so the live dropdown remains authoritative. Record a baseline, change one setting at a time and repeat the same questions before deciding that a new configuration is better.
Privacy-protected two-click player
Chatbot Settings Explained: Model, Prompt & Security
Learn the essential chatbot settings for model choice, system prompt, domain security, language and safe answer behavior, with two verified tests.
The YouTube player stays blocked until you choose Play. Loading it connects your browser to YouTube and may transfer technical data to Google.
Open directly on YouTubeWhat you will have at the end
- A clear public service name and internal chatbot name
- Exact allowed-website rules without accidental access
- A documented provider, model, quota and processing-location choice
- A focused answer style and role with a safe missing-information fallback
- A repeatable known-answer and unknown-answer test
Before you start
- An existing chatbot on a paid plan or active trial
- Every exact production hostname that should load the widget
- Your primary language and business timezone
- At least one known source fact and one deliberately unsupported question
How configuration affects every answer
Provider and model shape quality, latency, quota usage and processing location. Temperature changes variation, streaming changes how quickly text appears, while language, timezone and the system prompt add behavioral context.
Changing the provider also changes the embedding space. Existing data sources then require re-indexing before grounded answers work again.
01–09
Set it up step by step
Open the correct chatbot and select Settings
Confirm the assistant name before changing anything.
Open Northstar Support under AI Assistants and select Settings. The breadcrumb, page heading and highlighted sidebar entry must all show the same chatbot. The save indicator on the right distinguishes No unsaved changes, saving, saved and error states.
The thin Advanced settings link above the tabs is for developer-facing identifiers and is not the same as Advanced AI settings inside Answer behavior.
- Example chatbot: Northstar Support
- Expected status before editing: No unsaved changes
Separate the public service name from the internal chatbot name
One name helps visitors; the other keeps the dashboard organized.
Service Name / Company Name is the brand context used in responses and visitor-facing places. Chatbot name is the internal dashboard label used in lists, filters and reports.
In the example, Northstar Services is the company while Northstar Support identifies this one assistant. Avoid naming both fields simply Test or Assistant, especially when several bots share an account.
- Public service: Northstar Services
- Internal chatbot: Northstar Support
Allow every real hostname—and nothing else
Allowed websites prevents unauthorized widget use.
Enter hostnames without paths, one by one, and press Enter after each entry. A hostname and its subdomain are separate: `www.northstar-demo.example` does not automatically include `support.northstar-demo.example`.
Our first test intentionally returned HTTP 403 after the production list was active because the local preview hostname was not allowed. That proves the restriction works. Test before launch from every production and staging hostname you actually need; remove temporary local entries afterwards.
- Allowed: www.northstar-demo.example
- Also allowed: support.northstar-demo.example
- Not allowed: unrelated.example
Set timezone and fallback language from business reality
Dates, appointments and short messages need explicit context.
Timezone is used for appointment scheduling and time references. Choose the business location whose opening hours and calendar rules the bot represents, not automatically the administrator’s laptop timezone.
Default Language is the fallback for short or ambiguous input such as “Hi” and is especially important in WhatsApp. It does not stop visitors from writing another language when automatic language handling is enabled elsewhere.
- Example timezone: (GMT+2) Central European Summer Time – Berlin
- Example fallback: English (EN)
Choose the AI provider by policy, quota and location
The provider defines the model family and processing route.
Expand Advanced AI settings. The provider field tells you who generates answers and, for managed providers, how many included messages are available. EU labels indicate the documented processing option; verify your own legal and contractual requirements separately.
Northstar Support uses Google Vertex (EU) with 6,000 included messages per month on the Premium demo. A provider change can require re-indexing because the embedding space may change. Finish that re-index before judging answer quality.
- Selected provider: Google Vertex (EU)
- Observed usage: 4 / 6,000 internal messages
Choose the fastest economical model that passes the same tests
Speed hints and multipliers are decision aids, not quality guarantees.
The model dropdown belongs to the selected provider. Labels show speed or intelligence hints, EU status and—when applicable—a quota multiplier. A 5× model deducts five units for one answer, so compare the gain against latency and allowance.
The example uses `gemini-2.5-flash` EU because a support FAQ benefits from quick retrieval and it passes the two controlled questions. Model names change; use the current dropdown and store the exact selection with the test date.
- Baseline model: gemini-2.5-flash EU
- Upgrade only after the identical evaluation set improves materially
Tune answer creativity and streaming separately
Creativity changes variation; streaming changes presentation.
Answer creativity controls how much wording and choice may vary. For factual support, start low—the example uses 0.2—and raise it only for tasks that benefit from alternatives or a more expressive voice. Reset to default returns to the model-specific recommendation.
Enable Streaming makes text appear while it is generated. It usually improves perceived speed but does not make the model smarter or change the factual source. Test both fast and slow connections for stable rendering.
- Support baseline: Answer creativity 0.2
- Enable Streaming: on
Write an operational answer style and role
Define identity, evidence, format, fallback and forbidden behavior.
Open Answer style and role. Start from a preset only if it matches the job, then replace generic wording with the real service, desired tone, evidence rule, length, missing-information behavior and hard boundaries.
The Northstar prompt says to answer in clear friendly English, stay under 100 words, use indexed sources, admit missing information, offer a support handoff and never invent prices, warranties, availability or policies. A role shapes behavior; it cannot create facts that are absent from the knowledge base.
- Role: Northstar website support
- Evidence: indexed sources
- Fallback: admit the gap and offer handoff
- Never invent: prices, warranties, availability or policy details
Verify one known answer and one honest non-answer
The pair proves both usefulness and restraint.
Open a fresh conversation and ask a source-backed question first. Northstar Support answered the urgent incident question with the indexed code `NORDSTERN-42`. Then ask something deliberately missing: “What is the 2028 warranty policy?”
The verified reply says it does not have that information and offers to forward the question to support. That is the desired behavior: answer known facts precisely, refuse to manufacture unsupported policy, and provide a useful next step. Repeat this pair after every provider, model, source or role change.
- Known path: exact code NORDSTERN-42
- Unknown path: no invented warranty; support handoff offered
Example & result
See the practical test and its result
Every tutorial includes a fixed input, the expected outcome and a transparent record of what was actually verified locally.
Practical example: WebChatAgent chatbot settings explained: every field with examples
This exact scenario was completed with the temporary tutorial account.
Exact test input
What is the verification code in the uploaded tutorial PDF? Answer in one sentence and name the source.
Expected result
One concise sentence containing NORDSTERN-42 and the source document name.
What was actually verified
The live assistant answered: “The verification code is NORDSTERN-42, as stated in the Northstar Services Fictional Demo Knowledge Base.”
Tips & tricks
Make the setup reliable
Test with realistic examples, record your baseline and change one setting at a time. That makes real improvements visible.
Use Model Arena for comparisons
Compare models with identical questions and source content instead of relying on a single impressive answer.
Keep prompts operational
State what the bot should do when evidence is missing; do not hide missing knowledge behind personality instructions.
When something does not work
Troubleshooting
Check status, permissions and test data systematically before changing the model or prompt.
Sources show “Re-indexing required”
The provider changed. Re-index all sources before evaluating grounded answers.
Widget does not load on the site
Add the exact hostname, including its subdomain, to Allowed Domains and clear the site cache.
Ready for a production-style test
Save one configuration record with service name, hostnames, timezone, language, provider, model, creativity, streaming state, role version and test date. Re-run the known/unknown pair plus at least three representative customer questions after every major model, source or role change.