Why Your RAG Chatbot Setup Matters
GoodBarber's RAG Chatbot extension turns your app into an intelligent assistant that answers user questions based on your own content — articles, events, and map points. But getting the best results requires understanding a few key configuration choices.
Two questions come up repeatedly from GoodBarber users: "How far will my 5,000 monthly credits go?" and "Why isn't my chatbot finding the right content?" The answers lie in how you configure your setup. In this guide, we break down the credit system, explain the often-misunderstood Items Per Feed setting, and walk you through customizing your chatbot's personality with a well-crafted system prompt.
Whether you are evaluating the RAG Chatbot extension or have already subscribed, this article will help you avoid common pitfalls and get the most out of every credit.
Two questions come up repeatedly from GoodBarber users: "How far will my 5,000 monthly credits go?" and "Why isn't my chatbot finding the right content?" The answers lie in how you configure your setup. In this guide, we break down the credit system, explain the often-misunderstood Items Per Feed setting, and walk you through customizing your chatbot's personality with a well-crafted system prompt.
Whether you are evaluating the RAG Chatbot extension or have already subscribed, this article will help you avoid common pitfalls and get the most out of every credit.
1. Understanding Credits and Tokens: How Far Do 5,000 Credits Go?
What Are Credits?
Every GoodBarber RAG Chatbot subscription includes 5,000 credits per month, automatically renewed. Credits are GoodBarber's internal currency that covers all the AI operations your chatbot performs:
- Indexing your content (articles, events, map points)
- Processing user questions (converting them to vector representations)
- Generating responses (the most credit-intensive step)
Credits vs Tokens: Why GoodBarber Uses Credits Instead of Tokens
If you are familiar with OpenAI's pricing, you know that AI models charge per token (roughly a word or word fragment). So why doesn't GoodBarber just show you token counts?
The answer is simple: the financial cost of a token varies dramatically between models. A token processed by GPT-4o costs significantly more than the same token processed by GPT-4o-mini. GoodBarber's credit system normalizes this difference into a single, model-agnostic unit. One credit always costs the same to you, but the number of tokens you get per credit depends on which AI model you select.
Think of it this way: credits are like a universal currency, and tokens are the goods you buy — cheaper models give you more tokens per credit, while premium models give you fewer.
Real-World Estimates: How Many Messages Can You Send?
While exact consumption depends on conversation length, content complexity, and the model selected, here are practical ballpark figures based on support team guidance and user reports:
Pro tip: If you are just starting out, begin with GPT-4o-mini. It offers surprisingly good response quality for most use cases while stretching your credits 3–5x further than premium models. You can always switch to a more advanced model later once you understand your usage patterns.
What Happens When Credits Run Out?
When your monthly credits are depleted, the chatbot stops responding until either the monthly renewal kicks in or you purchase additional credits. Extra credit packs are available at any time: 5,000 additional credits for €20. GoodBarber sends you an email alert when you reach 20% remaining credits, giving you time to act before the chatbot goes silent.
How to Monitor Your Credit Usage
Navigate to your RAG Chatbot settings in the GoodBarber back office to view your current credit balance. Monitoring usage during your first weeks is essential to establish a baseline for your specific content and audience.
Every GoodBarber RAG Chatbot subscription includes 5,000 credits per month, automatically renewed. Credits are GoodBarber's internal currency that covers all the AI operations your chatbot performs:
- Indexing your content (articles, events, map points)
- Processing user questions (converting them to vector representations)
- Generating responses (the most credit-intensive step)
Credits vs Tokens: Why GoodBarber Uses Credits Instead of Tokens
If you are familiar with OpenAI's pricing, you know that AI models charge per token (roughly a word or word fragment). So why doesn't GoodBarber just show you token counts?
The answer is simple: the financial cost of a token varies dramatically between models. A token processed by GPT-4o costs significantly more than the same token processed by GPT-4o-mini. GoodBarber's credit system normalizes this difference into a single, model-agnostic unit. One credit always costs the same to you, but the number of tokens you get per credit depends on which AI model you select.
Think of it this way: credits are like a universal currency, and tokens are the goods you buy — cheaper models give you more tokens per credit, while premium models give you fewer.
Real-World Estimates: How Many Messages Can You Send?
While exact consumption depends on conversation length, content complexity, and the model selected, here are practical ballpark figures based on support team guidance and user reports:
| Model | Estimated Messages per 5,000 Credits | Best For |
|---|---|---|
| GPT-4o-mini | Up to ~10,000 messages | High volume, cost-efficient apps |
| GPT-4o | ~2,000–3,000 messages | Balanced quality and cost. |
| GPT-4.1 / GPT-5 | ~1,000–1,500 messages | Maximum response quality |
Pro tip: If you are just starting out, begin with GPT-4o-mini. It offers surprisingly good response quality for most use cases while stretching your credits 3–5x further than premium models. You can always switch to a more advanced model later once you understand your usage patterns.
