Chatbot Training Guide
Chatbot Basic Configurations
The Model page controls how your AI agent answers: its instructions (base prompt), the AI model, and how closely it sticks to your data. Set the base prompt first; it matters most.
Where: open your chatbot, then Settings > Model.

Standard or AI Agent
The page starts with How your agent answers. Pick one of two modes:
Most chatbots should use Standard. The rest of this page describes Standard. In AI Agent mode the base model and temperature settings are hidden, because the models are managed for you. See AI Agent Mode.
Settings at a glance
| Setting (label on the page) | What it does | Default / recommendation |
|---|---|---|
| Base Prompt (system message) | Your instructions: tone, focus, limits, what to say when it doesn't know | Always write one. Mention CONTEXT INFORMATION |
| Auto Prompt | Writes or improves the base prompt for you | Free, up to 10 requests per hour |
| Time of Chatbot | Lets the AI agent know the current date and time | On, set to UTC. Pick your time zone or User Local Time |
| Base model used in the RAG processing pipeline | The AI model that writes the answers | New chatbots start on GPT 5.4 mini |
| Enable RAG Enhancement By Reranking | Finds the most relevant parts of your data more accurately | On. Keep it on (+1 credit per message) |
| Ignore Training Data Sources | Answers from the model alone, without your data | Off. Turn on only for a generic chatbot |
| Temperature | How creative answers are, from Precise to Creative | Keep it at 0 to stay close to your data |
| Nested Chatbots | Lets this chatbot ask other chatbots | Optional |
Click Save after you change anything.
Base Prompt
The base prompt is the most important setting. It sets the AI agent's tone and tells it what you expect. Without one, the chatbot falls back to a default prompt and may give unexpected or off-target answers.
Write CONTEXT INFORMATION in the prompt to refer to your training data (the content you added in Sources). This is the name Chat Data uses for it behind the scenes.
Starter template. Copy it, then replace the parts in brackets:
You are a friendly and easygoing AI assistant for [Your Company].
Specialize in answering questions about [your products or services].
Answer using the CONTEXT INFORMATION.
If the answer is not in the CONTEXT INFORMATION, say: "I'm sorry, I don't have that information. Please contact our customer support at [email or phone]."
Avoid providing [financial advice, or any topic you want to rule out].
Each line uses one of four strategies:
| Strategy | Example wording |
|---|---|
| Set the personality | "Present yourself as a friendly and easygoing AI Assistant." |
| Handle unknown questions | Instead of "Hmm, I am not sure": "I apologize, but I don't possess the information you seek; kindly reach out to our customer support." |
| Focus on your topics | "Specialize in providing information about environmental sustainability as an AI Assistant." |
| Set limits | "Avoid providing financial advice or information." |
Match the prompt to the model. Smaller, cheaper models can miss complex logic, so keep their prompts simple and add a few examples of good answers (few-shot prompting). Larger models handle more complex instructions.
Auto Prompt
Click Auto Prompt next to the base prompt, describe what you want the AI agent to do, and it writes a new prompt or improves your current one. It costs no credits and is limited to 10 requests per hour.

Chatbot Time
Time of Chatbot lets the AI agent know the current date and time, so your base prompt can include time rules (for example, opening hours). It is on by default and set to UTC.
- Pick a fixed time zone, such as your business's time zone, or
- Turn on User Local Time to use each visitor's browser time instead.

Base model
Choose the AI model under Base model used in the RAG processing pipeline. This applies to the custom-data-upload, medical-chat-human and medical-chat-vet chatbot types.
- New chatbots start on GPT 5.4 mini.
- The list includes models from OpenAI, Anthropic (Claude), Google (Gemini) and others. Each one shows its cost in credits per message.
- Some models need a higher plan.
Larger models follow the base prompt more closely and make fewer things up, but they are slower and cost more credits. If accuracy matters most, especially sticking to your data, try a larger model and compare answers in the Playground.
RAG Enhancement
When a visitor asks a question, Chat Data first finds the parts of your training data that look relevant. Enable RAG Enhancement By Reranking adds a second pass with Cohere reranker models: it re-checks a few hundred candidate text chunks and keeps the most relevant ones.
Keep it on. Finding the right content is the basis of an accurate chatbot, and a smaller model with reranking can beat a larger model without it. It adds 1 credit per message.
Temperature
Temperature sets how creative the answers are, from Precise to Creative. Keep it at 0 when the AI agent must stick closely to your data. Higher values let it add creative guesses that may not match your content.
Ignore Your Uploaded Data
Turn on Ignore Training Data Sources to build a generic chatbot that answers from the selected base model and your base prompt only, without your training data. Leave it off for a chatbot that answers from your content.
Nested Chatbots
Sometimes one chatbot can't do everything. For example, you train one on your company data, but you also want answers from a medical-chat model, or from another chatbot whose data no longer fits because of character limits. Nested chatbots let your main chatbot pass questions on to other chatbots.

A nested chatbot can be any chatbot in Chat Data, and it can have nested chatbots of its own. To add one, go to Settings > Model > Nested Chatbots, click Add a chatbot, and fill in:
| Field | What to enter |
|---|---|
| Chatbot Id | The ID of the chatbot to call |
| Function Name | A short, meaningful name for it |
| Description | What that chatbot can answer and when to use it |
| API Key | An API key from the account page of the account that owns that chatbot |
You can add several nested chatbots.

Next step: test answers and fix weak ones with Debug and Optimization.
Frequently Asked Questions
What settings matter most in basic chatbot configuration?
The most important foundational settings are the base prompt, the base model, and temperature. They live on the Settings > Model page and are the core controls for tone, accuracy, and behavior when the chatbot uses your own uploaded data.
How should I configure the base prompt and model for better accuracy?
The guide treats the base prompt as the highest-priority setting and says to include `CONTEXT INFORMATION` when using uploaded data. New chatbots start on GPT 5.4 mini. Smaller models work better with simple prompts and a few examples, while larger models handle more complex instructions and follow the prompt more closely, though they are slower and cost more credits per message.
What do time settings, reranking, and temperature change in chatbot behavior?
Time settings let the chatbot reason about current time in a chosen time zone or from the user's local browser time. It recommends enabling RAG enhancement with the reranker for better retrieval quality, and it advises keeping temperature at `0` when you want the chatbot to stay close to the uploaded data and avoid creative drift.
What are nested chatbots used for in Chat Data?
Nested chatbots let one chatbot call other chatbots when a single bot is not enough for the task. For example, you can combine your own company-data bot with other specialized bots like medical models. Each nested chatbot needs a chatbot ID, function name, description, and API key.