What Happens When Credits Run Out?
When your monthly credits are depleted, the chatbot stops responding until either the monthly renewal kicks in or you purchase additional credits. Extra credit packs are available at any time: 5,000 additional credits for €20. GoodBarber sends you an email alert when you reach 20% remaining credits, giving you time to act before the chatbot goes silent.
How to Monitor Your Credit Usage
Navigate to your RAG Chatbot settings in the GoodBarber back office to view your current credit balance. Monitoring usage during your first weeks is essential to establish a baseline for your specific content and audience.
2. Items Per Feed: The Setting That Makes or Breaks Your Chatbot
What Does "Items Per Feed" Mean?
The Items Per Feed parameter defines how many content items the chatbot indexes from each section or subsection you have enabled. It directly controls the scope of your chatbot's knowledge base.
For example, if you have a section called "Blog" with 500 articles and set Items Per Feed to 24, the chatbot will only index the 24 most recent articles. The remaining 476 articles are invisible to the AI — it simply cannot find them, no matter how relevant they are to a user's question.
Why the Default Is Set Low (And Why That Is Actually Smart)
A common frustration: "My chatbot can't answer questions about content I know exists in my app." In many cases, the culprit is a low Items Per Feed value.
GoodBarber sets a conservative default on purpose. Here is why this is a smart design choice:
1. It promotes content freshness. Since the chatbot indexes the most recent items first, a controlled limit ensures that user queries are matched against your latest, most up-to-date content. This is particularly valuable for news, events, or any app where timeliness matters — your chatbot naturally prioritizes fresh information over older, potentially outdated articles.
2. Indexing consumes credits. Every item you index uses credits from your monthly allowance. If you have 10,000 articles and index them all on day one, you could burn through a significant portion of your credits before a single user asks a question.
3. It protects you during setup. When you are experimenting with different models, embedding sizes, and configurations, you do not want each test run to consume thousands of credits.
Good News: The Default Is Already 100 Items Per Feed
GoodBarber sets the default Items Per Feed to 100, which is a reasonable starting point for most apps. This means you can go live without touching this setting at all — your chatbot will index up to 100 items per section right out of the box.
However, if your content library is larger than that, you will want to increase this value after your initial setup is validated. If you want to index 10,000 articles, go ahead — the limit exists to protect your credits during the experimentation phase, not to restrict you permanently.
When to increase it:
- Once you have confirmed your chatbot configuration works well (correct model, good prompt, relevant answers)
- When you notice the chatbot cannot find content that you know exists in sections with more than 100 items
- After your initial testing phase, when you are ready to go into production
After changing the value, your content will be automatically re-indexed within the next cycle — GoodBarber runs automatic re-indexation every 2 hours, so there is nothing to trigger manually. Just update the setting, save, and your expanded content library will be picked up at the next sync.
The Items Per Feed parameter defines how many content items the chatbot indexes from each section or subsection you have enabled. It directly controls the scope of your chatbot's knowledge base.
For example, if you have a section called "Blog" with 500 articles and set Items Per Feed to 24, the chatbot will only index the 24 most recent articles. The remaining 476 articles are invisible to the AI — it simply cannot find them, no matter how relevant they are to a user's question.
Why the Default Is Set Low (And Why That Is Actually Smart)
A common frustration: "My chatbot can't answer questions about content I know exists in my app." In many cases, the culprit is a low Items Per Feed value.
GoodBarber sets a conservative default on purpose. Here is why this is a smart design choice:
1. It promotes content freshness. Since the chatbot indexes the most recent items first, a controlled limit ensures that user queries are matched against your latest, most up-to-date content. This is particularly valuable for news, events, or any app where timeliness matters — your chatbot naturally prioritizes fresh information over older, potentially outdated articles.
2. Indexing consumes credits. Every item you index uses credits from your monthly allowance. If you have 10,000 articles and index them all on day one, you could burn through a significant portion of your credits before a single user asks a question.
3. It protects you during setup. When you are experimenting with different models, embedding sizes, and configurations, you do not want each test run to consume thousands of credits.
Good News: The Default Is Already 100 Items Per Feed
GoodBarber sets the default Items Per Feed to 100, which is a reasonable starting point for most apps. This means you can go live without touching this setting at all — your chatbot will index up to 100 items per section right out of the box.
However, if your content library is larger than that, you will want to increase this value after your initial setup is validated. If you want to index 10,000 articles, go ahead — the limit exists to protect your credits during the experimentation phase, not to restrict you permanently.
When to increase it:
- Once you have confirmed your chatbot configuration works well (correct model, good prompt, relevant answers)
- When you notice the chatbot cannot find content that you know exists in sections with more than 100 items
- After your initial testing phase, when you are ready to go into production
After changing the value, your content will be automatically re-indexed within the next cycle — GoodBarber runs automatic re-indexation every 2 hours, so there is nothing to trigger manually. Just update the setting, save, and your expanded content library will be picked up at the next sync.
Good to know: GoodBarber automatically re-indexes your content every 2 hours. After changing the Items Per Feed value, simply save your settings and wait for the next indexation cycle — no manual action required.
How Items Per Feed Affects Search Quality
The Items Per Feed setting does not just affect "how many" results the chatbot can find — it fundamentally changes the quality of semantic search. RAG (Retrieval-Augmented Generation) works by finding the most semantically similar content to a user's question. If the relevant article was never indexed because it fell outside the Items Per Feed limit, the chatbot will either return a less relevant result or tell the user it could not find anything.
This is especially critical for apps with large content libraries. If you have 1,400 articles (like some GoodBarber users do) but only index 24, the chatbot is working with less than 2% of your knowledge base.
How Items Per Feed Affects Search Quality
The Items Per Feed setting does not just affect "how many" results the chatbot can find — it fundamentally changes the quality of semantic search. RAG (Retrieval-Augmented Generation) works by finding the most semantically similar content to a user's question. If the relevant article was never indexed because it fell outside the Items Per Feed limit, the chatbot will either return a less relevant result or tell the user it could not find anything.
This is especially critical for apps with large content libraries. If you have 1,400 articles (like some GoodBarber users do) but only index 24, the chatbot is working with less than 2% of your knowledge base.
3. Customizing Your Chatbot's Tone: The Power of the System Prompt
What Is the System Prompt?
The system prompt (also called "System Message" or "Writing Tone" in GoodBarber's interface) is a set of instructions that tells the AI how to behave, what personality to adopt, and how to structure its answers. It is the single most impactful setting for the quality and style of your chatbot's responses.
Without a system prompt, your chatbot will use a neutral, generic tone. With a well-crafted prompt, it can become a brand ambassador that matches your app's voice perfectly.
What NOT to Include in Your System Prompt
A common mistake is to include technical instructions in the system prompt. GoodBarber already handles the technical plumbing — the RAG retrieval, context injection, and response formatting are configured on the server side.
Remove these from your prompt if you have them:
- # Knowledge Source or data source instructions
- # Technical Variables or API-related parameters
- # Output Format or JSON/HTML formatting rules
The AI already knows where to look for data and how to format results. Adding these instructions can actually confuse the model or override GoodBarber's optimized settings.
What TO Include: The Three Pillars of a Good System Prompt
Focus your system prompt on three areas:
Pillar 1: Role and Identity
Tell the AI who it is. Give it a name, a role, and a purpose.
Pillar 2: Style and Tone Guidelines
Define how the AI should communicate. Be specific about language register, length, and personality.
Pillar 3: Response Behavior and Boundaries
Set guardrails for what the AI should and should not do.
Step-by-Step: How to Change Your Chatbot's Tone
Here is how to access and modify the system prompt in GoodBarber:
Step 1: Log into your GoodBarber back office
Step 2: Navigate to Settings → RAG Section → Select your chatbot section → Configuration
Step 3: Scroll down to the "System Message" or "Writing Tone" field
Step 4: Enter your customized system prompt following the three pillars above
The system prompt (also called "System Message" or "Writing Tone" in GoodBarber's interface) is a set of instructions that tells the AI how to behave, what personality to adopt, and how to structure its answers. It is the single most impactful setting for the quality and style of your chatbot's responses.
Without a system prompt, your chatbot will use a neutral, generic tone. With a well-crafted prompt, it can become a brand ambassador that matches your app's voice perfectly.
What NOT to Include in Your System Prompt
A common mistake is to include technical instructions in the system prompt. GoodBarber already handles the technical plumbing — the RAG retrieval, context injection, and response formatting are configured on the server side.
Remove these from your prompt if you have them:
- # Knowledge Source or data source instructions
- # Technical Variables or API-related parameters
- # Output Format or JSON/HTML formatting rules
The AI already knows where to look for data and how to format results. Adding these instructions can actually confuse the model or override GoodBarber's optimized settings.
What TO Include: The Three Pillars of a Good System Prompt
Focus your system prompt on three areas:
Pillar 1: Role and Identity
Tell the AI who it is. Give it a name, a role, and a purpose.
You are "Luna", a friendly and knowledgeable assistant for the HealthyLife app.
Your role is to help users find relevant health tips, recipes, and wellness advice
from our content library.
Pillar 2: Style and Tone Guidelines
Define how the AI should communicate. Be specific about language register, length, and personality.
- Use a warm, encouraging tone
- Keep answers concise (2-3 paragraphs maximum)
- Avoid medical jargon — explain concepts in simple terms
- Always end with an encouraging note
- Address users informally (use "you" and first names when possible)
Pillar 3: Response Behavior and Boundaries
Set guardrails for what the AI should and should not do.
- Only answer questions related to health, wellness, and nutrition
- If a user asks about something outside the app's content, politely redirect them
- Never invent information — if you cannot find relevant content, say so honestly
- When citing content, mention the article title so users can find the full piece
Step-by-Step: How to Change Your Chatbot's Tone
Here is how to access and modify the system prompt in GoodBarber:
Step 1: Log into your GoodBarber back office
Step 2: Navigate to Settings → RAG Section → Select your chatbot section → Configuration
Step 3: Scroll down to the "System Message" or "Writing Tone" field
Step 4: Enter your customized system prompt following the three pillars above
Step 5: Save your settings and test with a few questions
Example Prompts for Common Use Cases
News / Media app:
Community / Association app:
News / Media app:
You are a news assistant for [Publication Name]. Summarize articles clearly
and factually. Always mention the publication date so users know how recent
the information is. Maintain a neutral, journalistic tone. If multiple
articles cover the same topic, synthesize the key points.
Community / Association app:
You are the virtual concierge for [Association Name]. Help members find
information about upcoming events, meeting locations, and organizational
resources. Be warm and welcoming. Use the association's terminology and
refer to members by their appropriate titles when possible.
4. Choosing the Right AI Models: Embedding and Completion
Embedding Model: Small vs Large
The embedding model indexes your content — it converts text into vector representations that enable semantic search.
- Small: More economical in credit consumption. Good for apps with straightforward, text-heavy content.
- Large: More accurate at understanding nuance, synonyms, and context. Recommended for content with specialized vocabulary or when precision matters.
Recommendation: If your budget allows, use the Large embedding model. The quality difference in search results is significant, and the additional credit cost at indexing time is a one-time expense per content update — not per user query.
Completion Model: Balancing Cost and Quality
The completion model generates the actual response the user sees. This is where the credit-per-token trade-off matters most, because every user question triggers the completion model.
Refer to the table in Section 1 to estimate your monthly volume. For most apps, GPT-4o-mini offers the best quality-to-cost ratio.
The embedding model indexes your content — it converts text into vector representations that enable semantic search.
- Small: More economical in credit consumption. Good for apps with straightforward, text-heavy content.
- Large: More accurate at understanding nuance, synonyms, and context. Recommended for content with specialized vocabulary or when precision matters.
Recommendation: If your budget allows, use the Large embedding model. The quality difference in search results is significant, and the additional credit cost at indexing time is a one-time expense per content update — not per user query.
Completion Model: Balancing Cost and Quality
The completion model generates the actual response the user sees. This is where the credit-per-token trade-off matters most, because every user question triggers the completion model.
Refer to the table in Section 1 to estimate your monthly volume. For most apps, GPT-4o-mini offers the best quality-to-cost ratio.
Your RAG Chatbot Setup Checklist
Setting up a GoodBarber RAG Chatbot effectively comes down to three decisions: understanding your credit budget, configuring the right content scope, and giving your bot a personality that matches your brand.
Here is your quick-start checklist:
1. Start with GPT-4o-mini to maximize messages per credit during testing
2. Set Items Per Feed to 50–100 during initial configuration
3. Write a focused system prompt covering Role, Style, and Behavior — skip technical instructions
4. Test thoroughly with real user questions before scaling
5. Increase Items Per Feed to cover your full content library once configuration is validated
6. Monitor credit consumption weekly during the first month to establish your baseline
7. Switch models if needed once you understand your real-world usage patterns
Your chatbot is only as good as its setup. Take the time to configure it properly, and it will become one of the most valuable features in your app.
Here is your quick-start checklist:
1. Start with GPT-4o-mini to maximize messages per credit during testing
2. Set Items Per Feed to 50–100 during initial configuration
3. Write a focused system prompt covering Role, Style, and Behavior — skip technical instructions
4. Test thoroughly with real user questions before scaling
5. Increase Items Per Feed to cover your full content library once configuration is validated
6. Monitor credit consumption weekly during the first month to establish your baseline
7. Switch models if needed once you understand your real-world usage patterns
Your chatbot is only as good as its setup. Take the time to configure it properly, and it will become one of the most valuable features in your app.